Statistical inference for association studies using electronic health records: handling both selection bias and outcome misclassification

PATIENT RECRUITMENT USING ELECTRONIC HEALTH RECORDS UNDER SELECTION BIAS: A TWO-PHASE SAMPLING FRAMEWORK

Electronic health records (EHRs) are increasingly recognized as a cost-effective resource for patient recruitment in clinical research. However, how to optimally select a cohort from millions of individuals to answer a scientific question of interest remains unclear. Consider a study to estimate the mean or mean difference of an expensive outcome. Inexpensive auxiliary covariates predictive of the outcome may often be available in patients' health records, presenting an opportunity to recruit patients selectively, which may improve efficiency in downstream analyses. In this paper we propose a two-phase sampling design that leverages available information on auxiliary covariates in EHR data. A key challenge in using EHR data for multiphase sampling is the potential selection bias, because EHR data are not necessarily representative of the target population. Extending existing literature on two-phase sampling design, we derive an optimal two-phase sampling method that improves efficiency over random sampling while accounting for the potential selection bias in EHR data. We demonstrate the efficiency gain from our sampling design via simulation studies and an application evaluating the prevalence of hypertension among U.S. adults leveraging data from the Michigan Genomics Initiative, a longitudinal biorepository in Michigan Medicine.

To weight or not to weight? The effect of selection bias in 3 large electronic health record-linked biobanks and recommendations for practice

OBJECTIVES

To develop recommendations regarding the use of weights to reduce selection bias for commonly performed analyses using electronic health record (EHR)-linked biobank data.

MATERIALS AND METHODS

We mapped diagnosis (ICD code) data to standardized phecodes from 3 EHR-linked biobanks with varying recruitment strategies: All of Us (AOU; n = 244 071), Michigan Genomics Initiative (MGI; n = 81 243), and UK Biobank (UKB; n = 401 167). Using 2019 National Health Interview Survey data, we constructed selection weights for AOU and MGI to represent the US adult population more. We used weights previously developed for UKB to represent the UKB-eligible population. We conducted 4 common analyses comparing unweighted and weighted results.

RESULTS

For AOU and MGI, estimated phecode prevalences decreased after weighting (weighted-unweighted median phecode prevalence ratio [MPR]: 0.82 and 0.61), while UKB estimates increased (MPR: 1.06). Weighting minimally impacted latent phenome dimensionality estimation. Comparing weighted versus unweighted phenome-wide association study for colorectal cancer, the strongest associations remained unaltered, with considerable overlap in significant hits. Weighting affected the estimated log-odds ratio for sex and colorectal cancer to align more closely with national registry-based estimates.

DISCUSSION

Weighting had a limited impact on dimensionality estimation and large-scale hypothesis testing but impacted prevalence and association estimation. When interested in estimating effect size, specific signals from untargeted association analyses should be followed up by weighted analysis.

CONCLUSION

EHR-linked biobanks should report recruitment and selection mechanisms and provide selection weights with defined target populations. Researchers should consider their intended estimands, specify source and target populations, and weight EHR-linked biobank analyses accordingly.

Exposure to autoimmune disorders increases Alzheimer's disease risk in a multi-site electronic health record analysis

Molecular studies of Alzheimer's disease (AD) implicate potential links between autoimmunity and AD, but the underlying clinical relationships between these conditions remain poorly understood. Electronic health records (EHRs) provide an opportunity to determine the clinical risk relationship between autoimmune disorders and AD and understand whether specific disorders and disorder subtypes affect AD risk at the phenotypic level in human populations. We evaluated relationships between 26 autoimmune disorders and AD across retrospective observational case-control and cohort study designs in the EHR systems at UCSF and Stanford. We quantified overall and sex-specific AD risk effects that these autoimmune disorders confer. We identified significantly increased AD risk in autoimmune disorder patients in both study designs at UCSF and at Stanford. This pattern was driven by specific autoimmunity subtypes including endocrine, gastrointestinal, dermatologic, and musculoskeletal disorders. We also observed increased AD risk from autoimmunity in both women and men, but women with autoimmune disorders continued to have a higher AD prevalence than men, indicating persistent sex-specificity. This study identifies autoimmune disorders as strong risk factors for AD that validate across several study designs and EHR databases. It sets the foundation for exploring how underlying autoimmune mechanisms increase AD risk and contribute to AD pathogenesis.

Challenges in and Opportunities for Electronic Health Record-Based Data Analysis and Interpretation

Electronic health records (EHRs) have been increasingly adopted in clinical practices across the United States, providing a primary source of data for clinical research, particularly observational cohort studies. EHRs are a high-yield, low-maintenance source of longitudinal real-world data for large patient populations and provide a wealth of information and clinical contexts that are useful for clinical research and translation into practice. Despite these strengths, it is important to recognize the multiple limitations and challenges related to the use of EHR data in clinical research. Missing data are a major source of error and biases and can affect the representativeness of the cohort of interest, as well as the accuracy of the outcomes and exposures. Here, we aim to provide a critical understanding of the types of data available in EHRs and describe the impact of data heterogeneity, quality, and generalizability, which should be evaluated prior to and during the analysis of EHR data. We also identify challenges pertaining to data quality, including errors and biases, and examine potential sources of such biases and errors. Finally, we discuss approaches to mitigate and remediate these limitations. A proactive approach to addressing these issues can help ensure the integrity and quality of EHR data and the appropriateness of their use in clinical studies.

Estimation of the benefit from pre-emptive genotyping based on the nationwide cohort data in South Korea

Genetic variants affect drug responses, making pre-emptive genotyping crucial for averting serious adverse events (SAEs) and treatment failure. However, assessing the benefits of pre-emptive genotyping based on genetic distribution, drug exposure, and demographics is challenging. This study aimed to estimate the population-level benefits of pre-emptive genotyping in the Korean population using nationwide cohort data. We reviewed actionable gene-drug combinations recommended by both the Clinical Pharmacogenomics Implementation Consortium (CPIC) and the Dutch Pharmacogenetics Working Group (DPWG) as of February 2022, identifying high-risk phenotypes. We collected reported risk reduction from genotyping and standardized it into population attributable risks. Healthcare reimbursement costs for SAEs and treatment failures were obtained from the Health Insurance Review and Assessment Service Statistics in 2021. The benefits of pre-emptive genotyping for a specific group were determined by multiplying drug exposure from nationwide cohort data by individual genotyping benefits. We identified 31 gene-drug-event pairs, with CYP2D6 and CYP2C19 demonstrating the greatest benefits for both male and female patients. Individuals aged 65-70 years had the highest individual benefit from pre-emptive genotyping, with $84.40 for men and $100.90 for women. Pre-emptive genotyping, particularly for CYP2D6 and CYP2C19, can provide substantial benefits.

The Role of Health Insurance Type and Clinic Visit on Hypertension Status Among Multiethnic Chicago Residents

PURPOSE

To investigate the joint relationship of health insurance and clinic visit with hypertension among underserved populations.

DESIGN

Population-based cohort study.

SUBJECTS

Data from 1092 participants from the Chicago Multiethnic Prevention and Surveillance Study (COMPASS) between 2013 and 2020 were analyzed.

MEASURES

Five health insurance types were included: uninsured, Medicaid, Medicare, private, and other. Clinic visit over past 12 months were retrieved from medical records and categorized into 4 groups: no clinic visit, 1-3 visits, 4-7 visits, >7 visits.

ANALYSIS

Inverse-probability weighted logistic regression was used to estimate odds ratios (OR) and 95% confidence interval (CI) for hypertension status according to health insurance and clinic visit. Models were adjusted for individual socio-demographic variables and medical history.

RESULTS

The study population was predominantly Black (>85%) of low socioeconomic status. Health insurance was not associated with more clinic visit. Measured hypertension was more frequently found in private insurance (OR = 6.48, 95% CI: 1.92-21.85) compared to the uninsured group, while 1-3 clinic visits were associated with less prevalence (OR = .59, 95% CI: .35-1.00) compared to no clinic visit. These associations remained unchanged when health insurance and clinic visit were adjusted for each other.

CONCLUSION

In this study population, private insurance was associated with higher measured hypertension prevalence compared to no insurance. The associations of health insurance and clinic visit were independent of each other.

Biases in Electronic Health Records Data for Generating Real-World Evidence: An Overview

Electronic Health Records (EHR) are increasingly being perceived as a unique source of data for clinical research as they provide unprecedentedly large volumes of real-time data from real-world settings. In this review of the secondary uses of EHR, we identify the anticipated breadth of opportunities, pointing out the data deficiencies and potential biases that are likely to limit the search for true causal relationships. This paper provides a comprehensive overview of the types of biases that arise along the pathways that generate real-world evidence and the sources of these biases. We distinguish between two levels in the production of EHR data where biases are likely to arise: (i) at the healthcare system level, where the principal source of bias resides in access to, and provision of, medical care, and in the acquisition and documentation of medical and administrative data; and (ii) at the research level, where biases arise from the processes of extracting, analyzing, and interpreting these data. Due to the plethora of biases, mainly in the form of selection and information bias, we conclude with advising extreme caution about making causal inferences based on secondary uses of EHRs.

Integration of risk factor polygenic risk score with disease polygenic risk score for disease prediction

Polygenic risk score (PRS) is useful for capturing an individual's genetic susceptibility. However, previous studies have not fully exploited the potential of the risk factor PRS (RFPRS) for disease prediction. We explored the potential of integrating disease-related RFPRSs with disease PRS to enhance disease prediction performance. We constructed 112 RFPRSs and analyzed the association of RFPRSs with diseases to identify disease-related RFPRSs in 700 diseases, using the UK Biobank dataset. We uncovered 6157 statistically significant associations between 247 diseases and 109 RFPRSs. We estimated the disease PRSs of 70 diseases that exhibited statistically significant heritability, to generate RFDiseasemetaPRS-a combined PRS integrating RFPRSs and disease PRS-and compare the prediction performance metrics between RFDiseasemetaPRS and disease PRS. RFDiseasemetaPRS showed better performance for Nagelkerke's pseudo-R2, odds ratio (OR) per 1 SD, net reclassification improvement (NRI) values and difference of R2 considered by variance of R2 in 31 out of 70 diseases. Additionally, we assessed risk classification between two models by examining OR between the top 10% and remaining 90% individuals for the 31 diseases; RFDiseasemetaPRS exhibited better R2, NRI and OR than disease PRS. These findings highlight the importance of utilizing RFDiseasemetaPRS, which can provide personalized healthcare and tailored prevention strategies.

To weight or not to weight? Studying the effect of selection bias in three large EHR-linked biobanks

Objective

To explore the role of selection bias adjustment by weighting electronic health record (EHR)-linked biobank data for commonly performed analyses.

Materials and methods

We mapped diagnosis (ICD code) data to standardized phecodes from three EHR-linked biobanks with varying recruitment strategies: All of Us (AOU; n=244,071), Michigan Genomics Initiative (MGI; n=81,243), and UK Biobank (UKB; n=401,167). Using 2019 National Health Interview Survey data, we constructed selection weights for AOU and MGI to be more representative of the US adult population. We used weights previously developed for UKB to represent the UKB-eligible population. We conducted four common descriptive and analytic tasks comparing unweighted and weighted results.

Results

For AOU and MGI, estimated phecode prevalences decreased after weighting (weighted-unweighted median phecode prevalence ratio [MPR]: 0.82 and 0.61), while UKB's estimates increased (MPR: 1.06). Weighting minimally impacted latent phenome dimensionality estimation. Comparing weighted versus unweighted PheWAS for colorectal cancer, the strongest associations remained unaltered and there was large overlap in significant hits. Weighting affected the estimated log-odds ratio for sex and colorectal cancer to align more closely with national registry-based estimates.

Discussion

Weighting had limited impact on dimensionality estimation and large-scale hypothesis testing but impacted prevalence and association estimation more. Results from untargeted association analyses should be followed by weighted analysis when effect size estimation is of interest for specific signals.

Conclusion

EHR-linked biobanks should report recruitment and selection mechanisms and provide selection weights with defined target populations. Researchers should consider their intended estimands, specify source and target populations, and weight EHR-linked biobank analyses accordingly.

Bias correction models for electronic health records data in the presence of non-random sampling

Electronic health records (EHRs) contain rich clinical information for millions of patients and are increasingly used for public health research. However, non-random inclusion of subjects in EHRs can result in selection bias, with factors such as demographics, socioeconomic status, healthcare referral patterns, and underlying health status playing a role. While this issue has been well documented, little work has been done to develop or apply bias-correction methods, often due to the fact that most of these factors are unavailable in EHRs. To address this gap, we propose a series of Heckman type bias correction methods by incorporating social determinants of health selection covariates to model the EHR non-random sampling probability. Through simulations under various settings, we demonstrate the effectiveness of our proposed method in correcting biases in both the association coefficient and the outcome mean. Our method augments the utility of EHRs for public health inferences, as we show by estimating the prevalence of cardiovascular disease and its correlation with risk factors in the New York City network of EHRs.

Design, implementation, and inferential issues associated with clinical trials that rely on data in electronic medical records: a narrative review

Real world evidence is now accepted by authorities charged with assessing the benefits and harms of new therapies. Clinical trials based on real world evidence are much less expensive than randomized clinical trials that do not rely on "real world evidence" such as contained in electronic health records (EHR). Consequently, we can expect an increase in the number of reports of these types of trials, which we identify here as 'EHR-sourced trials.' 'In this selected literature review, we discuss the various designs and the ethical issues they raise. EHR-sourced trials have the potential to improve/increase common data elements and other aspects of the EHR and related systems. Caution is advised, however, in drawing causal inferences about the relationships among EHR variables. Nevertheless, we anticipate that EHR-CTs will play a central role in answering research and regulatory questions.

Impact of selection bias on polygenic risk score estimates in healthcare settings

BACKGROUND

Hospital-based biobanks are being increasingly considered as a resource for translating polygenic risk scores (PRS) into clinical practice. However, since these biobanks originate from patient populations, there is a possibility of bias in polygenic risk estimation due to overrepresentation of patients with higher frequency of healthcare interactions.

METHODS

PRS for schizophrenia, bipolar disorder, and depression were calculated using summary statistics from the largest available genomic studies for a sample of 24 153 European ancestry participants in the Mass General Brigham (MGB) Biobank. To correct for selection bias, we fitted logistic regression models with inverse probability (IP) weights, which were estimated using 1839 sociodemographic, clinical, and healthcare utilization features extracted from electronic health records of 1 546 440 non-Hispanic White patients eligible to participate in the Biobank study at their first visit to the MGB-affiliated hospitals.

RESULTS

Case prevalence of bipolar disorder among participants in the top decile of bipolar disorder PRS was 10.0% (95% CI 8.8-11.2%) in the unweighted analysis but only 6.2% (5.0-7.5%) when selection bias was accounted for using IP weights. Similarly, case prevalence of depression among those in the top decile of depression PRS was reduced from 33.5% (31.7-35.4%) to 28.9% (25.8-31.9%) after IP weighting.

CONCLUSIONS

Non-random selection of participants into volunteer biobanks may induce clinically relevant selection bias that could impact implementation of PRS in research and clinical settings. As efforts to integrate PRS in medical practice expand, recognition and mitigation of these biases should be considered and may need to be optimized in a context-specific manner.

Potential effect of antidepressants on remission from cocaine use disorder - A nationwide matched retrospective cohort study

BACKGROUND

Cocaine use disorder (CUD) is a significant public health issue for which there is no Food and Drug Administration-approved pharmacotherapy. Depressive disorders are common psychiatric comorbidity amongst individuals with CUD.

METHODS

A retrospective cohort study was conducted among 161,544 patients diagnosed with CUD and depression to evaluate the effectiveness of 13 antidepressants on CUD remission. For any antidepressant found to be associated with CUD remission that had an additional indication, we conducted an additional analysis to evaluate the effectiveness of the candidate drug in patients with CUD with that indication. We then analyzed publicly genomic and functional databases to identify potential explanatory mechanisms of action of the candidate drug in the treatment of CUD.

RESULTS

Among these antidepressants, bupropion was associated with higher rates of CUD remission compared to propensity-score matched patients prescribed other antidepressants: hazard ratio (HR) and 95% confidence interval (CI) 1.57 (95% CI: 1.27-1.94). Bupropion is also approved for smoking cessation. We identified CUD patients with co-occurring nicotine dependence and observed that patients prescribed bupropion displayed a higher rate of CUD remission compared to matched individuals prescribed other drugs for nicotine dependence: 1.38 (95% CI: 1.11-1.71). Genetic and functional analyses revealed that bupropion interacts with four protein-encoding genes (COMT, DRD2, SLC6A3, and SLC6A4) which are also associated with CUD and targets CUD-associated pathways including serotonergic synapses, cocaine addiction, and dopaminergic synapses.

CONCLUSIONS

Our findings suggest that bupropion might be considered a treatment for improving CUD remission in patients with CUD and co-occurring depression or nicotine dependence.

Diagnostic rate estimation from Medicare records: Dependence on claim numbers and latent clinical features

OBJECTIVE

International Classification of Disorders version 10 (ICD-10) codes contribute heavily to healthcare data. Medicare claims and other data-sources are used to constitute study populations and appraise healthcare processes. How variability in claims-per-beneficiary impacts diagnostic determinations is inadequately understood. The objective of this study is so assess distributional properties of Medicare claims, and examine claim rates impact on code utilization and rate determinations.

METHODS

The study population was Medicare beneficiaries aged 75-79.99 with claim(s) in the 5% standard analytical Carrier and Outpatient files, alive and participating in Medicare part B for all 12 months of 2017. Medicare beneficiary files were processed to create records containing all ICD-10 codes specified, key demographics, Part B and vital status, and the total claims for each 2017 beneficiary. Claim number cohorts were characterized.

RESULTS

Beneficiaries meeting inclusion criteria totaled 221,625, these having 7,617,503 claims; 96.4% had between 1 and 120 claims. Median claims were 24 for males (females 25); modal claims were 11 (13). Average distinct codes per beneficiary increased with claims number. The assignment of ICD-10 codes, i.e., 'diagnostic rate estimates' (DRE), increased as claim numbers increased for most codes among those most commonly utilized. For some conditions, mostly benign and age-related, DREs plateaued as claim numbers increased. For other conditions, typically associated with clinical acuity, e.g., chest pain, DREs increased steeply with claims.

CONCLUSIONS

Older adult Medicare beneficiaries aged 75-80 exhibited varying claims activity over the course of a year. Although DRE dependence on claim numbers varies across ICD-10 codes, rate estimates are higher for beneficiaries with claim numbers above the median.

Secondary data for global health digitalisation

Substantial opportunities for global health intelligence and research arise from the combined and optimised use of secondary data within data ecosystems. Secondary data are information being used for purposes other than those intended when they were collected. These data can be gathered from sources on the verge of widespread use such as the internet, wearables, mobile phone apps, electronic health records, or genome sequencing. To utilise their full potential, we offer guidance by outlining available sources and approaches for the processing of secondary data. Furthermore, in addition to indicators for the regulatory and ethical evaluation of strategies for the best use of secondary data, we also propose criteria for assessing reusability. This overview supports more precise and effective policy decision making leading to earlier detection and better prevention of emerging health threats than is currently the case.

Detecting and Adjusting for Hidden Biases due to Phenotype Misclassification in Genome-Wide Association Studies

With the advent of healthcare-based genotyped biobanks, genome-wide association studies (GWAS) leverage larger sample sizes, incorporate patients with diverse ancestries and introduce noisier phenotypic definitions. Yet the extent and impact of phenotypic misclassification on large-scale datasets is not currently well understood due to a lack of statistical methods to estimate relevant parameters from empirical data. Here, we develop a statistical method and scalable software, PheMED, Phenotypic Measurement of Effective Dilution, to quantify phenotypic misclassification across GWAS using only summary statistics. We illustrate how the parameters estimated by PheMED relate to the negative and positive predictive value of the labeled phenotype, compared to ground truth, and how misclassification of the phenotype yields diluted effect-sizes of variant-phenotype associations. Furthermore, we apply our methodology to detect multiple instances of statistically significant dilution in real-world data. We demonstrate how effective dilution biases downstream GWAS replication and heritability analyses despite utilizing current best practices, and provide a dilution-aware meta-analysis approach that outperforms existing methods. Consequently, we anticipate that PheMED will be a valuable tool for researchers to address phenotypic data quality issues both within and across cohorts.

Case studies in bias reduction and inference for electronic health record data with selection bias and phenotype misclassification

Electronic health records (EHR) are not designed for population-based research, but they provide easy and quick access to longitudinal health information for a large number of individuals. Many statistical methods have been proposed to account for selection bias, missing data, phenotyping errors, or other problems that arise in EHR data analysis. However, addressing multiple sources of bias simultaneously is challenging. We developed a methodological framework (R package, SAMBA) for jointly handling both selection bias and phenotype misclassification in the EHR setting that leverages external data sources. These methods assume factors related to selection and misclassification are fully observed, but these factors may be poorly understood and partially observed in practice. As a follow-up to the methodological work, we demonstrate how to apply these methods for two real-world case studies, and we evaluate their performance. In both examples, we use individual patient-level data collected through the University of Michigan Health System and various external population-based data sources. In case study (a), we explore the impact of these methods on estimated associations between gender and cancer diagnosis. In case study (b), we compare corrected associations between previously identified genetic loci and age-related macular degeneration with gold standard external summary estimates. These case studies illustrate how to utilize diverse auxiliary information to achieve less biased inference in EHR-based research.

Comparing mental and physical health of U.S. veterans by VA healthcare use: implications for generalizability of research in the VA electronic health records

OBJECTIVE

The Department of Veterans Affairs' (VA) electronic health records (EHR) offer a rich source of big data to study medical and health care questions, but patient eligibility and preferences may limit generalizability of findings. We therefore examined the representativeness of VA veterans by comparing veterans using VA healthcare services to those who do not.

METHODS

We analyzed data on 3051 veteran participants age ≥ 18 years in the 2019 National Health Interview Survey. Weighted logistic regression was used to model participant characteristics, health conditions, pain, and self-reported health by past year VA healthcare use and generate predicted marginal prevalences, which were used to calculate Cohen's d of group differences in absolute risk by past-year VA healthcare use.

RESULTS

Among veterans, 30.4% had past-year VA healthcare use. Veterans with lower income and members of racial/ethnic minority groups were more likely to report past-year VA healthcare use. Health conditions overrepresented in past-year VA healthcare users included chronic medical conditions (80.6% vs. 69.4%, d = 0.36), pain (78.9% vs. 65.9%; d = 0.35), mental distress (11.6% vs. 5.9%; d = 0.47), anxiety (10.8% vs. 4.1%; d = 0.67), and fair/poor self-reported health (27.9% vs. 18.0%; d = 0.40).

CONCLUSIONS

Heterogeneity in veteran sociodemographic and health characteristics was observed by past-year VA healthcare use. Researchers working with VA EHR data should consider how the patient selection process may relate to the exposures and outcomes under study. Statistical reweighting may be needed to generalize risk estimates from the VA EHR data to the overall veteran population.

Using genetic data to identify transmission risk factors: Statistical assessment and application to tuberculosis transmission

Identifying host factors that influence infectious disease transmission is an important step toward developing interventions to reduce disease incidence. Recent advances in methods for reconstructing infectious disease transmission events using pathogen genomic and epidemiological data open the door for investigation of host factors that affect onward transmission. While most transmission reconstruction methods are designed to work with densely sampled outbreaks, these methods are making their way into surveillance studies, where the fraction of sampled cases with sequenced pathogens could be relatively low. Surveillance studies that use transmission event reconstruction then use the reconstructed events as response variables (i.e., infection source status of each sampled case) and use host characteristics as predictors (e.g., presence of HIV infection) in regression models. We use simulations to study estimation of the effect of a host factor on probability of being an infection source via this multi-step inferential procedure. Using TransPhylo-a widely-used method for Bayesian estimation of infectious disease transmission events-and logistic regression, we find that low sensitivity of identifying infection sources leads to dilution of the signal, biasing logistic regression coefficients toward zero. We show that increasing the proportion of sampled cases improves sensitivity and some, but not all properties of the logistic regression inference. Application of these approaches to real world data from a population-based TB study in Botswana fails to detect an association between HIV infection and probability of being a TB infection source. We conclude that application of a pipeline, where one first uses TransPhylo and sparsely sampled surveillance data to infer transmission events and then estimates effects of host characteristics on probabilities of these events, should be accompanied by a realistic simulation study to better understand biases stemming from imprecise transmission event inference.

Sociodemographic characteristics differ across routine adult vaccine cohorts: An All of Us descriptive study

BACKGROUND

The National Institutes of Health All of Us (AoU) Research Program is currently building a database of 1million+ adult subjects. With it, we describe the characteristics of those with documented vaccinations.

OBJECTIVES

To describe the sociodemographic, health status, and lifestyle factors associated with vaccinations.

METHODS

This is a retrospective study involving data from the AoU program (R2020Q4R2, N = 315,297). Five vaccine cohorts [influenza, hepatitis B (HBV), pneumococcal <65 years old, pneumococcal ≥65 years old, and human papillomavirus (HPV)] were generated based on vaccination history. The influenza cohort comprised participants with documented influenza vaccinations in electronic health records (EHRs) from September 2017 to May 2018. Other vaccine cohorts comprised participants with ≥1 lifetime record(s) of vaccination documented in the EHR by December 2018. The vaccine cohorts were compared to the overall AoU cohort. Descriptive statistics were generated using EHR- and survey-based sociodemographic, health, and lifestyle information. The SAMBA (0.9.0) R package was utilized to adjust for EHR selection and outcome misclassification biases to infer sources of disparity for pneumococcal vaccinations in older adults.

RESULTS

Cohort counts were as follows: influenza (n = 15,346), HBV (n = 6323), pneumococcal <65 (n = 15,217), pneumococcal ≥65 (n = 15,100), and HPV (n = 2125). All vaccine cohorts had higher proportions of White and non-Hispanic/Latino participants compared to the overall AoU cohort. The largest differences were found in pneumococcal age ≥65, with 80.2% White participants compared to 52.9% in the overall study population. Multivariable analysis revealed that race/ethnic disparities in pneumococcal vaccination among older adults were explained by biological sex, income, health insurance, and education-related variables.

CONCLUSION

Racial, ethnic, education, and income characteristics differ across the vaccine cohorts among AoU participants. These findings inform future utilization of large health databases in vaccine epidemiology research and emphasize the need for more targeted interventions that address differences in vaccine uptake.

Causal inference under over-simplified longitudinal causal models

Many causal models of interest in epidemiology involve longitudinal exposures, confounders and mediators. However, repeated measurements are not always available or used in practice, leading analysts to overlook the time-varying nature of exposures and work under over-simplified causal models. Our objective is to assess whether - and how - causal effects identified under such misspecified causal models relates to true causal effects of interest. We derive sufficient conditions ensuring that the quantities estimated in practice under over-simplified causal models can be expressed as weighted averages of longitudinal causal effects of interest. Unsurprisingly, these sufficient conditions are very restrictive, and our results state that the quantities estimated in practice should be interpreted with caution in general, as they usually do not relate to any longitudinal causal effect of interest. Our simulations further illustrate that the bias between the quantities estimated in practice and the weighted averages of longitudinal causal effects of interest can be substantial. Overall, our results confirm the need for repeated measurements to conduct proper analyses and/or the development of sensitivity analyses when they are not available.

A Case-Crossover Phenome-wide Association Study (PheWAS) for Understanding Post-COVID-19 Diagnosis Patterns

BACKGROUND

Post COVID-19 condition (PCC) is known to affect a large proportion of COVID-19 survivors. Robust study design and methods are needed to understand post-COVID-19 diagnosis patterns in all survivors, not just those clinically diagnosed with PCC.

METHODS

We applied a case-crossover Phenome-Wide Association Study (PheWAS) in a retrospective cohort of COVID-19 survivors, comparing the occurrences of 1,671 diagnosis-based phenotype codes (PheCodes) pre- and post-COVID-19 infection periods in the same individual using a conditional logistic regression. We studied how this pattern varied by COVID-19 severity and vaccination status, and we compared to test negative and test negative but flu positive controls.

RESULTS

In 44,198 SARS-CoV-2-positive patients, we foundenrichment in respiratory,circulatory, and mental health disorders post-COVID-19-infection. Top hits included anxiety disorder (p=2.8e-109, OR=1.7 [95%CI: 1.6-1.8]), cardiac dysrhythmias (p=4.9e-87, OR=1.7 [95%CI: 1.6-1.8]), and respiratory failure, insufficiency, arrest (p=5.2e-75, OR=2.9 [95%CI: 2.6-3.3]). In severe patients, we found stronger associations with respiratory and circulatory disorders compared to mild/moderate patients. Fully vaccinated patients had mental health and chronic circulatory diseases rise to the top of the association list, similar to the mild/moderate cohort. Both control groups (test negative, test negative and flu positive) showed a different pattern of hits to SARS-CoV-2 positives.

CONCLUSIONS

Patients experience myriad symptoms more than 28 days after SARS-CoV-2 infection, but especially respiratory, circulatory, and mental health disorders. Our case-crossover PheWAS approach controls for within-person confounders that are time-invariant. Comparison to test negatives and test negative but flu positive patients with a similar design helped identify enrichment specific to COVID-19. This design may be applied other emerging diseases with long-lasting effects other than a SARS-CoV-2 infection. Given the potential for bias from observational data, these results should be considered exploratory. As we look into the future, we must be aware of COVID-19 survivors' healthcare needs.

Identifying who has long COVID in the USA: a machine learning approach using N3C data

BACKGROUND

Post-acute sequelae of SARS-CoV-2 infection, known as long COVID, have severely affected recovery from the COVID-19 pandemic for patients and society alike. Long COVID is characterised by evolving, heterogeneous symptoms, making it challenging to derive an unambiguous definition. Studies of electronic health records are a crucial element of the US National Institutes of Health's RECOVER Initiative, which is addressing the urgent need to understand long COVID, identify treatments, and accurately identify who has it-the latter is the aim of this study.

METHODS

Using the National COVID Cohort Collaborative's (N3C) electronic health record repository, we developed XGBoost machine learning models to identify potential patients with long COVID. We defined our base population (n=1 793 604) as any non-deceased adult patient (age ≥18 years) with either an International Classification of Diseases-10-Clinical Modification COVID-19 diagnosis code (U07.1) from an inpatient or emergency visit, or a positive SARS-CoV-2 PCR or antigen test, and for whom at least 90 days have passed since COVID-19 index date. We examined demographics, health-care utilisation, diagnoses, and medications for 97 995 adults with COVID-19. We used data on these features and 597 patients from a long COVID clinic to train three machine learning models to identify potential long COVID among all patients with COVID-19, patients hospitalised with COVID-19, and patients who had COVID-19 but were not hospitalised. Feature importance was determined via Shapley values. We further validated the models on data from a fourth site.

FINDINGS

Our models identified, with high accuracy, patients who potentially have long COVID, achieving areas under the receiver operator characteristic curve of 0·92 (all patients), 0·90 (hospitalised), and 0·85 (non-hospitalised). Important features, as defined by Shapley values, include rate of health-care utilisation, patient age, dyspnoea, and other diagnosis and medication information available within the electronic health record.

INTERPRETATION

Patients identified by our models as potentially having long COVID can be interpreted as patients warranting care at a specialty clinic for long COVID, which is an essential proxy for long COVID diagnosis as its definition continues to evolve. We also achieve the urgent goal of identifying potential long COVID in patients for clinical trials. As more data sources are identified, our models can be retrained and tuned based on the needs of individual studies.

FUNDING

US National Institutes of Health and National Center for Advancing Translational Sciences through the RECOVER Initiative.

Bias reduction and inference for electronic health record data under selection and phenotype misclassification: three case studies

Electronic Health Records (EHR) are not designed for population-based research, but they provide access to longitudinal health information for many individuals. Many statistical methods have been proposed to account for selection bias, missing data, phenotyping errors, or other problems that arise in EHR data analysis. However, addressing multiple sources of bias simultaneously is challenging. Recently, we developed a methodological framework (R package, SAMBA ) for jointly handling both selection bias and phenotype misclassification in the EHR setting that leverages external data sources. These methods assume factors related to selection and misclassification are fully observed, but these factors may be poorly understood and partially observed in practice. As a follow-up to the methodological work, we explore how these methods perform for three real-world case studies. In all three examples, we use individual patient-level data collected through the University of Michigan Health System and various external population-based data sources. In case study (a), we explore the impact of these methods on estimated associations between gender and cancer diagnosis. In case study (b), we compare corrected associations between previously identified genetic loci and age-related macular degeneration with gold standard external estimates. In case study (c), we evaluate these methods for modeling the association of COVID-19 outcomes and potential risk factors. These case studies illustrate how to utilize diverse auxiliary information to achieve less biased inference in EHR-based research.