Search

Abstract Increased left atrial volume and decreased left atrial function have long been associated with atrial fibrillation. The availability of large-scale cardiac magnetic resonance imaging data paired with genetic data provides a unique opportunity to assess the genetic contributions to left atrial structure and function, and understand their relationship with risk for atrial fibrillation. Here, we use deep learning and surface reconstruction models to measure left atrial minimum volume, maximum volume, stroke volume, and emptying fraction in 40,558 UK Biobank participants. In a genome-wide association study of 35,049 participants without pre-existing cardiovascular disease, we identify 20 common genetic loci associated with left atrial structure and function. We find that polygenic contributions to increased left atrial volume are associated with atrial fibrillation and its downstream consequences, including stroke. Through Mendelian randomization, we find evidence supporting a causal role for left atrial enlargement and dysfunction on atrial fibrillation risk.

Abstract A fundamental challenge in diagnostics is integrating multiple modalities to develop a joint characterization of physiological state. Using the heart as a model system, we develop a cross-modal autoencoder framework for integrating distinct data modalities and constructing a holistic representation of cardiovascular state. In particular, we use our framework to construct such cross-modal representations from cardiac magnetic resonance images (MRIs), containing structural information, and electrocardiograms (ECGs), containing myoelectric information. We leverage the learned cross-modal representation to (1) improve phenotype prediction from a single, accessible phenotype such as ECGs; (2) enable imputation of hard-to-acquire cardiac MRIs from easy-to-acquire ECGs; and (3) develop a framework for performing genome-wide association studies in an unsupervised manner. Our results systematically integrate distinct diagnostic modalities into a common representation that better characterizes physiologic state.

Different location of adipose tissue may have different consequences to cardiometabolic risk. Here the authors report that deep learning enabled accurate prediction of specific adipose tissue volumes, and that after adjustment for BMI, visceral adiposity was associated with increased risk of cardiometabolic disease, while gluteofemoral adiposity was associated with reduced risk.

Abstract Prediction models are commonly used to estimate risk for cardiovascular diseases, to inform diagnosis and management. However, performance may vary substantially across relevant subgroups of the population. Here we investigated heterogeneity of accuracy and fairness metrics across a variety of subgroups for risk prediction of two common diseases: atrial fibrillation (AF) and atherosclerotic cardiovascular disease (ASCVD). We calculated the Cohorts for Heart and Aging in Genomic Epidemiology Atrial Fibrillation (CHARGE-AF) score for AF and the Pooled Cohort Equations (PCE) score for ASCVD in three large datasets: Explorys Life Sciences Dataset (Explorys, n = 21,809,334), Mass General Brigham (MGB, n = 520,868), and the UK Biobank (UKBB, n = 502,521). Our results demonstrate important performance heterogeneity across subpopulations defined by age, sex, and presence of preexisting disease, with fairly consistent patterns across both scores. For example, using CHARGE-AF, discrimination declined with increasing age, with a concordance index of 0.72 [95% CI 0.72–0.73] for the youngest (45–54 years) subgroup to 0.57 [0.56–0.58] for the oldest (85–90 years) subgroup in Explorys. Even though sex is not included in CHARGE-AF, the statistical parity difference (i.e., likelihood of being classified as high risk) was considerable between males and females within the 65–74 years subgroup with a value of − 0.33 [95% CI − 0.33 to − 0.33]. We also observed weak discrimination (i.e.,

Abstract Inter-individual variation in fat distribution is increasingly recognized as clinically important but is not routinely assessed in clinical practice, in part because medical imaging has not been practical to deploy at scale for this task. Here, we report a deep learning model trained on an individual’s body shape outline—or “silhouette” —that enables accurate estimation of specific fat depots of interest, including visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) adipose tissue volumes, and VAT/ASAT ratio. Two-dimensional coronal and sagittal silhouettes are constructed from whole-body magnetic resonance images in 40,032 participants of the UK Biobank and used as inputs for a convolutional neural network to predict each of these quantities. Mean age of the study participants is 65 years and 51% are female. A cross-validated deep learning model trained on silhouettes enables accurate estimation of VAT, ASAT, and GFAT volumes (R 2: 0.88, 0.93, and 0.93, respectively), outperforming a comparator model combining anthropometric and bioimpedance measures (ΔR 2 = 0.05–0.13). Next, we study VAT/ASAT ratio, a nearly body-mass index (BMI)—and waist circumference-independent marker of metabolically unhealthy fat distribution. While the comparator model poorly predicts VAT/ASAT ratio (R 2: 0.17–0.26), a silhouette-based model enables significant improvement (R 2: 0.50–0.55). Increased silhouette-predicted VAT/ASAT ratio is associated with increased risk of prevalent and incident type 2 diabetes and coronary artery disease independent of BMI and waist circumference. These results demonstrate that body silhouette images can estimate important measures of fat distribution, laying the scientific foundation for scalable population-based assessment.

The inherited basis of body fat distribution is not fully understood. Here, the authors use genetic data and MRI-derived measures of local adiposity to highlight fat depot-specific genetic architecture with implications for cardiometabolic health.

Presence of islet autoantibodies precedes the onset of type 1 diabetes but it does not predict whether and how fast symptomatic disease appears. Here authors present a model to predict and visualize progression to diabetes by using a large longitudinal data set on autoantibodies and clinical parameters as input.

Abstract Background Polygenic scores—which quantify inherited risk by integrating information from many common sites of DNA variation—may enable a tailored approach to clinical medicine. However, alongside considerable enthusiasm, we and others have highlighted a lack of standardized approaches for score disclosure. Here, we review the landscape of polygenic score reporting and describe a generalizable approach for development of a polygenic score disclosure tool for coronary artery disease. Methods We assembled a working group of clinicians, geneticists, data visualization specialists, and software developers. The group reviewed existing polygenic score reports and then designed a two-page mock report for coronary artery disease. We then conducted a qualitative user-experience study with this report using an interview guide focused on comprehension, experience, and attitudes. Interviews were transcribed and analyzed for themes identification to inform report revision. Results Review of nine existing polygenic score reports from commercial and academic groups demonstrated significant heterogeneity, reinforcing the need for additional efforts to study and standardize score disclosure. Using a newly developed mock score report, we conducted interviews with ten adult individuals (50% females, 70% without prior genetic testing experience, age range 20–70 years) recruited via an online platform. We identified three themes from interviews: (1) visual elements, such as color and simple graphics, enable participants to interpret, relate to, and contextualize their polygenic score, (2) word-based descriptions of risk and polygenic scores presented as percentiles were the best recognized and understood, (3) participants had varying levels of interest in understanding complex genomic information and therefore would benefit from additional resources that can adapt to their individual needs in real time. In response to user feedback, colors used for communicating risk were modified to minimize unintended color associations and odds ratios were removed. All 10 participants expressed interest in receiving a polygenic score report based on their personal genomic information. Conclusions Our findings describe a generalizable approach to develop a polygenic score report understandable by potential patients. Although additional studies are needed across a wider spectrum of patient populations, these results are likely to inform ongoing efforts related to polygenic score disclosure within clinical practice.

Abstract To support point-of-care decision making by presenting outcomes of past treatment choices for cohorts of similar patients based on observational data from electronic health records (EHRs), a machine-learning precision cohort treatment option (PCTO) workflow consisting of (1) data extraction, (2) similarity model training, (3) precision cohort identification, and (4) treatment options analysis was developed. The similarity model is used to dynamically create a cohort of similar patients, to inform clinical decisions about an individual patient. The workflow was implemented using EHR data from a large health care provider for three different highly prevalent chronic diseases: hypertension (HTN), type 2 diabetes mellitus (T2DM), and hyperlipidemia (HL). A retrospective analysis demonstrated that treatment options with better outcomes were available for a majority of cases (75%, 74%, 85% for HTN, T2DM, HL, respectively). The models for HTN and T2DM were deployed in a pilot study with primary care physicians using it during clinic visits. A novel data-analytic workflow was developed to create patient-similarity models that dynamically generate personalized treatment insights at the point-of-care. By leveraging both knowledge-driven treatment guidelines and data-driven EHR data, physicians can incorporate real-world evidence in their medical decision-making process when considering treatment options for individual patients.

Summary: Rapid advances in artificial intelligence (AI) and availability of biological, medical, and healthcare data have enabled the development of a wide variety of models. Significant success has been achieved in a wide range of fields, such as genomics, protein folding, disease diagnosis, imaging, and clinical tasks. Although widely used, the inherent opacity of deep AI models has brought criticism from the research field and little adoption in clinical practice. Concurrently, there has been a significant amount of research focused on making such methods more interpretable, reviewed here, but inherent critiques of such explainability in AI (XAI), its requirements, and concerns with fairness/robustness have hampered their real-world adoption. We here discuss how user-driven XAI can be made more useful for different healthcare stakeholders through the definition of three key personas—data scientists, clinical researchers, and clinicians—and present an overview of how different XAI approaches can address their needs. For illustration, we also walk through several research and clinical examples that take advantage of XAI open-source tools, including those that help enhance the explanation of the results through visualization. This perspective thus aims to provide a guidance tool for developing explainability solutions for healthcare by empowering both subject matter experts, providing them with a survey of available tools, and explainability developers, by providing examples of how such methods can influence in practice adoption of solutions. The bigger picture: Adoption of AI tools in practical settings, such as for research/clinical tasks, has been hampered by a lack of transparency/interpretability of the models. After performing a review of different types of AI explainability (XAI) methods developed to better understand the predictions made by a model, we also develop a taxonomy to better classify the different approaches. We think that these XAI techniques are not sufficient to enhance practical implementations and illustrate via an example how user-driven XAI can be useful for different stakeholders in the healthcare domain. We identify and define three key personas involved in healthcare—data scientists, clinical researchers, and clinicians—and present an overview of the different approaches that can address their needs. The ultimate goal of adopting AI in medical practice and patient care goes beyond explainability and will need the development of extra layers of security and confidence, in particular regarding AI trustworthiness, as XAI transparent systems become prone to attacks that may reveal confidential information, and AI fairness, as systems developed and tested in diverse environments need to be expanded to real-world situations.

Genetic variation predisposes to disease via monogenic and polygenic risk variants. Here, the authors assess the interplay between these types of variation on disease penetrance in 80,928 individuals. In carriers of monogenic variants, they show that disease risk is a gradient influenced by polygenic background.

Abstract The ability to identify patients who are likely to have an adverse outcome is an essential component of good clinical care. Therefore, predictive risk stratification models play an important role in clinical decision making. Determining whether a given predictive model is suitable for clinical use usually involves evaluating the model’s performance on large patient datasets using standard statistical measures of success (e.g., accuracy, discriminatory ability). However, as these metrics correspond to averages over patients who have a range of different characteristics, it is difficult to discern whether an individual prediction on a given patient should be trusted using these measures alone. In this paper, we introduce a new method for identifying patient subgroups where a predictive model is expected to be poor, thereby highlighting when a given prediction is misleading and should not be trusted. The resulting “unreliability score” can be computed for any clinical risk model and is suitable in the setting of large class imbalance, a situation often encountered in healthcare settings. Using data from more than 40,000 patients in the Global Registry of Acute Coronary Events (GRACE), we demonstrate that patients with high unreliability scores form a subgroup in which the predictive model has both decreased accuracy and decreased discriminatory ability.

BACKGROUND & AIMS:A common limiting factor in the throughput of gastrointestinal endoscopy units is the availability of space for patients to recover post-procedure. This study sought to identify predictors of abnormally long recovery time after colonoscopy performed with procedural sedation. In clinical research, this type of study would be performed using only one regression modeling approach. A goal of this study was to apply various "machine learning" techniques to see if better prediction could be achieved. METHODS:Procedural data for 31,442 colonoscopies performed on 29,905 adult patients at Massachusetts General Hospital from 2011 to 2015 were analyzed to identify potential predictors of long recovery times. These data included the identities of hospital personnel, and the initial statistical analysis focused on the impact of these personnel on recovery time via multivariate logistic regression. Secondary analyses included more information on patient vitals both to identify secondary predictors and to predict long recoveries using more complex techniques. RESULTS:In univariate analysis, the endoscopist, procedure room nurse, recovery room nurse, and surgical technician all showed a statistically significant relationship to long recovery times, with p-value below 0.0001 in all cases. In the multivariate logistic regression, the most significant predictor of a long recovery time was the identity of the recovery room nurse, with the endoscopist also showing a statistically significant relationship with a weaker effect. Complex techniques led to a negligible improvement over simple techniques in prediction of long recovery periods. CONCLUSION:The hospital personnel involved in performing a colonoscopy show a strong association with the likelihood of a patient spending an abnormally long time recovering from the procedure, with the most pronounced effect for the nurse in the recovery room. The application of more advanced approaches to improve prediction in this clinical data set only yielded modest improvements.

BACKGROUND AND AIMS:Accurate assessment of the risk of mortality following a cirrhosis-related admission can enable health-care providers to identify high-risk patients and modify treatment plans to decrease the risk of mortality. METHODS:We developed a post-discharge mortality prediction model for patients with a cirrhosis-related admission using a population of 314,292 patients who received care either at Massachusetts General Hospital (MGH) or Brigham and Women's Hospital (BWH) between 1992 and 2010. We extracted 68 variables from the electronic medical records (EMRs), including demographics, laboratory values, diagnosis codes, and medications. We then used a regularized logistic regression to select the most informative variables and created a risk score that comprises the selected variables. To evaluate the potential for generalizability of our score, we applied it on all cirrhosis-related admissions between 2010 and 2015 at an independent EMR data source of more than 18 million patients, pooled from different health-care systems with EMRs. We calculated the areas under the receiver operating characteristic curves (AUROCs) to assess prediction performance. RESULTS:We identified 4,781 cirrhosis-related admissions at MGH/BWH hospitals, of which 778 resulted in death within 90 days of discharge. Nine variables were the most effective predictors for 90-day mortality, and these included all MELD-Na's components, as well as albumin, total cholesterol, white blood cell count, age, and length of stay. Applying our nine-variable risk score (denoted as "MELD-Plus") resulted in an improvement over MELD and MELD-Na scores in several prediction models. On the MGH/BWH 90-day model, MELD-Plus improved the performance of MELD-Na by 11.4% (0.78 [95% CI, 0.75-0.81] versus 0.70 [95% CI, 0.66-0.73]). In the MGH/BWH approximate 1-year model, MELD-Plus improved the performance of MELD-Na by 8.3% (0.78 [95% CI, 0.76-0.79] versus 0.72 [95% CI, 0.71-0.73]). Performance improvement was similar when the novel MELD-Plus risk score was applied to an independent database; when considering 24,042 cirrhosis-related admissions, MELD-Plus improved the performance of MELD-Na by 16.9% (0.69 [95% CI, 0.69-0.70] versus 0.59 [95% CI, 0.58-0.60]). CONCLUSIONS:We developed a new risk score, MELD-Plus that accurately stratifies the short-term mortality of patients with established cirrhosis, following a hospital admission. Our findings demonstrate that using a small set of easily accessible structured variables can help identify novel predictors of outcomes in cirrhosis patients and improve the performance of widely used traditional risk scores.