Development of a liver disease-specific large language model chat interface using retrieval-augmented generation.

BACKGROUND AND AIMS: Large language models (LLMs) have significant capabilities in clinical information processing tasks. Commercially available LLMs, however, are not optimized for clinical uses and are prone to generating hallucinatory information. Retrieval-augmented generation (RAG) is an enterprise architecture that allows the embedding of customized data into LLMs. This approach "specializes" the LLMs and is thought to reduce hallucinations. APPROACH AND RESULTS: We developed "LiVersa," a liver disease-specific LLM, by using our institution's protected health information-complaint text embedding and LLM platform, "Versa." We conducted RAG on 30 publicly available American Association for the Study of Liver Diseases guidance documents to be incorporated into LiVersa. We evaluated LiVersa's performance by conducting 2 rounds of testing. First, we compared LiVersa's outputs versus those of trainees from a previously published knowledge assessment. LiVersa answered all 10 questions correctly. Second, we asked 15 hepatologists to evaluate the outputs of 10 hepatology topic questions generated by LiVersa, OpenAI's ChatGPT 4, and Meta's Large Language Model Meta AI 2. LiVersa's outputs were more accurate but were rated less comprehensive and safe compared to those of ChatGPT 4. RESULTS: We evaluated LiVersa's performance by conducting 2 rounds of testing. First, we compared LiVersa's outputs versus those of trainees from a previously published knowledge assessment. LiVersa answered all 10 questions correctly. Second, we asked 15 hepatologists to evaluate the outputs of 10 hepatology topic questions generated by LiVersa, OpenAI's ChatGPT 4, and Meta's Large Language Model Meta AI 2. LiVersa's outputs were more accurate but were rated less comprehensive and safe compared to those of ChatGPT 4. CONCLUSIONS: In this demonstration, we built disease-specific and protected health information-compliant LLMs using RAG. While LiVersa demonstrated higher accuracy in answering questions related to hepatology, there were some deficiencies due to limitations set by the number of documents used for RAG. LiVersa will likely require further refinement before potential live deployment. The LiVersa prototype, however, is a proof of concept for utilizing RAG to customize LLMs for clinical use cases.

Applying human-centered design to the construction of a cirrhosis management clinical decision support system.

BACKGROUND: Electronic health record (EHR)-based clinical decision support is a scalable way to help standardize clinical care. Clinical decision support systems have not been extensively investigated in cirrhosis management. Human-centered design (HCD) is an approach that engages with potential users in intervention development. In this study, we applied HCD to design the features and interface for a clinical decision support system for cirrhosis management, called CirrhosisRx. METHODS: We conducted technical feasibility assessments to construct a visual blueprint that outlines the basic features of the interface. We then convened collaborative-design workshops with generalist and specialist clinicians. We elicited current workflows for cirrhosis management, assessed gaps in existing EHR systems, evaluated potential features, and refined the design prototype for CirrhosisRx. At the conclusion of each workshop, we analyzed recordings and transcripts. RESULTS: Workshop feedback showed that the aggregation of relevant clinical data into 6 cirrhosis decompensation domains (defined as common inpatient clinical scenarios) was the most important feature. Automatic inference of clinical events from EHR data, such as gastrointestinal bleeding from hemoglobin changes, was not accepted due to accuracy concerns. Visualizations for risk stratification scores were deemed not necessary. Lastly, the HCD co-design workshops allowed us to identify the target user population (generalists). CONCLUSIONS: This is one of the first applications of HCD to design the features and interface for an electronic intervention for cirrhosis management. The HCD process altered features, modified the design interface, and likely improved CirrhosisRx's overall usability. The finalized design for CirrhosisRx proceeded to development and production and will be tested for effectiveness in a pragmatic randomized controlled trial. This work provides a model for the creation of other EHR-based interventions in hepatology care.

Factors associated with anxiety during the first two years of the COVID-19 pandemic in the United States: An analysis of the COVID-19 Citizen Science study.

COVID-19 increased the prevalence of clinically significant anxiety in the United States. To investigate contributing factors we analyzed anxiety, reported online via monthly Generalized Anxiety Disorders-7 (GAD-7) surveys between April 2020 and May 2022, in association with self-reported worry about the health effects of COVID-19, economic difficulty, personal COVID-19 experience, and subjective social status. 333,292 anxiety surveys from 50,172 participants (82% non-Hispanic white; 73% female; median age 55, IQR 42-66) showed high levels of anxiety, especially early in the pandemic. Anxiety scores showed strong independent associations with worry about the health effects of COVID-19 for oneself or family members (GAD-7 score +3.28 for highest vs. lowest category; 95% confidence interval: 3.24, 3.33; p<0.0001 for trend) and with difficulty paying for basic living expenses (+2.06; 1.97, 2.15, p<0.0001) in multivariable regression models after adjusting for demographic characteristics, COVID-19 case rates and death rates, and personal COVID-19 experience. High levels of COVID-19 health worry and economic stress were each more common among participants reporting lower subjective social status, and median anxiety scores for those experiencing both were in the range considered indicative of moderate to severe clinical anxiety disorders. In summary, health worry and economic difficulty both contributed to high rates of anxiety during the first two years of the COVID-19 pandemic in the US, especially in disadvantaged socioeconomic groups. Programs to address both health concerns and economic insecurity in vulnerable populations could help mitigate pandemic impacts on anxiety and mental health.

Association of nirmatrelvir for acute SARS-CoV-2 infection with subsequent Long COVID symptoms in an observational cohort study.

Oral nirmatrelvir/ritonavir is approved as treatment for acute COVID-19, but the effect of treatment during acute infection on risk of Long COVID is unknown. We hypothesized that nirmatrelvir treatment during acute SARS-CoV-2 infection reduces risk of developing Long COVID and rebound after treatment is associated with Long COVID. We conducted an observational cohort study within the Covid Citizen Science (CCS) study, an online cohort study with over 100 000 participants. We included vaccinated, nonhospitalized, nonpregnant individuals who reported their first SARS-CoV-2 positive test March-August 2022. Oral nirmatrelvir/ritonavir treatment was ascertained during acute SARS-CoV-2 infection. Patient-reported Long COVID symptoms, symptom rebound and test-positivity rebound were asked on subsequent surveys at least 3 months after SARS-CoV-2 infection. A total of 4684 individuals met the eligibility criteria, of whom 988 (21.1%) were treated and 3696 (78.9%) were untreated; 353/988 (35.7%) treated and 1258/3696 (34.0%) untreated responded to the Long COVID survey (n = 1611). Among 1611 participants, median age was 55 years and 66% were female. At 5.4 ± 1.3 months after infection, nirmatrelvir treatment was not associated with subsequent Long COVID symptoms (odds ratio [OR]: 1.15; 95% confidence interval [CI]: 0.80-1.64; p = 0.45). Among 666 treated who answered rebound questions, rebound symptoms or test positivity were not associated with Long COVID symptoms (OR: 1.34; 95% CI: 0.74-2.41; p = 0.33). Within this cohort of vaccinated, nonhospitalized individuals, oral nirmatrelvir treatment during acute SARS-CoV-2 infection and rebound after nirmatrelvir treatment were not associated with Long COVID symptoms more than 90 days after infection.

Preceding Night Sleep Quality and Atrial Fibrillation Episodes in the I-STOP-AFIB Randomized Trial.

BACKGROUND: Chronic sleep disruption is associated with incident atrial fibrillation (AF), but it is unclear whether poor sleep quality acutely triggers AF. OBJECTIVES: The aim of this study was to characterize the relationship between a given night's sleep quality and the risk of a discrete AF episode. METHODS: Patients with symptomatic paroxysmal AF in the I-STOP-AFIB (Individualized Studies of Triggers of Paroxysmal Atrial Fibrillation) trial reported sleep quality on a daily basis. Participants were also queried daily regarding AF episodes and were provided smartphone-based mobile electrocardiograms (ECGs) (KardiaMobile, AliveCor). RESULTS: Using 15,755 days of data from 419 patients, worse sleep quality on any given night was associated with a 15% greater odds of a self-reported AF episode the next day (OR: 1.15; 95% CI: 1.10-1.20; P < 0.0001) after adjustment for the day of the week. No statistically significant associations between worsening sleep quality and mobile ECG-confirmed AF events were observed (OR: 1.04; 95% CI: 0.95-1.13; P = 0.43), although substantially fewer of these mobile ECG-confirmed events may have limited statistical power. Poor sleep was also associated with longer self-reported AF episodes, with each progressive category of worsening sleep associated with 16 (95% CI: 12-21; P < 0.001) more minutes of AF the next day. CONCLUSIONS: Poor sleep was associated with an immediately heightened risk for self-reported AF episodes, and a dose-response relationship existed such that progressively worse sleep was associated with longer episodes of AF the next day. These data suggest that sleep quality may be a potentially modifiable trigger relevant to the near-term risk of a discrete AF episode.

Electronic health record intervention to increase use of NSAIDs as analgesia for hospitalised patients: a cluster randomised controlled study.

BACKGROUND: Prescribing non-opioid pain medications, such as non-steroidal anti-inflammatory (NSAIDs) medications, has been shown to reduce pain and decrease opioid use, but it is unclear how to effectively encourage multimodal pain medication prescribing for hospitalised patients. Therefore, the aim of this study is to evaluate the effect of prechecking non-opioid pain medication orders on clinician prescribing of NSAIDs among hospitalised adults. METHODS: This was a cluster randomised controlled trial of adult (≥18 years) hospitalised patients admitted to three hospital sites under one quaternary hospital system in the USA from 2 March 2022 to 3 March 2023. A multimodal pain order panel was embedded in the admission order set, with NSAIDs prechecked in the intervention group. The intervention group could uncheck the NSAID order. The control group had access to the same NSAID order. The primary outcome was an increase in NSAID ordering. Secondary outcomes include NSAID administration, inpatient pain scores and opioid use and prescribing and relevant clinical harms including acute kidney injury, new gastrointestinal bleed and in-hospital death. RESULTS: Overall, 1049 clinicians were randomised. The study included 6239 patients for a total of 9595 encounters. Both NSAID ordering (36 vs 43%, p<0.001) and administering (30 vs 34%, p=0.001) by the end of the first full hospital day were higher in the intervention (prechecked) group. There was no statistically significant difference in opioid outcomes during the hospitalisation and at discharge. There was a statistically but perhaps not clinically significant difference in pain scores during both the first and last full hospital day. CONCLUSIONS: This cluster randomised controlled trial showed that prechecking an order for NSAIDs to promote multimodal pain management in the admission order set increased NSAID ordering and administration, although there were no changes to pain scores or opioid use. While prechecking orders is an important way to increase adoption, safety checks should be in place.

Email-Based Recruitment Into the Health eHeart Study: Cohort Analysis of Invited Eligible Patients.

BACKGROUND: Web- or app-based digital health studies allow for more efficient collection of health data for research. However, remote recruitment into digital health studies can enroll nonrepresentative study samples, hindering the robustness and generalizability of findings. Through the comprehensive evaluation of an email-based campaign on recruitment into the Health eHeart Study, we aim to uncover key sociodemographic and clinical factors that contribute to enrollment. OBJECTIVE: This study sought to understand the factors related to participation, specifically regarding enrollment, in the Health eHeart Study as a result of a large-scale remote email recruitment campaign. METHODS: We conducted a cohort analysis on all invited University of California, San Francisco (UCSF) patients to identify sociodemographic and clinical predictors of enrollment into the Health eHeart Study. The primary outcome was enrollment, defined by account registration and consent into the Health eHeart Study. The email recruitment campaign was carried out from August 2015 to February 2016, with electronic health record data extracted between September 2019 and December 2019. RESULTS: The email recruitment campaign delivered at least 1 email invitation to 93.5% (193,606/206,983) of all invited patients and yielded a 3.6% (7012/193,606) registration rate among contacted patients and an 84.1% (5899/7012) consent rate among registered patients. Adjusted multivariate logistic regression models analyzed independent sociodemographic and clinical predictors of (1) registration among contacted participants and (2) consent among registered participants. Odds of registration were higher among patients who are older, women, non-Hispanic White, active patients with commercial insurance or Medicare, with a higher comorbidity burden, with congestive heart failure, and randomized to receive up to 2 recruitment emails. The odds of registration were lower among those with medical conditions such as dementia, chronic pulmonary disease, moderate or severe liver disease, paraplegia or hemiplegia, renal disease, or cancer. Odds of subsequent consent after initial registration were different, with an inverse trend of being lower among patients who are older and women. The odds of consent were also lower among those with peripheral vascular disease. However, the odds of consent remained higher among patients who were non-Hispanic White and those with commercial insurance. CONCLUSIONS: This study provides important insights into the potential returns on participant enrollment when digital health study teams invest resources in using email for recruitment. The findings show that participant enrollment was driven more strongly by sociodemographic factors than clinical factors. Overall, email is an extremely efficient means of recruiting participants from a large list into the Health eHeart Study. Despite some improvements in representation, the formulation of truly diverse studies will require additional resources and strategies to overcome persistent participation barriers.

Digital Tool-Assisted Hospitalization Detection in the Tailored Antiplatelet Initiation to Lessen Outcomes due to Decreased Clopidogrel Response After Percutaneous Coronary Intervention Study Compared to Traditional Site-Coordinator Ascertainment: Intervention Study.

BACKGROUND: Accurate, timely ascertainment of clinical end points, particularly hospitalizations, is crucial for clinical trials. The Tailored Antiplatelet Initiation to Lessen Outcomes Due to Decreased Clopidogrel Response after Percutaneous Coronary Intervention (TAILOR-PCI) Digital Study extended the main TAILOR-PCI trial's follow-up to 2 years, using a smartphone-based research app featuring geofencing-triggered surveys and routine monthly mobile phone surveys to detect cardiovascular (CV) hospitalizations. This pilot study compared these digital tools to conventional site-coordinator ascertainment of CV hospitalizations. OBJECTIVE: The objectives were to evaluate geofencing-triggered notifications and routine monthly mobile phone surveys' performance in detecting CV hospitalizations compared to telephone visits and health record reviews by study coordinators at each site. METHODS: US and Canadian participants from the TAILOR-PCI Digital Follow-Up Study were invited to download the Eureka Research Platform mobile app, opting in for location tracking using geofencing, triggering a smartphone-based survey if near a hospital for ≥4 hours. Participants were sent monthly notifications for CV hospitalization surveys. RESULTS: From 85 participants who consented to the Digital Study, downloaded the mobile app, and had not previously completed their final follow-up visit, 73 (85.8%) initially opted in and consented to geofencing. There were 9 CV hospitalizations ascertained by study coordinators among 5 patients, whereas 8 out of 9 (88.9%) were detected by routine monthly hospitalization surveys. One CV hospitalization went undetected by the survey as it occurred within two weeks of the previous event, and the survey only allowed reporting of a single hospitalization. Among these, 3 were also detected by the geofencing algorithm, but 6 out of 9 (66.7%) were missed by geofencing: 1 occurred in a participant who never consented to geofencing, while 5 hospitalizations occurred among participants who had subsequently turned off geofencing prior to their hospitalization. Geofencing-detected hospitalizations were ascertained within a median of 2 (IQR 1-3) days, monthly surveys within 11 (IQR 6.5-25) days, and site coordinator methods within 38 (IQR 9-105) days. The geofencing algorithm triggered 245 notifications among 39 participants, with 128 (52.2%) from true hospital presence and 117 (47.8%) from nonhospital health care facility visits. Additional geofencing iterative improvements to reduce hospital misidentification were made to the algorithm at months 7 and 12, elevating the rate of true alerts from 35.4% (55 true alerts/155 total alerts before month 7) to 78.7% (59 true alerts/75 total alerts in months 7-12) and ultimately to 93.3% (14 true alerts/5 total alerts in months 13-21), respectively. CONCLUSIONS: The monthly digital survey detected most CV hospitalizations, while the geofencing survey enabled earlier detection but did not offer incremental value beyond traditional tools. Digital tools could potentially reduce the burden on study coordinators in ascertaining CV hospitalizations. The advantages of timely reporting via geofencing should be weighed against the issue of false notifications, which can be mitigated through algorithmic refinements.

Development of a Liver Disease-Specific Large Language Model Chat Interface using Retrieval Augmented Generation.

Background: Large language models (LLMs) have significant capabilities in clinical information processing tasks. Commercially available LLMs, however, are not optimized for clinical uses and are prone to generating incorrect or hallucinatory information. Retrieval-augmented generation (RAG) is an enterprise architecture that allows embedding of customized data into LLMs. This approach "specializes" the LLMs and is thought to reduce hallucinations. Methods: We developed "LiVersa," a liver disease-specific LLM, by using our institution's protected health information (PHI)-complaint text embedding and LLM platform, "Versa." We conducted RAG on 30 publicly available American Association for the Study of Liver Diseases (AASLD) guidelines and guidance documents to be incorporated into LiVersa. We evaluated LiVersa's performance by comparing its responses versus those of trainees from a previously published knowledge assessment study regarding hepatitis B (HBV) treatment and hepatocellular carcinoma (HCC) surveillance. Results: LiVersa answered all 10 questions correctly when forced to provide a "yes" or "no" answer. Full detailed responses with justifications and rationales, however, were not completely correct for three of the questions. Discussions: In this study, we demonstrated the ability to build disease-specific and PHI-compliant LLMs using RAG. While our LLM, LiVersa, demonstrated more specificity in answering questions related to clinical hepatology - there were some knowledge deficiencies due to limitations set by the number and types of documents used for RAG. The LiVersa prototype, however, is a proof of concept for utilizing RAG to customize LLMs for clinical uses and a potential strategy to realize personalized medicine in the future.

Institutionalizing Community-engaged Translational Science in an Academic Institution: A Community Stakeholder-Driven Process.

BACKGROUND: Although studies have described the power imbalance in academic-community partnerships, little has been published describing how community-based participatory research-informed practitioners can change academic institutions to promote more effective community-engaged research. OBJECTIVES: This paper describes a university-funded community-based participatory project in which academic researchers and their community partners worked together to articulate, develop and advocate for institutionalizing best practices for equitable partnerships throughout the university. METHODS: Findings derive from a collaborative ethnographic process evaluation. RESULTS: The study describes the integral steps proposed to promote equitable community-university research collaboration, the process by which these principles and best practice recommendations were developed, and the institutional change outcomes of this process. CONCLUSIONS: When universities make even small investments toward promoting and nurturing community-engaged research, the quality of the science can be enhanced to advance health equity and community-university relationships can improve, particularly if based on trust, mutual respect, and openness to accomplish a shared vision.

Associations between tobacco and cannabis use and anxiety and depression among adults in the United States: Findings from the COVID-19 citizen science study.

BACKGROUND: Little is known about whether people who use both tobacco and cannabis (co-use) are more or less likely to have mental health disorders than single substance users or non-users. We aimed to examine associations between use of tobacco and/or cannabis with anxiety and depression. METHODS: We analyzed data from the COVID-19 Citizen Science Study, a digital cohort study, collected via online surveys during 2020-2022 from a convenience sample of 53,843 US adults (≥ 18 years old) nationwide. Past 30-day use of tobacco and cannabis was self-reported at baseline and categorized into four exclusive patterns: tobacco-only use, cannabis-only use, co-use of both substances, and non-use. Anxiety and depression were repeatedly measured in monthly surveys. To account for multiple assessments of mental health outcomes within a participant, we used Generalized Estimating Equations to examine associations between the patterns of tobacco and cannabis use with each outcome. RESULTS: In the total sample (mean age 51.0 years old, 67.9% female), 4.9% reported tobacco-only use, 6.9% cannabis-only use, 1.6% co-use, and 86.6% non-use. Proportions of reporting anxiety and depression were highest for the co-use group (26.5% and 28.3%, respectively) and lowest for the non-use group (10.6% and 11.2%, respectively). Compared to non-use, the adjusted odds of mental health disorders were highest for co-use (Anxiety: OR = 1.89, 95%CI = 1.64-2.18; Depression: OR = 1.77, 95%CI = 1.46-2.16), followed by cannabis-only use, and tobacco-only use. Compared to tobacco-only use, co-use (OR = 1.35, 95%CI = 1.08-1.69) and cannabis-only use (OR = 1.17, 95%CI = 1.00-1.37) were associated with higher adjusted odds for anxiety, but not for depression. Daily use (vs. non-daily use) of cigarettes, e-cigarettes, and cannabis were associated with higher adjusted odds for anxiety and depression. CONCLUSIONS: Use of tobacco and/or cannabis, particularly co-use of both substances, were associated with poor mental health. Integrating mental health support with tobacco and cannabis cessation may address this co-morbidity.

Predicting post-liver transplant outcomes in patients with acute-on-chronic liver failure using Expert-Augmented Machine Learning.

Liver transplantation (LT) is a treatment for acute-on-chronic liver failure (ACLF), but high post-LT mortality has been reported. Existing post-LT models in ACLF have been limited. We developed an Expert-Augmented Machine Learning (EAML) model to predict post-LT outcomes. We identified ACLF patients who underwent LT in the University of California Health Data Warehouse. We applied the RuleFit machine learning (ML) algorithm to extract rules from decision trees and create intermediate models. We asked human experts to rate the rules generated by RuleFit and incorporated these ratings to generate final EAML models. We identified 1384 ACLF patients. For death at 1 year, areas under the receiver-operating characteristic curve were 0.707 (confidence interval [CI] 0.625-0.793) for EAML and 0.719 (CI 0.640-0.800) for RuleFit. For death at 90 days, areas under the receiver-operating characteristic curve were 0.678 (CI 0.581-0.776) for EAML and 0.707 (CI 0.615-0.800) for RuleFit. In pairwise comparisons, both EAML and RuleFit models outperformed cross-sectional models. Significant discrepancies between experts and ML occurred in rankings of biomarkers used in clinical practice. EAML may serve as a method for ML-guided hypothesis generation in further ACLF research.

Optimizing Uptake of Multimodal Pain Management After Surgery Using the Electronic Health Record.

This quality improvement study evaluates the effect of an electronic health record intervention on multimodal pain management following surgery in 2 randomized clinical trials.

Clinical decision support and electronic interventions to improve care quality in chronic liver diseases and cirrhosis.

Significant quality gaps exist in the management of chronic liver diseases and cirrhosis. Clinical decision support systems-information-driven tools based in and launched from the electronic health record-are attractive and potentially scalable prospective interventions that could help standardize clinical care in hepatology. Yet, clinical decision support systems have had a mixed record in clinical medicine due to issues with interoperability and compatibility with clinical workflows. In this review, we discuss the conceptual origins of clinical decision support systems, existing applications in liver diseases, issues and challenges with implementation, and emerging strategies to improve their integration in hepatology care.

Sex Differences in the Association Between Cumulative Use of Cannabis and Cognitive Function in Middle Age: The Coronary Artery Risk Development in Young Adults Study.

Background: Cannabis use may impair cognitive function (CF) differently in men and women, due to sex-specific differences in neurobiological mechanisms and environmental risk factors. Objective: Assess sex differences in the association between cumulative exposure to cannabis and cognitive performance in middle age. Methods: We studied participants from the Coronary Artery Risk Development in Young Adults (CARDIA) Study, including Black and White men and women 18-30 years old at baseline followed over 30 years. Our cross-sectional analysis of CF scores at year 30 was stratified by sex. We computed categories of cumulative exposure in "cannabis-years" (1 cannabis-year=365 days of use) from self-reported use every 2 to 5 years over 30 years. At years 25 and 30, we assessed CF with the Rey Auditory Verbal Learning Test (verbal memory), the Digit Symbol Substitution Test (processing speed), and the Stroop Interference Test (executive function). At year 30, additional measures included Category and Letter Fluency Test (verbal ability) and the Montreal Cognitive Assessment (global cognition). We computed standardized scores for each cognitive test and applied multivariable adjusted linear regression models for self-reported cumulative cannabis use, excluding participants who used cannabis within 24 h. In a secondary analysis, we examined the association between changes in current cannabis use and changes in CF between years 25 and 30. Results: By year 30, 1,352 men and 1,793 women had measures of CF; 87% (N=1,171) men and 84% (N=1,502) women reported ever cannabis use. Men had a mean cumulative use of 2.57 cannabis-years and women 1.29 cannabis-years. Self-reported cumulative cannabis use was associated with worse verbal memory in men (e.g., -0.49 standardized units [SU] for ≥5 cannabis-years of exposure; 95% CI=-0.76 to -0.23), but not in women (SU=0.02; 95% CI=-0.26 to 0.29). Other measures of CF were not associated with cannabis. Changes in current cannabis use between years 25 and 30 were not associated with CF in men or women. Conclusions: Self-reported cumulative cannabis exposure was associated with worse verbal memory in men but not in women. Researchers should consider stratified analyses by sex when testing the association between cannabis and cognition.

Learning health system, positive deviance analysis, and electronic health records: Synergy for a learning health system.

Introduction: Over the past decade, numerous efforts have encouraged the realization of the learning health system (LHS) in the United States. Despite these efforts, and promising aims of the LHS, the full potential and value of research conducted within LHSs have yet to be realized. New technology coupled with a catalyzing global pandemic have spurred momentum. In addition, the LHS has lacked a consistent framework within which "best evidence" can be identified. Positive deviance analysis, itself reinvigorated by recent advances in health information technology (IT) and ubiquitous adoption of electronic health records (EHRs), may finally provide a framework through which LHSs can be operationalized and optimized. Methods: We describe the synergy between positive deviance and the LHS and how they may be integrated to achieve a continuous cycle of health system improvement. Results: As we describe below, the positive deviance approach focuses on learning from high-performing teams and organizations. Conclusion: Such learning can be enabled by EHRs and health IT, providing a lens into how digital clinical interventions are successfully developed and deployed.

A common allele of HLA is associated with asymptomatic SARS-CoV-2 infection.

Studies have demonstrated that at least 20% of individuals infected with SARS-CoV-2 remain asymptomatic1-4. Although most global efforts have focused on severe illness in COVID-19, examining asymptomatic infection provides a unique opportunity to consider early immunological features that promote rapid viral clearance. Here, postulating that variation in the human leukocyte antigen (HLA) loci may underly processes mediating asymptomatic infection, we enrolled 29,947 individuals, for whom high-resolution HLA genotyping data were available, in a smartphone-based study designed to track COVID-19 symptoms and outcomes. Our discovery cohort (n = 1,428) comprised unvaccinated individuals who reported a positive test result for SARS-CoV-2. We tested for association of five HLA loci with disease course and identified a strong association between HLA-B*15:01 and asymptomatic infection, observed in two independent cohorts. Suggesting that this genetic association is due to pre-existing T cell immunity, we show that T cells from pre-pandemic samples from individuals carrying HLA-B*15:01 were reactive to the immunodominant SARS-CoV-2 S-derived peptide NQKLIANQF. The majority of the reactive T cells displayed a memory phenotype, were highly polyfunctional and were cross-reactive to a peptide derived from seasonal coronaviruses. The crystal structure of HLA-B*15:01-peptide complexes demonstrates that the peptides NQKLIANQF and NQKLIANAF (from OC43-CoV and HKU1-CoV) share a similar ability to be stabilized and presented by HLA-B*15:01. Finally, we show that the structural similarity of the peptides underpins T cell cross-reactivity of high-affinity public T cell receptors, providing the molecular basis for HLA-B*15:01-mediated pre-existing immunity.