Search

As an interventional cardiologist, my PhD research was motivated by the clinical problem of contrast-induced acute kidney injury (CI-AKI) that is a potential complication arising from the numerous coronary angiography and percutaneous coronary interventions that I do. Although most contrast media are considered safe, at-risk individuals can develop CI-AKI. CI-AKI is not an uncommon complication after coronary angiographic procedures with a reported incidence of 1-2% in the general population and could be as high as 50% in high-risk patient subgroups. My PhD, which is divided into 2 parts, is to provide new insights into risk prediction of CI-AKI and peri-procedural nephroprotective therapies. Specifically, my PhD addressed research questions related to firstly, the risk prediction of CI-AKI in patients undergoing percutaneous intervention and examine if patients could be identified for periprocedural intervention and secondly to test the potential benefits of a peri-procedural therapeutic intervention to prevent CI-AKI in patients undergoing percutaneous interventions. CI-AKI risk prediction scores have been developed but these often involved the use of contrast volume in score calculations that limits their application in identifying patients who might benefit from peri-procedural strategies that protect the kidney. In the first part of my PhD thesis, I evaluated 2 recently proposed pre-procedural risk scores that do not include procedural variables including the contrast volume: Modified Mehran Score and the score developed by Liu and colleagues (Liu Prediction Score) to determine their ability to identify at risk patients undergoing chronic total occlusion percutaneous intervention (CTO PCI). I prospectively enrolled 329 consecutive and eligible patients (mean age 60.9±8.6 years, 64.1% Males; baseline eGFR 76.1±22.5 ml/min, serum creatinine 1.0±0.2 mg/dL) undergoing CTO PCI at our institution from December 1st 2017 to October 28 2020. All patients provided written informed consent to participate in this clinical study. The Modified Pre-procedural Mehran Score utilized the variables in Mehran score except the contrast volume. The Liu Prediction Score utilized four key risk factors (age ≥75 years, LVEF < 40%, serum albumin < 30 g/L and Serum Creatinine (SCr) >1.5 mg/dL) for predicting CI-AKI. The performance of the Modified Mehran Score and Liu Prediction Score were compared with the performance of the original Mehran score using area under the receiver operator characteristic curve (AUROC). Sensitivity and specificity were also compared for different cut-off values. 29 (8.8%) cases developed CI-AKI, defined as a 25% increase or an absolute increase in SCr ≥0.5 mg/dL over baseline within 48-72 h after contrast medium exposure. The removal of contrast volume from Mehran score resulted in no loss of discrimination (AUROC 0.9591 vs 0.9514 for the original Mehran and the modified Mehran respectively, P=NS). Full Mehran score had an optimal cut-point of 11 that resulted in a 0.90 accuracy of predicting CI-AKI with a sensitivity of 93.1% and a specificity of 90.0%. Modified Mehran score had an optimal cut-point of 7 that resulted in a 0.85 accuracy of predicting CI-AKI with a sensitivity of 96.6% and a specificity of 83.3%. With an optimal cut point of 5.5, the Liu score that was determined in 208 patients that had available serum albumin resulted in a 0.90 accuracy of predicting CI-AKI with a sensitivity of 100.0% and a specificity of 89.3%. A recent meta-analysis of high-volume forced diuresis with matched hydration using the RenalGuard medical device system revealed a significant reduction in the risk of CI-AKI, major adverse cardiac event rate, and the need for renal replacement therapy. However, the RenalGuard device is not widely available and affordable especially in low-middle income countries such as Kurdistan. In the second part of my PhD thesis, I designed and conducted the CINEMA trial (ISRCTN Registry Number: 72194653) to evaluate the potential benefits of a non-automated matched hydration and forced diuresis (MHFD) protocol compared to current hydration protocol in the prevention of contrast induced AKI. A total of 1,205 consecutive patients with chronic kidney disease (CKD) undergoing coronary procedures were randomized to either non-automated MHFD (MHFD group, n = 799) or standard intravenous isotonic saline hydration (control group; n= 406). The MHFD group received 250 ml normal saline (NS) bolus over 30 minutes before the coronary procedure followed by an intravenous bolus (0.5 mg/kg) of furosemide. Hydration infusion rate was manually adjusted to replace the patient's urine output. When urine output rate reached >300 ml/h, patients underwent coronary procedure. Matched fluid replacement was maintained during the procedure and for 4 h post-treatment. CI-AKI was defined conventionally as ≥25% or ≥0.5 mg/dl rise in serum creatinine over baseline. In the MHFD group, less patients developed CI-AKI (MHFD vs control: CI-AKI, 64 (8.01%) patients vs 57 (14.04%), p < 0.001). Incidence of cumulative in-hospital clinical complications was similar in both MHFD-treated patients and controls. In conclusion, in the CINEMA trial, I have shown that a non- automated MHFD protocol is effective and safe method for the prevention of CI-AKI in patients with CKD. I have also shown that both a modified pre-procedural Mehran score as well as the Liu Risk Score have potential use to identify patients undergoing CTO PCI for peri-procedural nephroprotective therapies such as with the matched hydration protocol as in the CINEMA trial. Taken together, my research findings have shown the utility of pre-procedural risk prediction scores to identify at risk individuals who could be targeted for a non-automated MHFD to prevent CI-AKI.

Background: Lung cancer (LC) is the leading cause of cancer mortality worldwide. Poor survival is driven by late onset of non-specific symptoms, resulting in advanced stage diagnoses. Evidence for the efficacy of low-dose CT (LDCT) screening in detecting cancer earlier, thereby reducing lung-cancer specific mortality, is now well established. Attention has turned to developing and implementing screening programmes in the population. A key aspect of an effective screening programme is the successful selection of participants; this ensures a favourable benefit-to-harm ratio for participants and an efficient and cost-effective programme. This thesis aims to improve screening selection in socio-economically disadvantaged populations by identifying areas of sub-optimal performance and considering strategies for further improvement. The focus on socio-economically deprived populations is of particular importance, as lung cancer risk is often higher in these subgroups, positioning them to be an ideal target population for LDCT screening. Methods: I) A retrospective study of the Manchester Lung Health Check (LHC) pilot, a community-based LC screening programme, comparing the selection performance and calibration of National Lung Screening Trial (NLST) criteria and two risk prediction models (RPMs) (PLCOM2012 and LLPV2), as well as the comorbidity profile of the screening cohort. II) Retrospective modelling of a benefit-based selection approach (LYFS-CT) in the LHC pilot, comparing performance with a risk-based approach and examining the characteristics and outcomes of the screening cohort. III) A Manchester-based case-control study validating nine published polygenic risk score (PRS) tools and assessing if they could improve risk prediction. IV) A cross-sectional questionnaire study of LHC programme participants, examining risk perception, worry and disease knowledge. Results: There were significant differences in screening selection performance based on the method of selection used. RPMs contributed to increased screening efficiency compared to NLST, but under-estimated LC risk in this population and selected a screening cohort with high levels of comorbidity. Inclusion of spirometry (FEV1/FVC ratio) or coronary artery calcification in RPMs may improve risk prediction but would further increase participant comorbidity. LYFS-CT selected significantly younger and less comorbid participants but also directed screening away from the most socio-economically disadvantaged. Eight PRS tools were successfully validated in the Manchester cohort and two novel genetic loci were identified for possible inclusion in a future PRS. Participants' comparative risk perception was more accurate than absolute risk perception. Women and those at high LC risk were more likely to have adverse psychological indicators. Conclusion: Risk-based selection leads to high screening efficiency, but RPMs are not well calibrated for use in socio-economically deprived populations and the optimal RPM and risk threshold strategy is unclear. Benefit-based selection may be an important tool for maximising the screening benefit provided to participants. Prospective studies are required to further elucidate the most advantageous selection strategy. Inclusion of genetic risk factors in RPMs may improve both risk- and benefit-based screening selection. Comparative-based language and decision aids should be employed for communicating risk to screening participants and ensuring effective shared decision making.

Genetics contribute substantially to the ability of the immune system to respond appropriately to a challenge. Consequently, many infectious and inflammatory diseases have a heritable component. As genome-wide association studies provide increasing data linking genetic variants to disease, we can leverage this information to gain insights into disease biology if we can elucidate the mechanisms underlying an observed association. Data-driven bioinformatic approaches, in vitro mechanistic studies and whole-organism approaches to study integrated pathophysiological systems provide complementary information to help establish causal links between specific variants and their effects on a target gene, and between that gene and disease pathogenesis. Cells of the mononuclear phagocyte system play key roles in regulation of innate and adaptive immune responses, either via secreted mediators or through direct cell-to-cell contacts. To advance understanding of how these regulatory processes, and genetic variations therein, shape the course of disease, I have combined experimental and bioinformatic approaches to explore novel genetic associations between genes for macrophage surface receptors and human disease. Specifically, functional follow-up of an association between adhesion G-protein-coupled receptor CD97 and severe influenza showed, using a mouse model, that deficiency of this receptor reduces the efficiency of the CD8+ T-lymphocyte response, a process critical to clearance of infected cells. Secondly, I addressed the question of how to identify causal variants in a disease-associated linkage disequilibrium block, for an association between macrophage regulatory receptor SIRPα and schizophrenia. To achieve this, I developed novel methodology for targeted locus screening using CRISPR/Cas9 mutagenesis, and identified a number of plausible causal regulatory variants that could affect expression of this receptor. Combining variant-level information with gene-level studies of disease pathophysiology can provide valuable insights into genetic causation of immune dys-regulation leading to disease, which may be harnessed for improved personalised disease risk prediction, or to identify therapeutically targetable pathways.

Chronic kidney disease (CKD) is a highly prevalent condition worldwide that is associated with major health consequences. Patients with progressive and advanced disease can experience significant multi-system complications as well as an increased risk of end-stage renal disease (ESRD) and mortality. Using data from the Salford Kidney Study, an ongoing observational cohort study of non-dialysis patients with CKD stages 3 to 5, this thesis aimed to explore two major research themes related to optimising CKD care. The first was risk prediction in CKD: the capacity to accurately risk stratify patients in order to enable timely, targeted treatment to those most likely to sustain adverse outcomes. The second concerned better hyperkalaemia management: effectively negating the occurrence of hyperkalaemia can facilitate continuation and optimisation of renin-angiotensin-aldosterone system inhibitors (RAASi), which are well established renoprotective agents. Results chapters 4 to 8 focused on areas in the field of risk prediction. Chapter 4 highlighted the differential impact of risk determinants in different primary renal disease groups with emphasis on ESRD and mortality for those with rapid linear progression and those with stable disease. Those progressing rapidly can be further differentiated with respect to clinical endpoints by their pattern of progression, detailed in chapter 5. Chapter 6 reinforced the need to account for the rate and pattern of CKD progression when attempting to identify novel biomarkers of disease. Risk prediction tools exist such as the Kidney Failure Risk Equation (KFRE), designed to predict the 2- and 5-year risk of ESRD in patients with CKD stages 3-5. Chapter 7 demonstrated that the KFRE is not wholly accurate when used for risk prediction in transplant recipients but does have clinical utility in those with advanced CKD as analysed in chapter 8. This provides compelling evidence for shifting towards risk-based tools to guide decision making in clinical practice. With respect to hyperkalaemia management, chapter 9 emphasised the efficacy of patiromer, a novel potassium binder, in maintaining normokalaemia. This effect was harnessed to successfully up-titrate RAASi dosing in patients with symptomatic heart failure attending a bespoke hyperkalaemia clinic, which was described in chapter 10. Whilst all the studies in this thesis have helped to deepen our understanding in the field of CKD care, the strength of the KFRE for risk prediction in advanced CKD and the real-world experience of the effectiveness of oral potassium binders have emerged as major findings. This will hopefully provide optimism for future research into tackling the heterogeneity and complexity of CKD management so as to improve long-term patient prognosis.

Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality worldwide. Identifying individuals who are at higher risk of CVD is fundamental for effectively implementing prevention strategies with limited health care resources and subsequently reducing the burden of CVD. For this purpose, numerous prognostic cardiovascular risk prediction models have been developed in populations from different regions over the past two decades. However, there are limitations of existing risk prediction models. First, they are mostly based on single measurements of risk factors and there is limited evidence quantifying the value of longitudinal risk predictor measures. Therefore, the first aim of this thesis is to evaluate the role of repeated risk factor measures on CVD risk prediction, with a focus on people with type 2 diabetes who are regularly monitored and have more measurements. Second, few models have considered effect of post-baseline statin initiation, which may lead to an underestimation of an individual's future risk of disease. Thus, the second aim is to explore novel approaches to account for post-baseline statin initiation in CVD risk prediction models. Third, a single fixed risk threshold for treatment initiation is typically recommended in most guidelines, however, such strategy does not account for the large impact of age and sex on CVD risk. Consequently, the third aim is to investigate age- and sex-specific thresholds for CVD risk stratification. These questions are addressed using electronic health records (EHRs) from approximately two million individuals from the UK Clinical Practice Research Datalink (CPRD), together with the linked data from Hospital Episode Statistics (HES) and the Office for National Statistics (ONS). Key findings 1: By applying landmark modelling to EHRs for people with type 2 diabetes, models incorporating trajectories and variability of risk predictors demonstrated significant improvement in risk discrimination (C-index=0.659, 95% Confidence Interval: 0.654-0.663) as compared to using last observed values (0.651, 0.646-0.656) or means (0.650, 0.645-0.655). Inclusion of standard deviations (SDs) of systolic blood pressure yielded the greatest improvement in discrimination (C-index increase=0.005, 95% Confidence Interval: 0.004-0.007) in comparison to incorporating SDs of total cholesterol (0.002, 0.000-0.003), HbA1c (0.002, 0.000-0.003), or high-density lipoprotein cholesterol (0.003, 0.002-0.005). Given that repeat measures are readily available in EHRs especially for regularly monitored patients with diabetes, this improvement could easily be achieved. Key findings 2: To account for statin initiation in CVD risk prediction, I incorporated a time-dependent effect of statin initiation constrained to a 25% relative risk reduction (from trial results) into the risk prediction models. In models accounting for (versus ignoring) statin initiation, 10-year CVD risk predictions were slightly higher; predictive performance was moderately improved. However, few individuals were reclassified to a high-risk threshold, resulting in negligible improvements in number needed to screen to prevent one CVD event. In conclusion, incorporating statin effects from trial results into risk prediction models enabled statin-naïve CVD risk estimation and provides moderate gains in predictive ability but had a limited impact on treatment decision-making under current guidelines in this population. Key findings 3: Age- and sex-specific risk thresholds were specified as the minimum of 10% or the 90th percentile of the estimated risk distributions from the respective populations. Compared with the single threshold of 10%, using age- and sex-specific thresholds significantly improved the discriminatory ability to identify high-risk men and women at younger ages. The number needed to screen to prevent one CVD event was reduced by 58% and 89% for women and women aged 40 to 49. The gain in CVD-free life expectancy by age and sex was slightly higher when the strategy identified more people as high-risk for younger age groups, with a maximum increase of 0.16 years. In conclusion, the results suggest using age- and sex-specific thresholds can modestly enhance CVD risk stratification for allocation of statin therapy among younger people. Overall, these findings have identified achievable and pragmatic approaches to improve CVD risk prediction and risk stratification for allocating statin initiation by harnessing information from electronic medical records.

Fetal growth restriction (FGR) is a common complication of pregnancy and the single largest risk factor for stillbirth in high-resource settings. FGR can be classified by the gestation at diagnosis, using 32-34 weeks' gestation to differentiate early-onset from late-onset disease. Within early-onset FGR there appears to be an extreme subset, typically diagnosed before 28 weeks' and delivered prior to 33 weeks' gestation which is associated with more severe placental disease and worse outcomes. It is this extreme subtype, termed eFGR, which forms the focus of this thesis. With no therapeutic interventions available, eFGR management is based upon antenatal surveillance using ultrasonography and fetal heart rate monitoring. The goal of management is optimise outcome, whilst avoiding preventable stillbirth. However, there is a paucity of eFGR specific knowledge about survival and outcome, and further study is required to provide better information for clinicians and parents of affected babies. It was hypothesised that in cases of eFGR, antenatal factors, ultrasonography and FHR patterns can be used to better predict outcomes. This hypothesis was tested by: 1) Using population level and local data to determine the current incidence of eFGR and investigate the relationship between gestational age at delivery, birthweight and survival; 2) Performing a prognostic factor study to better characterise eFGR and identify if antenatal and ultrasound characteristics can be used to reliably predict outcome in individual cases of eFGR; 3) Prospectively examining the validity and potential value of ambulatory 24-hour fetal heart rate monitoring in cases of eFGR. The incidence of eFGR was confirmed to be 3 per 1000 births in the general maternity population. The relationship between gestational age, birthweight and neonatal death in preterm infants was explored to suggest the two can be combined to predict outcome. In the case of eFGR infants with static growth, there is likely to be little advantage by gaining gestation. On an individual level, a combination of ultrasound measurements at diagnosis can be used to predict the likelihood of stillbirth. In addition, longitudinal data collected over the course of an eFGR pregnancy relating to fetal growth and Doppler progression can be used to modify risk predictions as the pregnancy progresses. Finally, a comparison of computerised cardiotocography parameters in eFGR and normal pregnancies, although highlighting differences between the two, suggests that further work is required in this area to determine how analysis of fetal heart rate can be improved as a tool for predicting prognosis in these high-risk pregnancies. This set of studies has provided improved predictions of pregnancy outcomes for both clinicians and affected families. It has highlighted areas for further development which should translate in the future to improved management, subsequently reducing associated morbidity and mortality.

My thesis considers the theme of comorbidity between cardiometabolic disorders and schizophrenia by focussing on three key aspects: the nature of association between cardiometabolic disorders and schizophrenia; the potential for common underlying biological mechanisms for the comorbidity; and the prediction of cardiometabolic risk in young adults with psychosis. On the nature of association between cardiometabolic disorders and schizophrenia, using longitudinal repeat measure data from a large birth cohort, I found that disruption to glucose-insulin homeostasis through childhood/adolescence is associated with increased risk of psychosis in early-adulthood; may not be fully explained by common sociodemographic and lifestyle factors; and may be specific to it. On the mechanisms of association between cardiometabolic disorders and schizophrenia, I used a range of genetic and observational epidemiological methods to examine whether inflammation and shared genetic liability may be common underlying biological mechanisms for the comorbidity. Using birth cohort data, I show that genetic risk for type 2 diabetes is associated with psychosis-risk in adulthood, and vice versa. I also show that genetic risk for type 2 diabetes may influence psychosis risk by increasing systemic inflammation. Using summary data from large genome-wide association studies (GWAS), I show a thread of evidence for shared genetic overlap between schizophrenia, cardiometabolic and inflammatory traits. Finally, using Mendelian randomization, I show evidence supporting that inflammation may be a common cause for insulin resistance and schizophrenia. On the prediction of cardiometabolic risk in young adults with psychosis, I performed a systematic review of cardiometabolic risk prediction algorithms and explored their predictive performance in a sample of young people at risk of developing psychosis. In doing so, I show that none are likely to be suitable for this population. Then, using patient data, I developed and externally validated the Psychosis Metabolic Risk Calculator (PsyMetRiC), the first cardiometabolic risk prediction algorithm specifically tailored for young people with psychosis. Together, my work suggests that cardiometabolic disorders and schizophrenia share aetiologic mechanisms, namely inflammation and shared genetic liability. I have shown that it is possible to accurately predict cardiometabolic risk in young people with psychosis using a tool tailored for the population. Such tools can in future become valuable resources for clinicians to reduce the risk of long-term cardiometabolic morbidity and mortality in people with schizophrenia.

Background NHS Blood and Transplant (NHSBT), the sole provider of blood to the National Health Service (NHS) in England, UK, must ensure a safe and sustainable supply of blood, whilst protecting donor health and wellbeing. Objectives The primary objective of this thesis is to advance the evidence base for strategies aimed at optimising donor iron status and improving the efficiency and sustainability of the blood supply through reduced deferrals and improved return rates. To achieve this objective, this thesis provides a portfolio of complementary quantitative epidemiological studies, which aimed to: (i) assess temporal trends and contemporary patterns in blood collection and adverse events; (ii) identify individual and collection site characteristics associated with return behaviour; (iii) characterise in detail the cross-sectional associations of the three optimal indices for defining iron deficiency (i.e., haemoglobin, ferritin, and soluble transferrin receptor [sTfR]), particularly modifiable and genetic correlates; (iv) develop and validate risk prediction models for low haemoglobin deferral to determine the added value of different types of data. In addition, the extent to which participants from blood donor research studies are representative of the national donor population will be investigated. This thesis will leverage large-scale, detailed, and multi-purpose data sources to overcome some of the limitations of previous studies, reporting findings to a greater detail than has previously been possible, in order to inform future studies that will directly influence blood donation practice and policy. Data sources Data extracts from the NHSBT national donor database (N=3,400,000) contained anonymised individual-level blood donation records from 2007-19 and were used to assess temporal trends and adverse events and to identify determinants of return behaviour, as well as to define comparator populations for the INTERVAL trial and COMPARE study for the purpose of assessing the representativeness of these cohorts. Data from the INTERVAL trial (2012-16; N=45,000), a large, parallel group, pragmatic, individually randomised trial, embedded in the routine blood service, evaluating the efficiency and safety of varying the frequency of whole blood donation over a 2-year period, was used to complement the analysis of return behaviour through study of adherence to assigned inter-donation intervals, in addition to contributing to the cross-sectional individual-participant data meta-analysis of correlates of iron levels and the development of risk prediction models for low haemoglobin deferral. Data from the COMPARE study (2016-17; N=30,000), a two-stage, systematic, within-person comparison diagnostic-accuracy study of different haemoglobin screening methods, contributed to the meta-analysis and was used to validate risk prediction models. The meta-analysis also included individual-participant data from the Cambridge BioResource (2007-9; N=7,500), Cambridge CardioResource (2010; N=2,500), and UK Blood Services Common Controls (2005; N=3,600). Study participants were well-characterised with extensive genetic, haematological, biochemical, lifestyle, health, and serially-assessed clinical information, in addition to routinely available demographic, donation history, and collection site characteristics, together with geographical and socioeconomic information through linkage to national datasets. Polygenic risk scores (PRSs) were derived for key iron measures using published variants. Results In analyses of over 20 million donor attendances, overall blood collection declined by 24∙1% between 2009 and 2018, but collection trends differed at a blood group level with targeted increases in the proportion of donations comprising universal, rare, and minor blood groups. As a result, the donor population contracted, shifting toward an older and more ethnically diverse population, comprising fewer male donors. Deferral for low haemoglobin increased substantially following the introduction of revised screening procedures in 2017, whilst adverse events remained stable. INTERVAL trial and COMPARE study participants were broadly representative of the national donor population of England, with only minor differences in age, sex, and number of previous donations. In analyses of return behaviour, during a 2-year follow-up period, 67·0% of donors from the national donor comparator population and 93·3% of participants from the INTERVAL trial returned to donate, with male sex, older age, number of previous attendances, and number of opportunities to donate positively associated with return. By contrast, non-white ethnicity, distance to the donation venue, deferral and adverse events at the previous visit were negatively associated with return. Substantial deviations from model assumptions revealed time-dependent associations. To reliably quantify cross-sectional associations of iron measures, individual-participant data on 85,847 whole blood donors from five relevant independent UK studies was harmonised to establish a large national consortium, allowing concomitant assessment of haemoglobin, ferritin, and sTfR. The iron measures shared some important modifiable risk factors, but the overlap was imperfect. In aggregate, only around one sixth of the total variation in haemoglobin and sTfR levels was explained by all of the variables considered, whereas, for ferritin levels, this was approximately one third for men and one quarter for women, of which the majority was attributable to factors describing donation frequency. Modifiable lifestyle characteristics (such as smoking, alcohol consumption, physical activity) and dietary habits together explained only a few percent of the total variation in iron measures. Notably, for haemoglobin, the respective PRS accounted for the largest fraction of the total explained variation. The prevalence of all stages of iron deficiency increased with higher donation frequency and younger age in women. 2∙1% of men and 6∙0% of women were deferred for low haemoglobin at their first donation attempt during the INTERVAL trial. The cross-validated concordance (C)-index of a prediction model for low haemoglobin deferral including conventional risk factors alone was 0·937 (95% CI 0·927 to 0·947) for men and 0·859 (0·847 to 0·871) for women, which could largely be attributed to the previous haemoglobin level. The inclusion of additional data types did not significantly augment the predictive ability of the model. Conclusions Differential trends by blood group highlight the increasing challenge for NHSBT of needing to reduce blood collection capacity to match declining overall demand and blood usage, whilst retaining donors of certain blood groups where demand is not falling. Perpetual large-scale population-based studies of temporal trends and contemporary patterns in blood collection and adverse events are needed to inform donor recruitment and retention strategies, and to prioritise research. Despite the INTERVAL trial and COMPARE study participants being broadly representative of the national donor population of England, the exact degree of generalisability remains somewhat uncertain. Key determinants of return behaviour were identified and point to possible opportunities for targeted retention strategies within the national donor population, as well as providing valuable information for the effective planning and delivery of short-term policies to allow more frequent donation. Evaluation of the correlates and possible determinants of iron measures has advanced understanding of the potential for prevention efforts to reduce the high prevalence of iron deficiency in whole blood donors to further safeguard donor health. Prediction models based on conventional risk factors alone provided accurate and reliable discrimination between those with low and acceptable haemoglobin levels at a subsequent donation attempt. Increased personalisation of blood donation, in the form of tailoring the inter-donation interval to the donor's capacity to give blood safely, may reduce the number of deferrals for low haemoglobin, which in turn may yield improvements in operational efficiency and enhance donor health, satisfaction, and retention.

Risk prediction models have become embedded into the health system. They are used to guide clinical decision making in a variety of settings: risk of death following surgery (should we operate?), diagnostic models for cancer (should we screen?), or the probability of having a clinical event over a certain time period (should we take preventative measures?). Despite clear guidelines on the development and reporting of models, features of models developed for the same purpose often differ. Furthermore, in the field of cardiovascular disease (CVD), risk thresholds for initiating statin therapy vary across England, Scotland, the US and Europe, despite a large body of evidence on when treatment becomes cost effective. This results in uncertainty when using these models to guide treatment for a patient, as using different models or clinical guidelines may result in a different decision for an individual. This thesis focused on identifying sources of uncertainty associated with both parts of this process, generating risk predictions, and making clinical decisions based on these risk predictions. Case studies consider the primary prevention of CVD, which was chosen due to the high incidence of CVD, the saturation of CVD risk prediction models in the literature, and the fierce debate over the last 10 years about the best approach for the primary prevention of CVD. Chapter 3 found the impact of covariate selection on the risks of individuals to be small, apart from a large secular trend. Chapter 4 identified high levels of instability in risk scores when using sample sizes of widely used models, and when derived from recently published sample size formula. Chapter 5 found that the secular trend in CVD (identified in Chapter 3) caused over prediction of risks for patients in the present day and was not driven by increasing statin use. Chapter 6 highlighted that a small number of extra CVD events could be prevented by delaying statin initiation to when patients are at risks higher than 10% (given the high statin discontinuation rates identified in practice). Chapter 7 showed that the reduction of the risk threshold for initiating statins for the primary prevention of CVD, from a 10-year CVD risk of 20% to 10%, had little impact on clinical practice in England. This finding is contrary to current evidence. The findings in this thesis are a mix of methodological findings of interest to those developing models, and those that have a direct impact on the prevention of CVD in the UK.

The only established prostate cancer (PCa) risk factors are age, ethnic ancestry and family history. Several PCa genetic susceptibility variants have been identified, including rare moderate-to-high risk pathogenic variants (PVs) and common low-risk variants which jointly explain 40% of the excess familial risk. No risk prediction model exists that reflects the current knowledge on PCa susceptibility. The main objective of this thesis was to develop a comprehensive prediction model of future PCa risk in unaffected European ancestry men, that includes the effects of known genetic risk variants and explicitly modelled residual family history. Sub-objectives included characterisation of the PCa risks associated with rare HOXB13, BRCA1 and BRCA2 PVs. Through systematic review and meta-analysis of 20 publications, I estimated a pooled PCa relative risk (RR) of 3.42 (95% confidence interval [CI] 2.81-4.17) for HOXB13 G84E carriers. Based on complex segregation analysis of family history data on 11,983 genotyped PCa patients, the most parsimonious model was a multiplicative model for the effect of each HOXB13 G84E copy that assumed separate RRs by birth cohort. The estimated per-allele RRs were 3.09 (95% CI 2.03-4.71) for carriers born ≤1929 and 5.96 (95% CI 4.01-8.88) for carriers born ≥1930. The PCa risks varied by family history. Using a prospective cohort of 823 male BRCA1/2 carriers, I estimated age-specific RRs of 3.99 (95% CI 1.88-8.49) at age < 65 years and 4.64 (95% CI 2.91-7.41) at age≥65 years for BRCA2 carriers, and 3.57 (95% CI 1.68-7.58) at age < 65 years and 1.86 (95% CI 0.96-3.59) at age≥65 years for BRCA1 carriers. The risks varied by age, family history, and PV location within BRCA2. BRCA2 PVs were associated with moderate-to-high-grade PCa. Combining these and previous estimates in a systematic review and meta-analysis of 25 publications yielded pooled RR estimates of 7.04 (95% CI 5.30-9.35) at age < 65 years and 3.84 (95% CI 2.84-5.18) at age≥65 years for non-Ashkenazi Jewish European ancestry BRCA2 carriers and 1.95 (95% CI 1.23-3.09) at age < 65 years and 0.91 (95% CI 0.62-1.33) at age≥65 years for BRCA1 carriers. Using genotypic and family history data on 16,633 PCa patients, I developed a risk prediction model that incorporates the explicit effects of rare PVs in HOXB13, BRCA1 and BRCA2, and a polygenic risk score that summarises the effects of 174 common low-risk variants identified through genome-wide association studies. The residual familial risk is modelled by a hypothetical recessively inherited high-risk variant associated with a RR for homozygotes of 104 (95% CI 21.4-502) and a risk allele frequency of 4.44% (95% CI 2.14%-8.97%), and an age-dependent normally distributed polygenic component with a SD of 2.11 (95% CI 1.98-2.24) at age 70 that decreases at a rate of 0.985 (95% CI 0.981-0.990) per year of age. The predicted familial relative risks are consistent with those reported in large observational studies. Limitations include potential biases related to screening of relatives of affected men and known PV carriers, and use of self-reported family history information. The model has not been externally validated. This is the first risk prediction model that can be used to provide future PCa risk based on age, family history and all known genetic factors. The risk estimates for HOXB13, BRCA1 and BRCA2 carriers and those given by the comprehensive model will facilitate the genetic counselling of men seen in family clinics and may enable future risk-stratified screening approaches.

Breast cancer is the most common cancer among women and incidence is still rising. The majority of breast cancer is sporadic (non-familial), and a small proportion is familial. Only ~50% of high risk families have a genetic diagnosis at the present time. Currently, the majority of women are unaware of their personal breast cancer risk. Therefore better risk prediction is warranted as many women without a family history could have a high risk of breast cancer and more breast cancers could be prevented. In this study, two approaches were taken to improve risk prediction in the general population and in high risk families without a genetic diagnosis. The predictive ability of 18 single nucleotide variants (SNPs) associated with breast cancer risk was tested in a prospective cohort of 9363 women. Incorporation of both SNPs and mammographic density into a classical risk factor prediction model was performed to assess the improvement of risk prediction. Promoter methylation analysis and upstream sequencing of BRCA1 was performed in a high-risk cohort of 49 families with breast/ovarian cancer. Functional analysis of upstream variants was carried out using saturation genome editing and luciferase assays. Incorporation of SNPs and mammographic density into a classical risk factor model identified more women at high (>5%) and low (<2%) was identified in 2 of 49 high risk families, explaining their familial disease. The exact functional effect of this variant however, is not yet fully explained. Risk prediction was successfully improved with the addition of SNPs and mammographic density to the currently used risk prediction models. Similarly, identification of a novel pathogenic upstream methylation-associated BRCA1 variant in high risk breast cancer families allows better risk prediction in these families. Both approaches may lead to more personalized risk prediction and preventive management of breast cancer.

Chronic obstructive pulmonary disease (COPD) is one of the leading causes of death in the world. In addition to premature mortality, the consequent socio-economic burden is high, causing reduced quality of life, loss of productivity, and hospital admission. Diagnosis of COPD relies on lung function tests, which are inadequate and often leave the condition undiagnosed and thus untreated. There is a growing interest in the extra-pulmonary manifestations of COPD and assessing the predictive value of cardiovascular abnormalities, musculoskeletal weakness and plasma biomarkers for acute exacerbation of COPD, hospital admission and mortality, as there is currently no individual biomarker able to reliably identify or predict these common clinical outcomes. The aim of this research was to identify and evaluate the predictive value of existing and novel biomarkers for COPD, and determine if and how these biomarkers can predict the longer-term clinical outcomes using electronic health record data. Electronic databases were systematically searched and identified 61 studies, which were synthesised, including meta-analyses to estimate pooled hazard ratios of the associations between selected biomarkers and common clinical outcomes. Data derived from the Evaluating the Role of Inflammation in Chronic Airways disease (ERICA) study were linked to electronic health record data (i.e. hospital admissions) and survival data. Predictive models for mortality and cardiovascular related hospital admission were developed using stratified multivariable Cox regression, and assessed by C-indices with 10-fold cross-validation. Negative binomial regression was used to model the event rate of acute exacerbation of COPD and determine the risk of hospitalisation due to acute exacerbation of COPD, and the associated length of stay. Data from the UK Biobank were used to explore cause-specific deaths in COPD. Sex-specific all-cause and cause-specific mortality rates were age-standardised using the 2013 European Standard Population. Hazard ratios were estimated using Cox proportional hazards regression, adjusted for age and sex. Systematic review indicated that shorter six-minute walk distance, elevated heart rate, fibrinogen, C-reactive protein, and white cell count were associated with a higher risk of mortality. Shorter six-minute walk distance and elevated fibrinogen and C-reactive protein were associated with COPD exacerbation, and shorter six-minute walk distance and higher heart rate, C-reactive protein and interleukin-6 were associated with hospitalisation. Data from the ERICA cohort indicated no significant difference between the discriminative ability of a BODE Index with six-minute walk and BODE Index with short physical performance battery when predicting mortality. For most musculoskeletal measures, poorer performance was associated with higher rate or longer duration of hospitalised acute exacerbation of COPD. Measures of arterial stiffness and carotid intima-media thickness were not associated with cardiovascular events. Measures of exercise capacity were significantly associated with cardiovascular disease and improved the discriminative ability when added to Framingham risk factors. Data from the UK Biobank indicated COPD was associated with a higher risk of all-cause mortality, and cardiovascular death. In both men and women, COPD had an associated threefold higher risk of early mortality, including a fourfold higher risk of cardiovascular-related death in women, and threefold higher risk of cardiovascular-related death in men. Epidemiological evidence indicates that musculoskeletal measures have the potential to replace the six-minute walk in the BODE Index for predicting mortality in COPD. In addition, physical capacity should be considered as an important treatable trait in reducing risk of hospitalisations for acute exacerbation of COPD. Data from the ERICA cohort does not support the use of objective measures of arterial stiffness and carotid intima-media thickness in addition to Framingham risk factors for predicting cardiovascular events within COPD. Findings in the UK Biobank indicated that COPD is associated with a higher risk of cardiovascular death but cancer and respiratory disease to be the leading causes.

The incidence of degenerative aortic stenosis is increasing with an ageing population. Valve replacement is the only proven treatment, but this carries significant procedural risk in older people. Current guidelines advocate intervention in symptomatic severe aortic stenosis, but non-cardiac symptoms and comorbidity may obscure this assessment. Clinical biomarkers offer the potential for objective patient assessment. My aim was to firstly assess the validity and reproducibility of novel blood biomarkers of disease progression in aortic stenosis. Secondly, in older patients considered for valve replacement, my aim was to compare measures of frailty with conventional surgical risk assessment. In 265 patients with asymptomatic aortic stenosis and 46 healthy controls, I assessed serum concentrations of the sarcomeric protein cardiac myosin binding protein C (cMyC) and objective markers of disease progression and mortality. cMyC concentrations were independently associated with imaging evidence of left ventricular mass, fibrosis volume and extracellular volume. These relationships were not observed in healthy controls. cMyC concentrations were also associated with all-cause mortality over 11 years of follow-up. This suggests a role for cMyC as a novel objective biomarker of aortic stenosis disease severity. Other blood biomarkers including cardiac troponin, brain-type natriuretic peptide (BNP) and galectin-3 have been suggested as disease biomarkers in aortic stenosis. However, performance and precision of these assays has not been described in older patients. In a study of analytical and biological variability, I undertook repeated hourly and weekly blood sampling for cardiac troponin, BNP and galectin-3 in 14 subjects with severe asymptomatic aortic stenosis. These biomarkers demonstrated low indices of individuality, implying that interpretation requires serial testing for change rather than isolated elevated blood concentrations. The reference change values for weekly fresh sampling were 42% for cardiac troponin, 55% for BNP and 14% for galectin. These values for cardiac troponin and BNP were lower than equivalent studies in healthy controls and in stable heart failure. To assess the role of frailty in the assessment of patients for aortic valve replacement, I first performed a systematic review and meta-analysis of studies including frailty assessment before Transcatheter Aortic Valve Implantation (TAVI). This procedure is reserved for patients considered at prohibitive risk of complication from conventional open-heart surgery. Ten cohort studies with 4,592 TAVI patients were included. Frailty was associated with increased risk of early and late mortality, and use of an objective frailty tool rather than subjective assessment identified those at highest risk; these patients experienced greater than double the mortality risk of non-frail individuals. In 185 patients with severe aortic stenosis, I prospectively assessed frailty using four tools: the Fried phenotype, Edmonton Frail Scale, Short Physical Performance Battery and Clinical Frailty Scale. These measures were compared to surgical risk assessment scores from the Society of Thoracic Surgeons (STS) and EuroSCORE II. Agreement between frailty measures was moderate and unrelated to patient age or the degree of aortic valve severity. Frail patients had poorer physical and mental wellbeing. Frailty increased at higher STS and EuroSCORE estimates, but using principal components analysis I demonstrated divergence between frailty measures and surgical risk estimates. Outcomes after aortic valve replacement are now required to establish if this observation is meaningful for the improved prediction of outcomes after surgery. My findings suggest that serial testing of blood biomarkers of myocardial injury in patients with aortic stenosis may detect meaningful disease progression prior to decompensation. In patients considered for valve replacement, measurement of frailty differs from existing surgical risk tools and may add to the holistic assessment of older patients.

Objectives: Around 1000 patients die each year in the UK due to complications suffered following cardiac surgery. Early identification of those at risk of specific complications would allow targeted interventions aimed at reducing the harm caused by those complications. However, commonly used risk models only quantify overall mortality risk for groups of patients based on analyses of pre- and intra-operative data. These models do not predict specific complications and are unable to update risk estimates as postoperative events unfold. This thesis aims to advance understanding of postoperative risk prediction following cardiac surgery by assessing the performance of existing risk prediction tools in this population and developing novel risk models. Methods: Postoperative physiological monitoring data, blood test results, medication administration data and demographics for over 3000 patients were cleaned, analysed using computerised processing algorithms and entered into a comprehensive database. The database was used to validate three mortality models, the Sepsis-3 diagnostic criteria and the KDIGO AKI criteria. A novel dynamic Bayesian model which analyses an individual's urine output to predict their risk of severe oliguria was developed and validated. Finally, the potential usefulness of potassium and magnesium concentrations when predicting atrial fibrillation (AF) was assessed. Results: While the logistic Cardiac Surgery Score (logCASUS), Rapid Clinical Evaluation (RACE) and Sequential Organ Failure Assessment (SOFA) score all discriminated well between survivors and those who died, calibration of the models was inadequate. The Sepsis-3 criteria identified patients at increased risk of adverse outcomes. The KDIGO staging criteria were poorly calibrated, overestimating the risk associated with mild oliguria following cardiac surgery. The Bayesian urine output model discriminated excellently between those who did and did not go on to suffer severe oliguria and was well calibrated. Postoperative potassium and magnesium concentrations were similar for those who did and did not suffer AF. Conclusion: The clinical usefulness of existing risk stratification methods has been assessed and weaknesses identified. It has been demonstrated that serum potassium and magnesium concentrations are unlikely to be useful when predicting AF following cardiac surgery. A novel approach to modelling urine output has been described and validated. The model's performance should be assessed in other settings and then the clinical usefulness of the model could be assessed in clinical trials. The methodology described in this thesis should be replicated to improve postoperative prediction of other complications following cardiac surgery.

Chronic heart failure (HF) is a condition that affects more than 900,000 people in the UK. Mortality rates associated with the condition are high, with nearly 20% of patients dying within one year of diagnosis. Continuous monitoring and risk stratification can help identify patients at risk of deterioration and may consequently improve patients' likelihood of survival. Current repeated-measure risk stratification techniques for HF patients often rely on subjective perception of symptoms, such as breathlessness, and markers of fluid retention in the body (e.g. weight). Despite the common use of such markers, studies have shown that they offer limited effectiveness in predicting HF-related events. This thesis set out to identify and evaluate new markers for repeated-measure risk stratification of HF patients. It started with an exploration of traditional HF measurements, including weight, blood pressure, heart rate and symptom scores, and aimed to improve the performance of these measurements using a data-driven approach. A multi-variate model was developed from data acquired during a randomised controlled trial of remotely-monitored HF patients. The rare occurrence of HF-related adverse events during the trial required the developement of a careful methodology. This methodology helped identify the markers with most predictive ability, which achieved moderate performance at identifying patients at risk of HF-related adverse events, clearly outperforming commonly-used thresholds. Subsequently, this thesis explored the potential value of additional, accelerometer-derived physical activity (PA) and sleep markers. For this purpose, the ability of accelerometer-derived markers to differentiate between individuals with and without HF was evaluated. It was found that markers that summarise the frequency and duration of different PA intensities performed best at differentiating between the two groups and may therefore be most suitable for future use in repeated-measure applications. As part of the analysis of accelerometer-derived HF markers, a gap in the methodology of automated accelerometer processing was identified, namely the need for self-reported sleep-onset and wake-up information. As a result, Chapter 5 of this thesis describes the development and evaluation of a data-driven solution for this problem. In summary, this thesis explored both traditional and new, accelerometer-derived markers for the early detection of HF deterioration. It utilised sound methodology to overcome limitations faced by sparse and unbalanced datasets and filled a methodological gap in the processing of signals from wrist-worn accelerometers.

Cardiovascular disease contributes significantly to global morbidity and mortality and is particularly prevalent among individuals with Type 2 diabetes, which is thought to in part be due to the association between diabetes and the metabolic syndrome. Traditional cardiovascular risk prediction scores perform well in the general population but their use in people with Type 2 diabetes is limited as they are thought to underperform in high risk groups. Indeed, the use of any risk prediction in people with Type 2 diabetes is a point of discussion among clinicians as people with diabetes are thought by some to be at immediate high risk of CVD, whereas others view them as having a degree of modifiable risk which can be addressed using risk prediction. In the general population, novel markers such as cIMT and carotid plaque, as well as other potential biomarkers of cardiovascular risk, have been explored as possible adjuncts to risk scores in the prediction of cardiovascular disease. The evidence for their use in general populations has been established, although there have been no firm conclusions with regard to recommendations for their use, which is partly due to the high degree of variability in cIMT measurement. However, the evidence for their use in people with Type 2 diabetes is sparse, despite the use of such markers as surrogate CV endpoints in clinical trials. This thesis aimed to describe the frequency, distribution and change of cIMT and carotid plaque, as well as to explore the relationship of cIMT and carotid plaque with cardiovascular risk factors, prevalent cardiovascular disease and future cardiovascular events in older people with Type 2 diabetes. The association between cIMT, carotid plaque and other novel risk markers was also explored. The analysis was performed using data from the Edinburgh Type 2 Diabetes Study (ET2DS). This study is a large, prospective cohort study of 1066 men and women with Type 2 diabetes, aged 60-75 years at recruitment, living in Edinburgh and the Lothians. cIMT and carotid plaque were measured at year 1 follow up of the study. Variables concerning cardiovascular risk factors used in this thesis were obtained from the data collection performed at baseline and year 1. A mean of 3.5 years of follow up was available for analysis and is complete for the baseline cohort as data linkage was performed. Mean values of cIMT in the ET2DS were comparable with other studies of cIMT in people with Type 2 diabetes and may indeed be higher than cIMT in the general population. Measurement of cIMT by the sonographer was comparable with computer aided measurements. Increasing cIMT was independently associated (although only modestly) with increasing age, male sex and raised systolic blood pressure. Mean cIMT was associated with prevalent vascular disease and was predictive of incident global cardiovascular events and coronary artery events (but not stroke) over and above UKPDS risk factors, although the clinical impact of this on the reclassification of vascular risk (as demonstrated by net reclassification index (NRI)) was limited. There was a high prevalence of carotid plaque, and in particular “high risk” plaque, in the ET2DS. Different measures of carotid plaque were independently associated with several cardiovascular risk factors. Carotid plaque thickness was independently associated, albeit modestly, with increasing age, male sex, duration of diabetes and hypertension, plaque score with increasing age, hypertension, smoking and low BMI, and high risk plaque with hypertension and low BMI. All measures of carotid plaque were associated with prevalent vascular disease. However, despite these associations, carotid plaque did not have any additional predictive value for incident cardiovascular events over and above UKPDS risk factors. Finally, measures of cIMT and carotid plaque in the ET2DS were associated with the biomarkers ankle brachial index (ABI) and NTproBNP. In addition these markers were significantly higher in those individuals with prevalent vascular disease, suggesting a more extensive exploration of the association of these markers in relation to cardiovascular disease in the ET2DS may be warranted. cIMT and carotid plaque are modestly associated with traditional cardiovascular risk factors and prevalent cardiovascular disease in older adults with Type 2 diabetes. cIMT has been shown to be predictive of incident events while carotid plaque was not, in people with Type 2 diabetes, over and above traditional cardiovascular risk factors, although its impact on risk reclassification may only be small. Further evidence is required from the longer follow up of the ET2DS before firm conclusions can be drawn on the usefulness of cIMT and carotid plaque as risk markers in people with Type 2 diabetes. In addition, large collaborative studies could be used to further explore the relationship of carotid plaque, and change in cIMT with incident cardiovascular events, as well as exploring the additive effect of cIMT and plaque on risk prediction.

Objectives: Cardiovascular disease is the leading cause of mortality and morbidity in the developed world. Surgery can improve prognosis and relieve symptoms. Risk prediction models are increasingly being used to inform clinicians and patients about the risks of surgery, to facilitate clinical decision making and for the risk-adjustment of surgical outcome data. The importance of risk prediction models in cardiovascular surgery has been highlighted by the publication of cardiovascular surgery outcome data and the need for risk-adjustment. The overall objective of this thesis is to advance risk prediction modelling in cardiovascular surgery with a focus on the development of models for elective AAA repair and assessment of models for cardiac surgery. Methods: Three large clinical databases (two elective AAA repair and one cardiac surgery) were utilised. Each database was cleaned prior to analysis. Logistic regression was used to develop both regional and national risk prediction models for mortality following elective AAA repair. A regional model to identify the risk of developing renal failure following elective AAA repair was also developed. The performance of a widely used cardiac surgery risk prediction model (the logistic EuroSCORE) over time was evaluated using a national cardiac database. In addition an updated model version (EuroSCORE II) was validated and both models’ performance in emergency cardiac surgery was evaluated. Results: Regional risk models for mortality following elective AAA repair (VGNW model) and a model to predict post-operative renal failure were developed. Validation of the model for mortality using a national dataset demonstrated good performance compared to other available risk models. To improve generalisability a national model (the BAR score) with better discriminatory ability was developed. In a prospective validation of both models using regional data, the BAR score demonstrated excellent discrimination overall and good discrimination in procedural sub-groups. The EuroSCORE was found to have lost calibration over time due to a fall in observed mortality despite an increase in the predicted mortality of patients undergoing cardiac surgery. The EuroSCORE II demonstrated good performance for contemporary cardiac surgery. Both EuroSCORE models demonstrated inadequate performance for emergency cardiac surgery. Conclusions: Risk prediction models play an important role in cardiovascular surgery. Two accurate risk prediction models for mortality following elective AAA repair have been developed and can be used to risk-adjust surgical outcomes and facilitate clinical decision making. As surgical practice changes over time risk prediction models may lose accuracy which has implications for their application. Cardiac risk models may not be sufficiently accurate for high-risk patient groups such as those undergoing emergency surgery and specific emergency models may be required. Continuing research into new risk factors and model outcomes is needed and risk prediction models may play an increasing role in clinical decision making in the future.

Mammographic density, the amount of radiodense tissue on a mammogram, is a strong risk factor for breast cancer, with properties that could be an asset in screening and prevention programmes. Its use in risk prediction contexts is currently limited, however, mainly due to di culties in measuring and interpreting density. This research investigates rstly, the properties of density as an independent marker of breast cancer risk and secondly, how density should be measured. The rst question was addressed by analysing data from a chemoprevention trial, a trial of hormonal treatment, and a cohort study of women with a family history of breast cancer . Tamoxifen-induced density reduction was observed to be a good predictor of breast cancer risk reduction in high-risk una ected subjects. Density and its changes did not predict risk or treatment outcome in subjects with a primary invasive breast tumour. Finally absolute density predicted risk better than percent density and showed a potential to improve existing risk-prediction models, even in a population at enhanced familial risk of breast cancer. The second part of thesis focuses on density measurement and in particular evaluates two fully-automated volumetric methods, Quantra and Volpara. These two methods are highly correlated and in both cases absolute density (cm3) discriminated cases from controls better than percent density. Finally, we evaluated and compared di erent measurement methods. Our ndings suggested good reliability of the Cumulus and visual assessments. Quantra volumetric estimates appeared negligibly a ected by measurement error, but were less variable than visual bi-dimensional ones, a ecting their ability to discriminate cases from controls. Overall, visual assessments showed the strongest association with breast cancer risk in comparison to computerised methods. Our research supports the hypothesis that density should have a role in personalising screening programs and risk management. Volumetric density measuring methods, though promising, could be improved.

Background: Despite several decades of research, the relevance of body fat and body fat distribution to the risk of cardiovascular disease remains unclear. This thesis aims to investigate associations of body-mass index (BMI), waist circumference (WC) and waist-to-hip ratio (WHR) with risk of first-onset cardiovascular disease under a range of different circumstances. Methods: This thesis used individual records from the Emerging Risk Factors Collaboration to calculate risk ratios, and measures of discrimination and reclassification. 118 prospective studies, involving 1,064,541 participants without known history of cardiovascular disease, had information on BMI at baseline examination. 58 of these studies, involving 221,934 participants, had additional information on waist and hip circumference at baseline examination. Serial measurements made in 42,300 participants from 12 studies with concomitant information on these adiposity measures enabled quantification of within-person variability in BMI, WC and WHR. Results: Cross-sectional analyses demonstrated that although the correlations of adiposity measures differed with one another, BMI, WC and WHR were similarly and importantly associated with mediating cardiovascular risk factors, such as blood pressure, fasting glucose and lipids. Within-person variability was lower in BMI (regression dilution ratio: 0.96) than in WC (0.88) and WHR (0.66). The variability of adiposity measures was not materially influenced by several characteristics, although the variability of WHR varied somewhat by sex, diabetes status and baseline WHR values. 1,064,541 individuals with information on BMI recorded 161,903 deaths or non-fatal cardiovascular outcomes during 15.0 million person-years of follow-up. In analyses adjusted for age, sex and smoking status, BMI had positive and nearly loge-linear associations with coronary heart disease and ischaemic stroke (except at BMI values below 20 kg/m2), which were largely explained by the intermediate risk factors noted above. The association between BMI and non-vascular mortality was curvilinear. Data on 221,934 individuals with complete information on weight, height, and waist and hip circumference (14,297 incident cardiovascular outcomes; 1.87 million person-years of follow-up) demonstrated that BMI, WC and WHR were substantially and similarly related to risk of coronary heart disease and ischaemic stroke. For cardiovascular risk prediction, additional information on BMI, WC or WHR to a prediction model containing conventional risk factors did not importantly improve risk discrimination, nor classification of participants to risk categories of predicted 10-year risk. Conclusions: BMI, WC and WHR are similarly associated with risk of cardiovascular disease, with much of the risk explained by intermediate risk factors. These clinical measures of adiposity do not importantly improve cardiovascular risk prediction when additional information is available on blood pressures, history of diabetes and lipids.

Cardiovascular risk scores predict an individual’s risk of developing cardiovascular disease. Many were developed and validated in study cohorts on risk-factor lowering treatment – a cause of inaccuracy. In addition, risk scores are criticised as being biased towards the elderly due to the prominence of age as a risk predictor. Although present guidelines advocate the use of short-term (5-10 year) absolute risk scores, other approaches to redress this perceived imbalance such as lifetime risk scores are being considered. The overall objective of this thesis is to identify the most appropriate cardiovascular risk score for use in general practice, taking account of the impact of treatment and time on assessed risk. This objective was met by three different methods. First, a systematic review of cardiovascular risk scores was conducted. This explored the derivation of each score, including the extent of treatment. Next, doctors were interviewed in depth to understand their perception and use of risk scores. Finally, mathematical models were devised to determine whether a true difference in life expectancy exists at different ages but the same short-term cardiovascular risk. The models incorporated age-specific case fatality rates, competing risks and time preference to estimate the potential years of life lost due to a five-year treatment delay in different age groups with the same short-term coronary heart disease risk. The findings demonstrate that cardiovascular risk scores do not take account of treatment effects. This significantly affects their application in clinical practice. In addition, there is little difference in potential life years lost between ages at the same risk level because of higher case-fatalities in older people. When time preference is considered, any residual case for treating the same level of short-term risk differently at different ages is abolished. The overall conclusion is that the five to ten-year absolute cardiovascular risk score is the most appropriate approach to primary cardiovascular disease prevention. By overestimating risk in the young, other approaches benefit the few at the expense of the many.

In this dissertation, we propose novel Deep Neural Network (DNN) based statistical learning models that can provide accurate predictions and clear interpretations simultaneously. Chapter 1 presents an introduction to the DNN as a nonparametric approximator to complex, non-linear functions. Chapter 2 introduces the Interpretable Neural Network Regression (INNER), a logistic regression model with nonparametric covariate-dependent coefficients constructed by DNNs. Applied to the individualized risk assessment of preoperative opioid use, the proposed INNER model can predict preoperative opioid use based on the preoperative characteristics and estimate the individual-level odds of opioid use induced by overall body pain, leading to straightforward interpretations of the tendency to use opioids. Applying INNER to Analgesic Outcomes Study (AOS), we identify patient characteristics strongly associated with opioid use. Chapter 3 develops the Penalized Deep Partially Linear Cox Model (Penalized DPLC) that incorporates the SCAD penalty to select significant features and employs the DNN to estimate the nonparametric component of the partially linear Cox model. An efficient alternating optimization algorithm is used for model estimation. We also prove the convergence and asymptotic properties of the estimator. The merits of this method are shown through intensive simulations. Finally, the Penalized DPLC is applied to the National Lung Screening Trial (NLST) to uncover the effects of critical clinical and imaging risk factors on patients' survival. Chapter 4 presents the Deep Survival Learner (DSL) for estimating the Conditional Average Treatment Effects (CATEs) in survival settings. DSL adapts the Doubly-Robust Learner to right-censored data by Inverse Probability of Censoring Weights (IPCW). DNNs are used as base learners to account for the complex relationships between baseline characteristics and survival outcomes. Large-scale simulation experiments are conducted to assess the performance of the proposed model under various scenarios. We then use DSL to study the treatment heterogeneity of perioperative chemotherapy for patients from the Boston Lung Cancer Study (BLCS).

Panel germline tests simultaneously detect multiple pathogenic variants, typically grouped by association with a higher risk of developing certain cancers or syndromes. Next-generation sequencing has made testing for many genes faster and more affordable than ever, but there are nevertheless financial and structural barriers that prevent panel testing from being offered for all patients. Additionally, because clinical guidelines for associations between harmful mutations and heritable diseases are not always well-established, it may not always be beneficial to make panels arbitrarily large. This dissertation explores tools to aid medical decision-making for optimally deploying panel gene testing. Using patient family history and population-level parameter estimates, Mendelian risk prediction models calculate the probability of carrying a pathogenic variant and the future risk of developing cancer. These quantitative risk measures can then be used to develop personalized prevention plans for as further testing or intervention. Existing Mendelian models are widely adopted as part of clinical risk assessment, but are usually limited to specific genes and syndromes. Chapter 1 introduces a generalizable, computationally efficient framework for Mendelian models that incorporates an arbitrary number of genes, cancers, and gene-cancer associations. This more comprehensive approach can flexibly account for new gene-cancer associations as they are discovered, to better reflect the field as it evolves. We rigorously evaluate our approach on simulations and a validation data set. Chapter 2 addresses some of the pragmatic challenges of implementing the framework presented in Chapter 1 by examining Mendelian models that aggregate across genes and cancers. The multi-gene, multi-cancer model proposed in Chapter 1 requires robust estimation of many model parameters, which may be difficult for rare genes and cancers, as well as accurate and detailed patient family history for a large number of cancers. We investigate simplifying modeling assumptions to help overcome uncertainties in parameter estimation and family history reporting. Motivated by the clinical context of pre-screening for a test of any cancer, we apply our aggregate approach to several simulation studies and two high-risk clinical cohorts. In Chapter 3, we continue engaging with the issue of uncertainty in parameter estimation for gene-cancer associations, but switch our attention to the question of determining which genes to include in a panel. We present a utility tool for evaluating the addition and removal of genes in a panel test, based on both quantitative parameter estimates and individualized utility costs. Our approach provides credible intervals to reflect uncertainty in the disease parameter estimates used in its calculation. We illustrate our multi-gene, multi-disease aggregate utility with a hereditary breast cancer panel application and explore of the effects of varying parameter inputs.

Abstract: Thanks to the advancement in cancer treatments, most children diagnosed with cancer are long-term survivors. However, it is estimated that two-thirds of childhood cancer survivors will develop chronic diseases later in life due to their cancer treatments. A reproductive late effect in female cancer survivors is primary ovarian insufficiency (POI), defined as menopause occurring naturally before age 40. Approximately 10% of female survivors of childhood cancer experience POI, far more frequent than the general population (about 1%). Women experiencing POI often face physical and mental health problems and have a decreased quality of life. A consequence of POI is not able to have biological children. Fertility preservation options are available, but these procedures are invasive and expensive. It is difficult to evaluate the need for fertility preservation without knowing the risk of POI. Currently, treatment-related risk factors have been established and used for risk prediction. In recent decades, genome-wide association (GWA) studies have identified genetic variants associated with menopause-related phenotypes. In this study, I evaluated the genetic risk for POI in the form of a polygenic risk score (PRS) and investigated the incremental value of the PRS in the clinical risk prediction model. A total of 1985 participants in the Childhood Cancer Survivor Study (CCSS) original cohort were used in this study. The published GWA studies for age at natural menopause conducted in the general population were used to construct a general population-based PRS (gPRS), while top genetic risk associations (P

INTRODUCTION: In patients who receive treatment for cancer, acute kidney injury (AKI) is arguably one of the most dangerous toxicities that results from cisplatin (CP), a chemotherapeutic agent. While AKI is a common occurrence amongst people who receive cisplatin (CP-AKI), the current risk assessment, intervention methods and understanding the role of magnesium in AKI, are either limited or understudied. OBJECTIVES: We aimed to build on previous CP-AKI risk prediction models, and establish a relationship between serum magnesium levels and AKI. Additionally, we used a feasibility study to test if intravenous magnesium sulfate in patients receiving intraoperative chemotherapy with cisplatin (HIOCC) for malignant mesothelioma can attenuate CP-AKI. This feasibility study was also used to determine a proper dosing regimen to achieve serum magnesium levels of 3 - 4.8 mg/dl. METHODS: We defined acute kidney injury as a 1.5-fold increase in serum creatine, or use of renal replacement therapy (RRT). Using clinical and demographic information from Memorial Sloan Kettering’s database, we conducted multivariable and univariable regression was used to identify the most significant demographic and clinical lab values. Using the information from the statistical analysis we built on previous risk prediction models for cisplatin associated kidney injury. Using the same statistical analysis, we further explored the relationship between serum magnesium values and AKI. In the feasibility study, we recruited patients from Brigham and Woman’s hospital who were receiving HIOCC treatment for mesothelioma. They received an infusion of intravenous magnesium sulfate during surgery. Serum magnesium levels were measured pre-operatively and post-operatively along with serum creatinine values. These were used to obtain pharmacokinetic information to further adjust the infusion rate in patients, lab values were also used to identify any AKI. CONCLUSION: A score-based model created using patient’s age, serum magnesium, albumin, hemoglobin, platelets, cisplatin dose and hypertension is predictive of cisplatin associated acute kidney injury. The feasibility study allowed us to inform phase 2 of an upcoming feasibility study that will include a bolus of 6g Mg/hr and an infusion of 2 g/hr after the bolus. This will work to increase the serum magnesium levels to the therapeutic range.

Evidence suggests that breast cancer, long recognized as a hormonal disease, also has a metabolic component influencing its development. However, etiologic mechanisms of breast cancer, including those related to metabolism, remain elusive given its heterogenous nature. Investigation of metabolomics can provide information on biochemical processes key to breast cancer development. Further, because some breast cancer risk factors differ by menopausal status at diagnosis, evaluating metabolomics and risk factors by menopausal groups is important to capture underlying disease mechanisms. Research into risk factors with opposing associations with pre- vs. postmenopausal breast cancers will allow better understanding of the biologic development of this disease. In this dissertation, I explore potential etiologic mechanisms for pre- and postmenopausal breast cancer through metabolomics and examine differences in risk factor profiles based on menopausal status at diagnosis. In Chapter 1 I took an agnostic approach to examine metabolite associations with breast cancer risk in a nested matched case-control study (N cases=939, N controls=939) in the Nurses’ Health Study (NHS). I uncovered several metabolite groups associated with risk of breast cancer, including cholesteryl esters (inversely associated with risk), triacylglycerols (TAGS) with double bonds (positively associated with risk), and TAGs with ≥ 3 double bonds (inversely associated with risk). I noted changing associations between metabolite groups and breast cancer risk dependent on timing of measurement. In Chapter 2, I took a closer look at adiposity as a risk factor for pre- and postmenopausal breast cancer and evaluated metabolomic profiles for adiposity measures. I found that metabolomic scores were often more informative than self-reported measures for adiposity in determining breast cancer risk, suggesting metabolic dysregulation as a key etiologic component to breast cancer. I also found that metabolite profiles for adiposity measures were similar by menopausal status, though associations with risk were opposite. In Chapter 3, I developed a risk prediction model for premenopausal breast cancer to identify risk factors important to development of this disease. I found that many risk factors identified as significant in primarily postmenopausal cohorts were similarly important for premenopausal breast cancer, including premenopausal duration, ever having a birth, time from menarche to first birth, height, alcohol use, BMI in young adulthood, family history of breast cancer, and history of benign breast disease. Overall, this dissertation suggests that mechanisms of metabolic dysregulation are key components for breast cancer development. Differences seen in risk of pre- vs. postmenopausal breast cancer with respect to certain risk factors, such as adiposity, may be explained by differential action of metabolic dysregulation across the life-course. Discovery that most reproductive and hormonal risk factors for premenopausal breast cancer align with those for postmenopausal breast cancer supports the notion that metabolic dysregulation may be a key element to distinguish the two diseases. Evaluation of the factors of metabolic dysregulation that may be responsible for differences in pre-vs. postmenopausal breast cancer risk can enhance knowledge about etiologic mechanisms for development of breast cancer over time.

Precision medicine approaches have promise to improve the prevention and treatment of cardiovascular disease, which is the leading cause of death in the United States and globally. As the size and number of electronic health record (EHR)-linked biobanks with paired genetic information continue to increase globally, so too do the opportunities for clinical utility of genetic discoveries. My research focuses on the optimal use of rich genetic and phenotypic information from biobanks to translate genetic discoveries to clinical applications. First, I utilize exome sequencing to identify thoracic aortic dissection patients within the Cardiac Health Improvement Project (CHIP) biobank that carry pathogenic genetic changes. Patients with monogenic causes of dissection fit a clinical profile of onset less than 50 years of age with no history of hypertension, and a family history of aortic disease. We conclude that aortic dissection patients in this demographic should be prioritized for clinical genetic testing followed by cascade screening of family members to guide clinical decision-making such as enhanced surveillance of aortic diameter and earlier surgical intervention. Second, I illustrate the promises and challenges of family health history in the context of genetic research studies, with examples from the Trøndelag Health (HUNT) Study and UK Biobank. Individuals who report having a first-degree relative with heart disease have a genetic burden of disease risk alleles intermediate between cases and controls. Family history captures shared genetic and environmental factors, and self-reported family history ascertained in biobank questionnaires is a significant predictor of disease. Self-reported family history remains a significant predictor in the context of polygenic scores, which quantify the genetic risk for disease. Self-reported family history demonstrates some interesting time-varying effects that should be considered. Intuitively, young individuals who likely have younger family members report lower rates of family history of disease, whereas older individuals who have higher rates of positive family history benefit less from preventive interventions. This work motivates biobanks to survey for self-reported family history at multiple time points for a variety of complex diseases. Finally, I examine how polygenic scores improve upon existing risk prediction models used in the clinic, such as the Pooled Cohorts Equation, to aid in earlier identification and treatment of people at high risk. By examining the HUNT Study and the UK Biobank, I systematically compare published polygenic scores for coronary artery disease (CAD) with and without conventional risk factors such as cholesterol, smoking, and hypertension. When the top performing polygenic score for CAD, a metaGRS (Inouye et al, 2018), is added to a model with conventional risk factors, it allows re-classification of 3% of individuals to the high-risk category recommended for therapeutic intervention. Over 10 years, 10.5% of the group re-classified into the high-risk category experienced a CAD event. These are patients who would benefit from implementing preventive lifestyle and medication changes if polygenic scores were added to existing clinical approaches for risk stratification. This dissertation illustrates the use of genetic variation, polygenic scores, and self-reported family history in EHR-linked biobanks with deep phenotyping. I establish criteria for prioritized genetic screening in thoracic aortic dissection, explore the relationship between genetic risk and self-reported family history in complex disease association, and benchmark polygenic scores for better and earlier disease classification. In total, this research aims to harness extensive genetic data for precision medicine approaches that prevent and treat cardiovascular disease.

In many areas of healthcare, clinical prediction models are used to assess disease risk and guide decisions about prevention and treatment. Accurate risk stratification is key to reducing morbidity and mortality through the effective delivery of precision medicine. This dissertation proposes and compares methods for improving the accuracy of risk prediction models through the integration of different models and/or datasets and the adaptation of machine learning algorithms that have achieved high accuracy in other prediction problems. Chapters 1 and 2 focus on cancer risk prediction, while Chapter 3 addresses general settings where multiple studies are available for training and validation. In Chapter 1, we propose to combine existing breast cancer risk prediction models that embed complementary information. Numerous models have been developed, but they often give predictions with conflicting clinical implications. Integrating information from different models can potentially improve the accuracy of risk predictions. BRCAPRO and BCRAT are two widely used models that are based on different risk factors and methodologies. BRCAPRO is a Mendelian model that uses detailed family history information to estimate the probability of carrying a BRCA1/2 mutation, as well as future risk of breast and ovarian cancer, based on mutation prevalence and penetrance (age-specific probability of developing cancer given genotype). BCRAT uses a relative hazard model based on first-degree family history and non-genetic risk factors. We consider two approaches for combining BRCAPRO and BCRAT: 1) modifying the penetrance functions in BRCAPRO using relative hazard estimates from BCRAT, and 2) training an ensemble model that takes as input BRCAPRO and BCRAT predictions. We assess the performance of the combination models in simulations and data from the Cancer Genetics Network, and show that they achieve performance gains over BRCAPRO and BCRAT among individuals with a strong family history of cancer. In Chapter 2, we propose to adapt neural networks for family history-based breast cancer risk prediction. The prevailing models for assessing familial risk of breast cancer are Mendelian models, but these models rely on many assumptions about cancer susceptibility genes. Training more flexible models, such as neural networks, on large datasets can potentially lead to accuracy gains. While there is an extensive literature on neural networks and their state-of-the-art performance in many tasks, there is little work applying them to family history data. The neural network models we propose eliminate the need to explicitly specify the effects of cancer susceptibility genes, overcoming one of the main limitations of Mendelian models. In data simulated under Mendelian inheritance, we demonstrate that neural networks are able to achieve nearly optimal prediction performance. Moreover, when the data generated from a Mendelian model are subject to misreporting of cancer diagnoses, neural networks are able to outperform the Mendelian model. Using a large dataset of over 200,000 family histories from the Risk Service, we train neural networks to predict future risk of breast cancer. We validate them using data from the Cancer Genetics Network and show that they achieve competitive performance with BRCAPRO. In Chapter 3, we compare methods for training prediction models using multiple studies. In precision medicine and other settings, systematic data sharing and data curation initiatives are opening opportunities for developing and validating models on multiple studies, which can lead to improved generalizability. Two general approaches for integrating information across studies are: 1) merging all of the studies and training a single model and 2) multi-study ensembling, which involves training a separate model on each study and combining the resulting predictions. We provide theoretical and empirical analyses comparing the performance of these approaches in the presence of potential heterogeneity in predictor-outcome relationships across studies. In a linear regression setting, we show analytically and confirm via simulations that merging yields lower prediction error than ensembling when the effects of the predictors are relatively homogeneous across studies. However, as heterogeneity increases, there exists a transition point beyond which ensembling outperforms merging. We provide analytic expressions for the transition point in various scenarios, study asymptotic properties, and illustrate how transition point theory can be used for deciding when to merge versus when to ensemble using metagenomic data.

Acute kidney injury (AKI) is the sudden decrease or loss of kidney function caused by direct kidney injury or functional impairment. Many patients do not recover renal function, leading to poor quality of life and high healthcare costs. Previous work has been done to evaluate urinary biomarkers associated with AKI, but these studies have focused on a few proteins with questionable diagnostic ability. Due to the complex pathophysiology of AKI, it may be necessary to create a panel of biomarkers for diagnostic and prognostic assessment of AKI. We explored metabolic biomarkers of AKI in combat casualties using metabolomics. In this study, we used proton nuclear magnetic resonance (1H-NMR) spectroscopy to identify urinary metabolic biomarkers associated with the following outcomes: AKI diagnosis, injury severity score (ISS), AKI stage, or a primary outcome of death or need for renal replacement therapy (RRT).

Abstract: Childhood cancer survivorship has increased drastically over the previous several decades, consequently increasing the frequency of chronic conditions in survivors. Female childhood cancer survivors are at an increased risk of developing nonsurgical premature menopause (NSPM) due to toxicities from their treatment. NSPM occurs when ovarian function is retained for at least 5 years following cancer diagnosis, but menopause develops naturally before age 40. Such a condition can negatively impact quality of life and reduce potential reproductive years. The literature details risk factors including an older age at cancer diagnosis, and treatment with high doses of alkylating agents and radiation. In order to aid physicians, patients and their families have informed discussions regarding fertility preservation, I aimed to develop prediction algorithms of the absolute risk an individual has of developing NSPM. The Childhood Cancer Survivor Study cohort was the primary data source for this project. Due to the presence of both stratified random sampling and participant loss to follow-up within the cohort, I initially investigated methods for combining sampling and censoring weights in the estimation of model accuracy measures to aid in model evaluation. I designed and implemented four simulation studies, varying the relationship between sampling design, censoring distribution and risk score distribution, and assessed weighting scenarios with distinct combinations of censoring and sampling weights. Depending on the study setting, different weighting scenarios gave reasonable estimates, and ignoring or inadequately accounting for weights resulted in biased accuracy estimates.Candidate risk prediction models were developed on a training set of 4,054 observations from the Childhood Cancer Survivor Study cohort using a time-specific logistic regression model with competing risks (TLR-CR), a Fine-Gray regression (FGR) model and a random survival forest model with competing risks (RSF-CR). Model performance and accuracy were measured using the time-specific area under the ROC curve (AUCt), the time-specific average positive predictive value (APt), and calibration curves on both the training set and an internal validation set of 1,454 observations.Model accuracy values and curves were presented for 15 years post cancer diagnosis as an illustration of overall model performance. All three models performed similarly on the training set. The estimated AUCt values decreased when internal validation was conducted; however APt values were still larger than the event rate. The APt / Event Rate ratio for the TLR-CR model increased from the training set performance. AUCt and APt values on the test set calculated over 10-20 years post cancer diagnosis displayed similar findings. The models were well calibrated for low risk patients, however only the TLR-CR model was consistently well calibrated for high risk patients on both datasets. Moving forward, model performance on individuals with clinically verified ovarian status will be assessed through validation on an external cohort. The future practical application of the risk estimates as a risk scoring system aims to have a positive impact on the quality of life of survivors well into their adulthood.

The aim of this study was to describe the clinical and sonographic features of ovarian metastases originating from colorectal cancer (mCRC), and to discriminate mCRC from primary ovarian cancer (OC). We conducted a multi-institutional, retrospective study of consecutive patients with ovarian mCRC who had undergone ultrasound examination using the International Ovarian Tumor Analysis (IOTA) terminology, with the addition of evaluating signs of necrosis and abdominal staging. A control group included patients with primary OC. Clinical and ultrasound data, subjective assessment (SA), and an assessment of different neoplasias in the adnexa (ADNEX) model were evaluated. Fisher’s exact and Student’s t-tests, the area under the receiver–operating characteristic curve (AUC), and classification and regression trees (CART) were used to conduct statistical analyses. In total, 162 patients (81 with OC and 81 with ovarian mCRC) were included. None of the patients with OC had undergone chemotherapy for CRC in the past, compared with 40% of patients with ovarian mCRC (p < 0.001). The ovarian mCRC tumors were significantly larger, a necrosis sign was more frequently present, and tumors had an irregular wall or were fixed less frequently; ascites, omental cake, and carcinomatosis were less common in mCRC than in primary OC. In a subgroup of patients with ovarian mCRC who had not undergone treatment for CRC in anamnesis, tumors were larger, and had fewer papillations and more locules compared with primary OC. The highest AUC for the discrimination of ovarian mCRC from primary OC was for CART (0.768), followed by SA (0.735) and ADNEX calculated with CA-125 (0.680). Ovarian mCRC and primary OC can be distinguished based on patient anamnesis, ultrasound pattern recognition, a proposed decision tree model, and an ADNEX model with CA-125 levels.

Public health practice and quality of medical care rely heavily on the accuracy, precision, and robustness of risk prediction models. Health care providers use risk prediction models to assess a patient’s risk of developing an event during a specified time frame given the patient’s specific characteristics, and subsequently recommend a course of treatment or preventative action. In public health research, risk prediction models are often constructed with common statistical modeling techniques, such as logistic regression for binary outcomes or Cox proportional hazard regression for time-to-event outcomes, and the performance of the model is assessed through internal or external validation, or some combination. Model validation requires statistical and clinical significance and satisfactory baseline or improvement in model calibration and discrimination: calibration quantifies how close predictions are to observed outcomes while discrimination quantifies the model’s ability to distinguish correctly between events and nonevents. Measures for evaluating these qualities include (but are not limited to) Brier score, calibration-in-the-large, proportion of variation (R2), sensitivity and specificity, area under the receiver operating characteristic curve (AUC), discrimination slope, net reclassification index (NRI), integrated discrimination improvement (IDI), and decision theory analytic measures such as net benefit and relative utility. Among these measures exist several interrelationships under certain assumptions, and their estimation and interpretation is an active area of research. The first two parts of this thesis focus on studying the empirical distributions and improving confidence interval (CI) estimation of ∆AUC, NRI, and IDI for both binary event data and time-to-event data. Through data simulation and the comparison of several CI types derived with bootstrapping techniques, we make recommendations for proper estimation in future work and apply our recommendations to real-life Framingham Heart Study data. The third part of this thesis summarizes the many interrelationships and possible redundancies among the measures listed, extends theoretical formulas assuming normal variables for ∆AUC, NRI, and IDI from nested models to non-nested models and to Brier score, and explores the impact of varying discrimination and calibration assumptions on Yates’ and Sanders’ decomposed versions of Brier score through simulation. Lastly, overall conclusions and future directions are presented at the end.

This dissertation offers three studies on criminal behavior and criminal risk forecasting. The first two chapters present theoretical models and empirical evidence on the nature and motives of criminal behavior among high risk and high rate offenders and the implications these models have for criminal justice policy. The third chapter is a study of judicial and administrative decision-making with regard to assessments of crime and violence risk.Chapter 1 gives an account of the 1970s and 80s rise and dominance of the policy of incapacitation through incarceration. The chapter describes the deterministic model of criminal behavior underpinning the theory and policy of incapacitation, and the policy implications of recent re-conceptualizations that view criminal behavior much more like other human choices - a question of contingencies and opportunities. The chapter focuses on Franklin Zimring's (2011) account of the New York City crime decline as evidence against the notion that criminal propensities are fixed and predictable over periods of years. Rather, it appears, relatively modest and superficial changes in circumstances and environmental features can lead to vastly different rates of criminal engagement.The second chapter turns to a particular aspect of the New York City crime decline and the question of future criminal risk prediction. Specifically, the chapter examines the more than two decade drop in the rate of prison return for a new felony among New York City offenders. The chapter assesses whether the declining prison return rate is indeed an indicator of significant behavior change, or is a reflection of changes in criminal justice system actor practices. Further, to the extent that the statistics are an indicator of behavior change, the chapter evaluates whether this can interpreted as the result of the changing New York City crime environment over the last two and a half decades, or is better understood as a reflection of changes in the individual criminal propensities of those leaving prison over this period. To tease apart these competing accounts, the chapter analyzes a unique dataset involving individual records of four cohorts of prisoners from New York City released in the years 1990, 1995, 2000, and 2008. The analysis suggests a mixed picture - all three accounts are at work. Finally, the third chapter of the dissertation turns from models of criminal behavior and criminal decision-making, to judges and administrator decisions regarding criminal risk. The chapter uses machine-learning procedures for prediction and causal estimation to analyze release decisions in California Parole hearings for inmates serving life sentences with the possibility of parole. Using an original dataset generated from all parole suitability hearings conducted since 2011 (over 8,000 transcripts), the chapter offers an empirical analysis of the current system, evaluating the rationality, uniformity, and defensibility of the criteria and decision-making applied to each claim for release. Finally, using the parole analysis as proof of concept, the chapter considers the promises and pitfalls of algorithmic assisted decision-making in the criminal justice system.

Assessment of risk has been a key element in efforts to identify factors associated with disease, to assess potential targets of therapy and enhance disease prevention and treatment. Considerable work has been done to develop methods to identify markers, construct risk prediction models and evaluate such models. This dissertation aims to develop robust approaches for these tasks. In Chapter 1, we present a robust, flexible yet powerful approach to identify genetic variants that are associated with disease risk in genome-wide association studies when some subjects are related. In Chapter 2, we focus on identifying important genes predictive of survival outcome when the number of covariates greatly exceeds the number of observations via a nonparametric transformation model. We propose a rank-based estimator that poses minimal assumptions and develop an efficient

The importance of developing personalized risk prediction estimates has become increasingly evident in recent years. In general, patient populations may be heterogenous and represent a mixture of different unknown subtypes of disease. When the source of this heterogeneity and resulting subtypes of disease are unknown, accurate prediction of survival may be difficult. However, in certain disease settings the onset time of an observable intermediate event may be highly associated with these unknown subtypes of disease and thus may be useful in predicting long term survival. Throughout this dissertation, we examine an approach to incorporate intermediate event information for the prediction of long term survival: the landmark model. In Chapter 1, we use the landmark modeling framework to develop procedures to assess how a patient’s long term survival trajectory may change over time given good intermediate outcome indications along with prognosis based on baseline markers. We propose time-varying accuracy measures to quantify the predictive performance of landmark prediction rules for residual life and provide resampling-based procedures to make inference about such accuracy measures. We illustrate our proposed procedures using a breast cancer dataset. In Chapter 2, we aim to incorporate intermediate event time information for the prediction of survival. We propose a fully non-parametric procedure to incorporate intermediate event information when only a single baseline discrete covariate is available for prediction. When a continuous covariate or multiple covariates are available, we propose to incorporate intermediate event time information using a flexible varying coefficient model. To evaluate the performance of the resulting landmark prediction rule and quantify the information gained by using the intermediate event, we use robust non-parametric procedures. We illustrate these procedures using a dataset of post-dialysis patients with end-stage renal disease. In Chapter 3, we consider improving efficiency by incorporating intermediate event information in a randomized clinical trial setting. We propose a semi-nonparametric two-stage procedure to estimate survival by incorporating intermediate event information observed before the landmark time. In addition, we present a testing procedure using these resulting estimates to test for a difference in survival between two treatment groups. We illustrate these proposed procedures using an AIDS dataset.

Cardiovascular disease (CVD) risk is negatively associated with socioeconomic status (SES) independent of traditional CVD risk factors but clinicians lack the tools to incorporate this SES gradient into their calculation of CVD risk; and policymakers infrequently evaluate the impact of CVD interventions on summary measures of health inequity. Current CVD risk calculators in the United States under-estimate CVD risk among low SES populations and may exacerbate SES inequity in CVD outcomes. The goals of this dissertation were to develop and evaluate a clinically practical equation that improves CVD risk prediction among middle age low SES groups, and to estimate the potential impact of using the equation on the income-related inequity in CVD in a community-based middle-age cohort.Primary analyses were conducted 12,218 participants in the Atherosclerosis Risk in Communities (ARIC) who were free vascular disease and diabetes at baseline and followed for up to 22 years. Geocoded observations were linked to a neighborhood deprivation index calculated from six census tract-level variables of the U.S. Census. Cox proportional hazards and competing risk regression models were used to predict 10 and 20 year risk of coronary heart disease or ischemic stroke (CVD), respectively. Applying methods recommended by the American Heart Association for the evaluation of novel CVD risk factors, the dissertation demonstrates that the addition of an SES-5 variable comprised of education and neighborhood SES is associated with CVD, predicts development of future outcomes, adds incremental value to established risk factors and has demonstrated clinical utility by changing predicted risk sufficiently to modify treatment decisions, in a middle-age cohort free of CVD at baseline in the late 1980s.The analyses also demonstrate how the Framingham risk function systematically under-estimates the socioeconomic gradient in CVD at 10 and 20 years, for multiple CVD outcomes; and that inclusion of an SES-5 variable attenuates this bias. We report results specific to low SES participants with low income or educational attainment, and demonstrate that the SES-5 variable improves reclassification to a greater extent in this low SES group than in the cohort as a whole. We further demonstrated that additional improvement in net reclassification is possible when single risk equations are replaced by a hybrid approach that uses the higher of two predicted risks to classify individuals into treatment categories. The individuals reclassified into higher risk categories experience elevated risk of CVD, and the expected numbers needed to treat with a statin for ten years (NNT) compares favorably with the NNT of traditional approaches. The SES-5 and hybrid approaches would begin to ameliorate CVD inequity by minimizing under-treatment among low SES groups and increasing the proportion of CVD events potentially averted by approximately 7.6% percentage points, representing a relative increase in events averted of 15% to 50%.We further explore the population-wide impact of adding the SES-5 variable to traditional CVD risk equations by calculating observed and predicted values for the concentration index of incident CVD, a summary measure of relative CVD inequity that accounts for differences across the entire spectrum of household income. We demonstrate that the traditional FRS risk function underestimates the observed CI in incident CVD at 10 and 20 years by nearly 50%. In contrast, a risk function that includes the SES-5 variable attenuates this under-estimation, which is no longer significantly different than zero.We also model the impact on health equity and achievement of competing treatment approaches based on risk equations that include or exclude the SES-5 variable. Although all CVD risk-based treatment approaches would improve CVD inequity and inequity-weighted CVD achievement, as measured by the concentration (CI) and achievement (AI) indices, respectively, the SES-5 treatment approach would do so to a greater extent than and approach based exclusively on traditional FRS risk factors.In order to illustrate how social attitudes towards inequity can be incorporated in the evaluation of CVD interventions, we add a parameter reflecting society's aversion to inequity to the equation used to calculate the CI and AI. We calculate the inequity-weighted absolute risk reduction (iARR) for each treatment approach and find that if inequity does not matter then society will be indifferent to the SES-5 and FRS treatment strategies that prevent a similar number of CVD events overall. However, if inequality does matter and if the treatment threshold falls in the range of 5-20% then a risk equation that includes SES-5 should allocate treatment in a manner that is preferable to a strategy based on traditional risk factors alone. A hybrid strategy will avert more CVD events than either the SES-5 and or treatment strategies alone, irrespective of society's aversion to inequity or choice of treatment threshold. Implications for clinical practice and policy are discussed.

Diverse forces have shaped the genomes of individuals throughout the world. It is crucial to understand those historical processes to study the genetics of individuals alive today. Nowhere else is this more important than in the study of admixed populations. A majority of individuals across the Americas are admixed, having received ancestry from sub-Saharan Africans, Europeans, and Indigenous Americans. However, to this day, admixed populations remain understudied, particularly because harnessing all information from their genomes requires in-depth population genetic analysis. This is not typically part of standard practice in genome-wide association studies. In this work I focus on two important aspects of understanding the history of admixed populations of the Americas to identify important associations with medical traits not possible using standard genetic analysis techniques. This work consists of two primary parts:1) I develop a framework for genome-wide admixture mapping meta-analysis from high density SNP genotyping data. I use it to identify a novel, heritable risk factor for asthma in over 7,000 Latinos at the SMAD2 locus that could not be discovered using standard genotype association techniques. I then demonstrate the downstream use of blood-based expression of SMAD2 as a biomarker for both risk of exacerbation and poor response to bronchodilators in people with asthma.2) Along with collaborators I developed the first fine-scale genetic map of indigenous and admixed populations across the country of Mexico, to determine how fine-scale differentiation of indigenous populations impacts the local communities of mixed ancestry. Using novel extensions of Principal Component Analysis we identify striking geographic correlations with the indigenous component of ancestry in admixed individuals, and use this data to identify for the first time a clinically meaningful association between indigenous American origins and lung function.This work also includes an Introduction and Best Practices recommendations on local ancestry estimation and methods for admixture mapping, and an Appendix on generating ancestry informative marker panels. Relevant code and important functions for running admixture mapping and meta-analyzing output will be made publicly available online.

Understanding the relationship between genomic markers and complex disease could have a profound impact on medicine, but the large number of potential markers can make it hard to differentiate true biological signal from noise and false positive associations. A standard approach for relating genetic markers to complex disease is to test each marker for its association with disease outcome by comparing disease cases to healthy controls. It would be cost-effective to use control groups across studies of many different diseases; however, this can be problematic when the controls are genotyped on a platform different from the one used for cases. Since different platforms genotype different SNPs, imputation is needed to provide full genomic coverage, but introduces differential measurement error. In Chapter 1, we consider the effects of this differential error on association tests. We quantify the inflation in Type I Error by comparing two healthy control groups drawn from the same cohort study but genotyped on different platforms, and assess several methods for mitigating this error. Analyzing genomic data one marker at a time can effectively identify associations, but the resulting lists of significant SNPs or differentially expressed genes can be hard to interpret. Integrating prior biological knowledge into risk prediction with such data by grouping genomic features into pathways reduces the dimensionality of the problem and could improve models by making them more biologically grounded and interpretable. The kernel machine framework has been proposed to model pathway effects because it allows nonlinear associations between the genes in a pathway and disease risk. In Chapter 2, we propose kernel machine regression under the accelerated failure time model. We derive a pseudo-score statistic for testing and a risk score for prediction using genes in a single pathway. We propose omnibus procedures that alleviate the need to prespecify the kernel and allow the data to drive the complexity of the resulting model. In Chapter 3, we extend methods for risk prediction using a single pathway to methods for risk prediction model using multiple pathways using a multiple kernel learning approach to select important pathways and efficiently combine information across pathways.

Colorectal cancer (CRC) is the third leading cause of cancer mortality among both men and women in the United States despite prevention efforts. A better tool for predicting colorectal cancer risk may help identify high risk patients who could benefit from chemopreventive medication, but the best statistical method for reducing survival models in large datasets is controvertible.The first project compared several variable selection methodologies in their ability to produce parsimonious, accurate prediction models in large datasets. The second investigation created a model for predicting colorectal cancer risk using a large prospective dataset of almost 200,000 participants. The final project was a cost-effectiveness analysis that explored the potential benefits of sulindac-difluoromethylornithine as an adjunct to colonoscopy for the prevention of CRC. A modified version of Forward Stepwise Regression based on model discrimination (Forward Stepwise C-Statistic) produced slightly more accurate prediction models than traditional forms of stepwise regression based on the F Statistic or Harrell’s Stepdown. New, gender-specific colorectal cancer risk calculators were created using Forward Stepwise C-statistic and achieved bias corrected C-Statistics (95% CI) of 0.69 (0.673-0.698) and 0.68 (0.668-0.694) in men and women, respectively. Age was the overriding risk factor for both men and women. The cost effectiveness analysis revealed that screening colonoscopy is both less expensive and results in more quality adjusted life years than any of the following strategies: no prevention, colonoscopy + chemoprevention, and chemoprevention alone. This finding persisted across one-way sensitivity analyses as well high risk patient scenarios. Forward Stepwise C-Statistic can produce reduced statistical models that are slightly more accurate than traditional stepwise regression or Harrell’s Stepdown, but the slight gain in accuracy may not be worth the additional computational burden in some instances. Head to head comparisons between the new colorectal cancer risk calculator and existing calculators are necessary to prove that the new model is better. The overriding influence of age in the new model suggests that current strategies to screen patients based on age may not be unreasonable. The results demonstrate that chemoprevention with sulindac-difluoromethylornithine should not be considered for patients without a hereditary cancer syndrome.

In this thesis, we argue for the use of loss-based semi-parametric estimation methods as an alternative to traditional parametric models in causal inference and prediction. We present a brief discussion on "black box" epidemiology in the first chapter and argue that risk factor epidemiology can be improved by using semi-parametric estimation methods. We demonstrate the use of semi-parametric methods by applying them to two different problems: one in causal inference and another in prediction. In each case, we demonstrate the process one would follow to define the question of interest, parameterize this question, and estimate it using semi-parametric methods. In the second chapter we introduce a formal concept of a perception effect, and define unmasking and placebo effects in the context of randomized trials. We employ modern tools from causal inference to derive semi-parametric estimators of such effects. The methods are illustrated on a motivating example from a recent pain trial where the occurrence of treatment-related side effects acts as a proxy for unmasking. In the third chapter, we redefine perception and unmasking effects for a longitudinal setting, and explore various causal graphs for the gabapentin trial. We demonstrate application of the semi-parametric methods in this more general setting by assuming a more complicated causal graph. To estimate the parameters, we use Maximum Likelihood Estimation and two different versions of Targeted Maximum Likelihood Estimation. Finally, in chapter four, we approach coronary heart disease risk prediction modeling from a semi-parametric perspective using data from the Framingham study. The "super learner" is used with a library of machine learning algorithms to create an ensemble risk prediction model for coronary heart disease. We define relative risk importance parameters for various risk factors and estimate them with semi-parametric methods used in earlier chapters. The results are compared to the Framingham study and those obtained by fitting a parametric model to the Framingham dataset.

Nearly ten percent of adults in the U.S. aged twenty years and older live with type 2 diabetes. Improving risk stratification and prediction facilitates the identification of those most likely to develop diabetes and informs prevention efforts. The aim of this dissertation was to evaluate the predictive validity of established definitions of high risk across population subgroups, and to examine the impact of modifiable and non-modifiable factors that predict type 2 diabetes risk, presented in three related manuscripts. In the first manuscript, the predictive validity of fasting glucose and glycated hemoglobin A1c (HbA1c) measures were evaluated in the Multi-Ethnic Study of Atherosclerosis (MESA), a community-based observational cohort study that enrolled 6,814 Caucasian Americans, Chinese Americans, African Americans, and Hispanic Americans aged 45-84 years. The association between fasting glucose levels at baseline and cumulative incidence of diabetes did not differ by self-reported ethnic group. However, the strength of association between HbA1c and diabetes incidence differed significantly by ethnicity. The second manuscript examined predictors of incident diabetes among individuals with pre-diabetes in the Coronary Artery Risk Development in Young Adults (CARDIA) study, a community-based longitudinal observational study of 5,115 African Americans and Caucasian Americans aged 18-30 years at baseline. Body mass index was the strongest predictor of diabetes; other predictors included race/ethnicity, diastolic blood pressure, and HDL-cholesterol. A final manuscript explored interactions between single nucleotide polymorphisms (SNPs), the trait of cardiorespiratory fitness, and diabetes-related outcomes in the CARDIA cohort. SNPs of rs8050136 of the fat mass and obesity associated (FTO) gene and rs3856806 of the peroxisome proliferator-activated receptor gamma (PPARG) gene were associated with HbA1c levels; the association between cardiorespiratory fitness level and HbA1c differed by genotype of rs8050136 in Caucasians and rs3856806 in African Americans. This dissertation evaluated the predictive value of current screening tools and markers of risk for type 2 diabetes. Improving diabetes prevention in population subgroups that are disproportionately affected -- particularly those with pre-diabetes and in certain race/ethnic groups -- is dependent upon effective screening tools validated across populations, and a more comprehensive understanding of the role of modifiable and non-modifiable risk factors in influencing one's diabetes risk profile.

WN virus has spread for over 60 years creating endemic and epidemic areas throughout Africa, Asia, and Europe, affecting human, bird, and equine populations. Its 1999 appearance in New York shows the ability of the virus to cross barriers and travel great distances, emerging into new territories previously free of infection. Spreading much faster than expected, WN virus has infected thousands of birds, equine, and humans throughout the conterminous United States (US). Case and serological studies performed in the Eastern hemisphere prior to 1999 offer detailed descriptions of endemic and epidemic locations in regards to geography, land cover, land use, population, climate, and weather patterns. Based on the severity of WN activity within each study area, the patterns associated with these environmental factors allow for the identification of values associated with different levels of risk. We can then model the landscape of the disease within the US and identify areas of high risk for infection. State and county public health officials can use this model as a decision-making tool to allocate funding for disease prevention and control. Dynamic factors associated with increased transmission, such as above average temperature and precipitation, can be closely monitored and measures of prevention can be implemented when necessary. In turn, detailed information from higher resolution analyses can be documented to an online GIS (Geographic Information System) that would contribute to a global collaboration on outbreaks and prevention of disease.

Integrating clinical and genetic information to improve clinicians’ ability to estimate an individual’s disease risk is an important biomedical research challenge. This dissertation develops a “risk index” procedure that combines clinical data and genome-wide genotypes to make predictions about individuals’ risk of disease. For a set of 100 simulated datasets containing 1,000 individuals, 8 clinical covariates, 500 Single nucleotide polymorphisms, and an outcome prevalence of 30%, the average area under the receiver operating characteristics (ROC) curve (AUC) for a risk index model built with clinical covariates and SNPs was significantly higher than a model built with clinical covariates alone (0.846 vs. 0.832, p=0.0002). A risk index model that includes the principal components that account for 90% of the variability in the SNPs also significantly increased the average AUC compared to a clinical covariates only model (0.839 vs. 0.826, p=0.008). For a set of 25 simulated datasets containing 10,000 individuals, 29 clinical covariates, and 38,835 SNPs, a significant difference in average AUC was observed between clinical and clinical+genotype models (0.939 vs. 0.926, p=0.001), using the 500 SNPs most highly associated with the outcome. A risk index model including the 500 largest principal components of the 38,835 SNPs did not significantly increase the mean AUC beyond the clinical model (0.931 vs. 0.931, p=0.98). The risk index methodology was then applied to individuals from the Framingham Heart Study using 27 clinical covariates and 48,071 SNPs. Clinical+genotype risk index models built to predict ten-year incident hypertension, ten-year incident diabetes, or prevalent hypertension had AUCs of 0.475, 0.682, and 0.692, respectively, using the 500 SNPs most highly associated with the outcome, and AUCs of 0.563, 0.782, and 0.712, respectively, using the 500 largest principal components of the SNPs. The results from these analyses suggest that the risk index methodology has utility for predicting an individual’s risk of developing a chronic disease, and that the use of principal components of a large set of SNPs in place of a smaller selected set of associated SNPs provides the best predictive performance.