Search

AIMS: Heart failure (HF) is a clinical syndrome with no definitive diagnostic tests. HF registries are often based on manual reviews of medical records of hospitalized HF patients identified using International Classification of Diseases (ICD) codes. However, most HF patients are not hospitalized, and manual review of big electronic health record (EHR) data is not practical. The US Department of Veterans Affairs (VA) has the largest integrated healthcare system in the nation, and an estimated 1.5 million patients have ICD codes for HF (HF ICD-code universe) in their VA EHR. The objective of our study was to develop artificial intelligence (AI) models to phenotype HF in these patients. METHODS AND RESULTS: The model development cohort (n = 20 000: training, 16 000; validation 2000; testing, 2000) included 10 000 patients with HF and 10 000 without HF who were matched by age, sex, race, inpatient/outpatient status, hospital, and encounter date (within 60 days). HF status was ascertained by manual chart reviews in VAs External Peer Review Program for HF (EPRP-HF) and non-HF status was ascertained by the absence of ICD codes for HF in VA EHR. Two clinicians annotated 1000 random snippets with HF-related keywords and labelled 436 as HF, which was then used to train and test a natural language processing (NLP) model to classify HF (positive predictive value or PPV, 0.81; sensitivity, 0.77). A machine learning (ML) model using linear support vector machine architecture was trained and tested to classify HF using EPRP-HF as cases (PPV, 0.86; sensitivity, 0.86). From the HF ICD-code universe, we randomly selected 200 patients (gold standard cohort) and two clinicians manually adjudicated HF (gold standard HF) in 145 of those patients by chart reviews. We calculated NLP, ML, and NLP + ML scores and used weighted F scores to derive their optimal threshold values for HF classification, which resulted in PPVs of 0.83, 0.77, and 0.85 and sensitivities of 0.86, 0.88, and 0.83, respectively. HF patients classified by the NLP + ML model were characteristically and prognostically similar to those with gold standard HF. All three models performed better than ICD code approaches: one principal hospital discharge diagnosis code for HF (PPV, 0.97; sensitivity, 0.21) or two primary outpatient encounter diagnosis codes for HF (PPV, 0.88; sensitivity, 0.54). CONCLUSIONS: These findings suggest that NLP and ML models are efficient AI tools to phenotype HF in big EHR data to create contemporary HF registries for clinical studies of effectiveness, quality improvement, and hypothesis generation.

Artificial intelligence (AI), the computerized capability of doing tasks, which until recently was thought to be the exclusive domain of human intelligence, has demonstrated great strides in the past decade. The abilities to play games, provide piloting for an automobile, and respond to spoken language are remarkable successes. How are the challenges and opportunities of medicine different from these challenges and how can we best apply these data-driven techniques to patient care and outcomes? A New England Journal of Medicine paper published in 1980 suggested that more well-defined “specialized” tasks of medical care were more amenable to computer assistance, while the breadth of approach required for defining a problem and narrowing down the problem space was less so, and perhaps, unachievable. On the other hand, one can argue that the modern version of AI, which uses data-driven approaches, will be the most useful in tackling tasks such as outcome prediction that are often difficult for clinicians and patients. The ability today to collect large volumes of data about a single individual (eg, through a wearable device) and the accumulation of large datasets about multiple persons receiving medical care has the potential to apply to the care of individuals. As these techniques of analysis, enumeration, aggregation, and presentation are brought to bear in medicine, the question arises as to their utility and applicability in that domain. Early efforts in decision support were found to be helpful; as the systems proliferated, later experiences have shown difficulties such as alert fatigue and physician burnout becoming more prevalent. Will something similar arise from data-driven predictions? Will empowering patients by equipping them with information gained from data analysis help? Patients, providers, technology, and policymakers each have a role to play in the development and utilization of AI in medicine. Some of the challenges, opportunities, and tradeoffs implicit here are presented as a dialog between a clinician (SJN) and an informatician (QZT).

BackgroundWhile only one drug is known to slow the progress of amyotrophic lateral sclerosis (ALS), numerous drugs can be used to treat its symptoms. However, very few randomized controlled trials have assessed the efficacy, safety, and side effects of these drugs. Due to this lack of randomized controlled trials, consensus among clinicians on how to treat the wide range of ALS symptoms and the efficacy of these treatments is low. Given the lack of clinical trials data, the wide range of reported symptoms, and the low consensus among clinicians on how to treat those symptoms, data on the prevalence and efficacy of treatments from a patient’s perspective could help advance the understanding of the symptomatic treatment of ALS. ObjectiveTo compare clinicians’ and patients’ perspectives on the symptomatic treatment of ALS by comparing data from a traditional survey study of clinicians with data from a patient social network. MethodsWe used a survey of clinicians’ perceptions by Forshew and Bromberg as our primary data source and adjusted the data from PatientsLikeMe to allow for comparisons. We first extracted the 14 symptoms and associated top four treatments listed by Forshew and Bromberg. We then searched the PatientsLikeMe database for the same symptom–treatment pairs. The PatientsLikeMe data are structured and thus no preprocessing of the data was required. ResultsAfter we eliminated pairs with a small sample, 15 symptom–treatment pairs remained. All treatments identified as useful were prescription drugs. We found similarities and discrepancies between clinicians’ and patients’ perceptions of treatment prevalence and efficacy. In 7 of the 15 pairs, the differences between the two groups were above 10%. In 3 pairs the differences were above 20%. Lorazepam to treat anxiety and quinine to treat muscle cramps were among the symptom–treatment pairs with high concordance between clinicians’ and patients’ perceptions. Conversely, amitriptyline to treat labile emotional effect and oxybutynin to treat urinary urgency displayed low agreement between clinicians and patients. ConclusionsAssessing and comparing the efficacy of the symptomatic treatment of a complex and rare disease such as ALS is not easy and needs to take both clinicians’ and patients’ perspectives into consideration. Drawing a reliable profile of treatment efficacy requires taking into consideration many interacting aspects (eg, disease stage and severity of symptoms) that were not covered in the present study. Nevertheless, pilot studies such as this one can pave the way for more robust studies by helping researchers anticipate and compensate for limitations in their data sources and study design.

BackgroundConsumer health vocabularies (CHVs) have been developed to aid consumer health informatics applications. This purpose is best served if the vocabulary evolves with consumers’ language. ObjectiveOur objective was to create a computer assisted update (CAU) system that works with live corpora to identify new candidate terms for inclusion in the open access and collaborative (OAC) CHV. MethodsThe CAU system consisted of three main parts: a Web crawler and an HTML parser, a candidate term filter that utilizes natural language processing tools including term recognition methods, and a human review interface. In evaluation, the CAU system was applied to the health-related social network website PatientsLikeMe.com. The system’s utility was assessed by comparing the candidate term list it generated to a list of valid terms hand extracted from the text of the crawled webpages. ResultsThe CAU system identified 88,994 unique terms 1- to 7-grams (“n-grams” are n consecutive words within a sentence) in 300 crawled PatientsLikeMe.com webpages. The manual review of the crawled webpages identified 651 valid terms not yet included in the OAC CHV or the Unified Medical Language System (UMLS) Metathesaurus, a collection of vocabularies amalgamated to form an ontology of medical terms, (ie, 1 valid term per 136.7 candidate n-grams). The term filter selected 774 candidate terms, of which 237 were valid terms, that is, 1 valid term among every 3 or 4 candidates reviewed. ConclusionThe CAU system is effective for generating a list of candidate terms for human review during CHV development.

Background Inpatient discharge instructions are a mandatory requirement of the Centers for Medicare and Medicaid Services and Joint Commission on Accreditation of Healthcare Organizations. The instructions include all the information relevant to post-discharge patient care. Prior studies show that patients often cannot fully understand or remember all the instructions. To address this issue, we have previously conducted a pilot study in which pictographs were created through a participatory design process to facilitate the comprehension and recall of discharge instructions. Objective The main objective of this study was to verify the individual effectiveness of pictographs created through a participatory design process. Methods In this study, we included 20 pictographs developed by our group and 20 pictographs developed by the Robert Wood Johnson Foundation as a reference baseline for pictographic recognition. To assess whether the participants could recognize the meaning of the pictographs, we designed an asymmetrical pictograph–text label-linking test. Data collection lasted for 7 days after the email invitation. A total of 44 people accessed the survey site. We excluded 7 participants who completed less than 50% of the survey. A total of 719 answers from 37 participants were analyzed. ResultsThe analysis showed that the participants recognized the pictographs developed in-house significantly better than those included in the study as a baseline (P< .001). This trend was true regardless of the participant’s gender, age, and education level. The results also revealed that there is a large variance in the quality of the pictographs developed using the same design process—the recognition rate ranged from below 50% to above 90%. Conclusions This study confirmed that the majority of the pictographs developed in a participatory design process involving a small number of nurses and consumers were recognizable by a larger number of consumers. The variance in recognition rates suggests that pictographs should be assessed individually before being evaluated within the context of an application.

Introduction: Presence of a firearm is associated with increased risk of violence and suicide. United States military veterans are at disproportionate risk of suicide. Routine healthcare provider screening of firearm access may prompt counseling on safe storage and handling of firearms. The objective of this study was to determine the frequency with which Veterans Health Administration (VHA) healthcare providers document firearm access in electronic health record (EHR) clinical notes, and whether this varied by patient characteristics.Methods: The study sample is a post-9-11 cohort of veterans in their first year of VHA care, with at least one outpatient care visit between 2012-2017 (N = 762,953). Demographic data, veteran military service characteristics, and clinical comorbidities were obtained from VHA EHR. We extracted clinical notes for outpatient visits to primary, urgent, or emergency clinics (total 105,316,004). Natural language processing and machine learning (ML) approaches were used to identify documentation of firearm access. A taxonomy of firearm terms was identified and manually annotated with text anchored by these terms, and then trained the ML algorithm. The random-forest algorithm achieved 81.9% accuracy in identifying documentation of firearm access.Results:The proportion of patients with EHR-documented access to one or more firearms during their first year of care in the VHA was relatively low and varied by patient characteristics. Men had significantly higher documentation of firearms than women (9.8% vs 7.1%; P < .001) and veterans >50 years old had the lowest (6.5%). Among veterans with any firearm term present, only 24.4% were classified as positive for access to a firearm (24.7% of men and 20.9% of women).Conclusion: Natural language processing can identify documentation of access to firearms in clinical notes with acceptable accuracy, but there is a need for investigation into facilitators and barriers for providers and veterans to improve a systemwide process of firearm access screening. Screening, regardless of race/ethnicity, gender, and age, provides additional opportunities to protect veterans from self-harm and violence.

Guideline concordant opioid therapy is a key part of the concerted effort to address the opioid crisis in the United States. The study aimed to compare the rates of guideline concordant care between veterans who solely used VA services (mono users) and veterans who used both VA services and community care (dual-system users). We used electronic health record data from the Washington DC and Baltimore VA Medical Centers from 2015 to 2019. We provided descriptive statistics as well as generalized estimating equations models to find associations between mono vs. dual-system users and each guideline outcome, controlling for demographic factors and comorbid conditions. The study found that overall rates of guideline concordant care were high in both mono and dual-system users with over 90% adherence rates for the majority of recommendations. However, there were variations in adherence to specific guidelines, with urine drug screening at initiation being the least commonly followed recommendation (8.9% of mono-user opioid initiators and 11.2% of dual-user initiators). This study also found that there was no consistent pattern of higher guideline adherence in mono vs. dual-system users but did show that through the course of this study (2015–2019) overall rates of guideline concordance increased. Future research will explore additional guideline recommendations and potential coordination issues among dual-system users. [ABSTRACT FROM AUTHOR]

Aims: According to the Kidney Disease: Improving Global Outcomes (KDIGO) guideline, the definition of chronic kidney disease (CKD) requires the presence of abnormal kidney structure or function for >3 months with implications for health. CKD in patients with heart failure (HF) has not been defined using this definition, and less is known about the true health implications of CKD in these patients. The objective of the current study was to identify patients with HF who met KDIGO criteria for CKD and examine their outcomes. Methods and results: Of the 1 419 729 Veterans with HF not receiving kidney replacement therapy, 828 744 had data on ≥2 ambulatory serum creatinine >90 days apart. CKD was defined as estimated glomerular filtration rate (eGFR) <60 ml/min/1.73 m2 (n = 185 821) or urinary albumin‐to‐creatinine ratio (uACR) >30 mg/g (n = 32 730) present twice >3 months apart. Normal kidney function (NKF) was defined as eGFR ≥60 ml/min/1.73 m2, present for >3 months, without any uACR >30 mg/g (n = 365 963). Patients with eGFR <60 ml/min/1.73 m2 were categorized into four stages: 45–59 (n = 72 606), 30–44 (n = 74 812), 15–29 (n = 32 077), and <15 (n = 6326) ml/min/1.73 m2. Five‐year all‐cause mortality occurred in 40.4%, 57.8%, 65.6%, 73.3%, 69.7%, and 47.5% of patients with NKF, four eGFR stages, and uACR >30mg/g (albuminuria), respectively. Compared with NKF, hazard ratios (HR) (95% confidence intervals [CI]) for all‐cause mortality associated with the four eGFR stages and albuminuria were 1.63 (1.62–1.65), 2.00 (1.98–2.02), 2.49 (2.45–2.52), 2.28 (2.21–2.35), and 1.22 (1.20–1.24), respectively. Respective age‐adjusted HRs (95% CIs) were 1.13 (1.12–1.14), 1.36 (1.34–1.37), 1.87 (1.84–1.89), 2.24 (2.18–2.31) and 1.19 (1.17–1.21), and multivariable‐adjusted HRs (95% CIs) were 1.11 (1.10–1.12), 1.24 (1.22–1.25), 1.46 (1.43–1.48), 1.42 (1.38–1.47), and 1.13 (1.11–1.16). Similar patterns were observed for associations with hospitalizations. Conclusion: Data needed to define CKD using KDIGO criteria were available in six out of ten patients, and CKD could be defined in seven out of ten patients with data. HF patients with KDIGO‐defined CKD had higher risks for poor outcomes, most of which was not explained by abnormal kidney structure or function. Future studies need to examine whether CKD defined using a single eGFR is characteristically and prognostically different from CKD defined using KDIGO criteria. [ABSTRACT FROM AUTHOR]

Opioid use disorder is known to be under-coded as a diagnosis, yet problematic opioid use can be documented in clinical notes, which are included in electronic health records. We sought to identify problematic opioid use from a full range of clinical notes and compare the demographic and clinical characteristics of patients identified as having problematic opioid use exclusively in clinical notes to patients documented through ICD opioid use disorder diagnostic codes. We developed and applied a natural language processing (NLP) tool that combines rule-based pattern analysis and a trained support vector machine to the clinical notes of a patient cohort (n = 222,371) from two Veteran Affairs service regions to identify patients with problematic opioid use. We also used a set of ICD diagnostic codes to identify patients with opioid use disorder from the same cohort. The NLP tool achieved 96.6% specificity, 90.4% precision/PPV, 88.4% sensitivity/recall, and 94.4% accuracy on unseen test data. NLP exclusively identified 57,331 patients; 6997 patients had positive ICD code identifications. Patients exclusively identified through NLP were more likely to be women. Those identified through ICD codes were more likely to be male, younger, have concurrent benzodiazepine prescriptions, more comorbidities, and more care encounters, and were less likely to be married. Patients in both these groups had substantially elevated comorbidity levels compared with patients not documented through either method as experiencing problematic opioid use. Clinicians may be reluctant to code for opioid use disorder. It is therefore incumbent on the healthcare team to search for documentation of opioid concerns within clinical notes. [ABSTRACT FROM AUTHOR]

Introduction: According to the US Renal Data System (USRDS), patients with end-stage kidney disease (ESKD) on maintenance dialysis had higher mortality during early COVID-19 pandemic. Less is known about the effect of the pandemic on the delivery of outpatient maintenance hemodialysis and its impact on death. We examined the effect of pandemic-related disruption on the delivery of dialysis treatment and mortality in patients with ESKD receiving maintenance hemodialysis in the Veterans Health Administration (VHA) facilities, the largest integrated national healthcare system in the USA. Methods: Using national VHA electronic health records data, we identified 7,302 Veterans with ESKD who received outpatient maintenance hemodialysis in VHA healthcare facilities during the COVID-19 pandemic (February 1, 2020, to December 31, 2021). We estimated the average change in the number of hemodialysis treatments received and deaths per 1,000 patients per month during the pandemic by conducting interrupted time-series analyses. We used seasonal autoregressive moving average (SARMA) models, in which February 2020 was used as the conditional intercept and months thereafter as conditional slope. The models were adjusted for seasonal variations and trends in rates during the pre-pandemic period (January 1, 2007, to January 31, 2020). Results: The number (95% CI) of hemodialysis treatments received per 1,000 patients per month during the pre-pandemic and pandemic periods were 12,670 (12,525–12,796) and 12,865 (12,729–13,002), respectively. Respective all-cause mortality rates (95% CI) were 17.1 (16.7–17.5) and 19.6 (18.5–20.7) per 1,000 patients per month. Findings from SARMA models demonstrate that there was no reduction in the dialysis treatments delivered during the pandemic (rate ratio: 0.999; 95% CI: 0.998–1.001), but there was a 2.3% (95% CI: 1.5–3.1%) increase in mortality. During the pandemic, the non-COVID hospitalization rate was 146 (95% CI: 143–149) per 1,000 patients per month, which was lower than the pre-pandemic rate of 175 (95% CI: 173–176). In contrast, there was evidence of higher use of telephone encounters during the pandemic (3,023; 95% CI: 2,957–3,089), compared with the pre-pandemic rate (1,282; 95% CI: 1,241–1,324). Conclusions: We found no evidence that there was a disruption in the delivery of outpatient maintenance hemodialysis treatment in VHA facilities during the COVID-19 pandemic and that the modest rise in deaths during the pandemic is unlikely to be due to missed dialysis. [ABSTRACT FROM AUTHOR]

Background and Aims: In deep learning, a major difficulty in identifying suicidality and its risk factors in clinical notes is the lack of training samples given the small number of true positive instances among the number of patients screened. This paper describes a novel methodology that identifies suicidality in clinical notes by addressing this data sparsity issue through zero‐shot learning. Our general aim was to develop a tool that leveraged zero‐shot learning to effectively identify suicidality documentation in all types of clinical notes. Methods: US Veterans Affairs clinical notes served as data. The training data set label was determined using diagnostic codes of suicide attempt and self‐harm. We used a base string associated with the target label of suicidality to provide auxiliary information by narrowing the positive training cases to those containing the base string. We trained a deep neural network by mapping the training documents' contents to a semantic space. For comparison, we trained another deep neural network using the identical training data set labels, and bag‐of‐words features. Results: The zero‐shot learning model outperformed the baseline model in terms of area under the curve, sensitivity, specificity, and positive predictive value at multiple probability thresholds. In applying a 0.90 probability threshold, the methodology identified notes documenting suicidality but not associated with a relevant ICD‐10‐CM code, with 94% accuracy. Conclusion: This method can effectively identify suicidality without manual annotation. Key points: Due to data sparsity, suicidality documentation is difficult to identify in clinical notes. Zero‐shot learning addresses the data sparsity issue. Zero‐shot learning enables identification of suicidality and its risks in clinical notes, and associated patients, where no diagnostic code relevant to suicide or self‐harm has been recorded. [ABSTRACT FROM AUTHOR]