George Hripcsak

Objective: Propensity score (PS) models are commonly used for addressing confounding in observational studies. Researchers are increasingly incorporating large numbers of covariates into PS models, using techniques like large-scale propensity score (LSPS) adjustment. It remains unclear if the inclusion of many covariates affects the subsequent choice of adjustment strategy. Methods: In this paper, we evaluate commonly used adjustment strategies using large-scale models, including matching, stratification, and inverse probability of treatment weighting (IPTW), to estimate treatment effects under a new-user cohort design. Evaluations employ real-world data across four national healthcare databases and incorporate 3,840 treatment effect estimates per model specification for 24 different treatment cohort comparisons against 160 negative control outcomes. We assess adjustment strategies for covariate balance between treatment cohorts and type-I and type-II error, mean-squared error, and precision of the adjusted effect estimate. Results: Across all model specifications, IPTW with symmetric Crump trimming and 1:1 matching perform best overall, achieving strong balance and low type-I error. Additionally, empirical calibration of effect estimates can substantially reduce differences between strategies. Discussion: No single strategy is clearly superior in all settings. We recommend incorporating the top-performing strategies at least as sensitivity analyses in studies even if researchers anticipate that some other strategy may excel due to expectation of their study’s operating characteristics. We also recommend use of diagnostics like covariate balance to check strategy performance. Conclusion: IPTW with Crump trimming and 1:1 matching are strong default choices for large-scale PS adjustment, but strategy selection should be guided by study context and validated with balance and error diagnostics. We broadly recommend empirical calibration to reduce sensitivity to model choice.

Background: Large-scale international real-world evidence generation benefits from terminology harmonization. Despite widespread adoption of standardized vocabularies, their effective use and long-term sustainability at scale remain poorly understood. Objective: This study aimed to examine real-world terminology and code use in data across a federated network of observational data sources. Methods: We conducted a 2-part survey of researchers and data owners within the Observational Health Data Sciences and Informatics community on their terminology use, challenges, and needs, accompanied by the analysis of code use across a subset of real-world data sources. Results: The survey covered 144 institutions across the United States, the United Kingdom, Europe, Asia, and Africa. Data on terminology use covered 60 sources, including 22 data sources that provided detailed code-level use information. We observed significant variations in terminology use, with 61 out of 89 terminologies used in the data present in less than 10% of the data sources. Code use even after data harmonization was also highly variable: less than 1% (95/742,337) of codes were found in all data sources. Mapping and hierarchy completeness, terminology coverage, versioning, and terminology changes were among the most common challenges. We outlined several of our subsequent process improvements: community contribution and stewardship pipelines, metadata for relationships, and informatics tools for assessment of the impact of terminology change. Conclusions: Terminology and coding inconsistencies across observational data sources require a standardized terminology system. Such a system is complex and time-consuming and needs community contribution and informatics solutions for harmonization to be scalable and sustainable. Even with a common reference standard, high heterogeneity of terminology and code use across different observational data sources remains.

Clinical decisions for determining optimal patient-specific interventions are complicated prediction tasks that rely on health care professionals’ understanding of physiological mechanisms and their dynamics. These decisions are challenged by (a) observational data sparsity and (b) patient heterogeneity. Here, we focus on estimating and forecasting specific physiological properties—that are not explicitly present in clinical observations—to provide additional features using only data available bedside at the time of decision-making. Mechanistic models of physiological system(s), e.g., physiological ordinary differential equation (ODE) models, provide pathways to compensate for data sparsity by synchronizing the model with observations of an individual patient using data assimilation (DA). However, DA used in a standard computational workflow to estimate constant model parameters from presently-known data is less effective at optimizing state forecasts of the model governed by physiological processes that evolve before new observations are available. Stated simply, we cannot forecast the future evolution of the model because we cannot forecast model parameters. To support next-generation clinical decision support, we develop a new DA and machine learning (ML) hybrid pipeline to estimate and forecast individual future physiological processes by forecasting ODE model parameters. This pipeline overcomes model and DA workflow limitations by stacking a DA-estimated posterior empirical distribution of physiological parameters with longitudinal ML forecasting models. We work within the context of glycemic management in an ICU using EHR data to construct and test a use case. We use synthetic data and real-world clinical data to validate the integrated pipeline and quantify uncertainties.

Objective: This study compares the contents of two data standards; the Observational Medical Outcomes Partnership (OMOP) and Fast Healthcare Interoperability Resources (FHIR), highlighting their strength and weaknesses and serve as an initial step toward understanding how each standard supports secondary data analysis. Materials and Methods: Participant electronic health record data in both OMOP and FHIR formats from the All of Us Research Program (AoURP) were compared, including codeable event volume, healthcare encounters, and person timelines. A phenotype-based assessment was also conducted using Type-II Diabetes Mellitus (T2DM). Results: Among 29 512 participants identified with overlapping FHIR and OMOP data, Median codeable event counts were comparable between FHIR and OMOP within the Measurement (OMOP = 846; FHIR = 832), Drug (OMOP = 92; FHIR = 90), and Observation (OMOP = 65; FHIR = 100) domains, but were higher in OMOP within the Condition (OMOP = 258; FHIR = 11) and Procedure (OMOP = 72; FHIR = 4) domains. Within the T2DM cohort, OMOP contained more data, except for medications. Very few participants had encounters in FHIR (1.5%) relative to OMOP (97.0%). On average, only 15.9% of visit dates overlapped in both standards, with most visit dates occurring only in OMOP (65.3%) or only FHIR (18.4%). Discussion: Data in OMOP showed a higher volume of observation and procedure codeable events, reported encounters, and T2DM symptoms, complications, and comorbidities. FHIR data was able to capture data across multiple providers and health systems. Conclusion: Based on AoURP data, both standards were shown to support healthcare data capture, although OMOP shows greater utility for research purposes as its extract, transform, and load process enables more flexible data capture relative to extracting data from FHIR payloads. FHIR, however, captures patient-level data from beyond the health system and can thus be used to supplement OMOP.

Aims: Understanding heterogeneous responses to glucose‐lowering therapies is crucial for advancing personalized treatment and optimizing outcomes in type 2 diabetes mellitus (T2DM). While average treatment effects are established for many drug classes, responses may vary by clinical and demographic factors. Here, we evaluated whether major glucose‐lowering drug classes exhibit heterogeneous treatment effects (HTE) across key patient characteristics. Materials and Methods: This large‐scale observational cohort study was replicated in six data‐sources across the Observational Health Data Sciences and Informatics network. New‐user, active‐comparator cohorts were constructed for patients with T2DM initiating one of the nine antihyperglycemic drug classes. Large‐scale propensity score adjustment, empirical calibration using negative controls, and random‐effects meta‐analysis were used to estimate calibrated hazard ratios (HRs). HTE was assessed by comparing differences in log HRs across clinical and demographic subgroups. Results: Nominal signals of HTE were observed in the hyperlipidemia, hypertension, obesity, and sex subgroups: Biguanides (vs. DPP‐4 inhibitors) were associated with lower risk of acute myocardial infarction in hyperlipidemia and heart failure hospitalization in obesity. SGLT‐2 inhibitors (vs. GLP‐1 receptor agonists) were associated with reduced stroke risk only in non‐obese patients. Sex‐specific patterns included higher risk of diarrhoea with GLP‐1 receptor agonists and lower risk of stroke with SGLT‐2 inhibitors (both vs. DPP‐4 inhibitors) in women. None of the signals demonstrated statistical significance after multiple testing correction. Several subgroups (e.g., diabetic ketoacidosis and retinopathy) did not meet diagnostic criteria for clinical reliability and were not evaluated. Conclusions: This hypothesis‐generating study identified limited signals of HTE. These findings are exploratory and require cautious interpretation and validation.

BACKGROUND: Proton pump inhibitors (PPIs) may potentially reduce clopidogrel’s antiplatelet effect and increase cardiovascular risk. The degree of CYP (cytochrome P450) 2C19 inhibition varies among PPIs. Few studies have evaluated individual PPIs by CYP2C19 inhibition strength across countries. This study aimed to compare the incidence of cardiovascular events between strong CYP2C19-inhibiting potency and weak CYP2C19-inhibiting potency (weak or non-CYP2C19–inhibiting PPIs) in patients receiving clopidogrel. METHODS: We conducted an international observational cohort study using 14 databases from the United States, South Korea, and Taiwan. We included patients aged ≥18 years who received clopidogrel with PPIs from 1985 to 2023. PPIs were classified into strong CYP2C19-inhibiting PPIs and weak or non-CYP2C19–inhibiting PPIs based on CYP2C19 inhibition. We compared the hazard ratios and 95% CIs for major adverse cardiovascular events, including myocardial infarction, stroke, and cardiovascular mortality, using the Cox proportional hazards model after 1:1 propensity score matching. Secondary outcomes included cardiovascular mortality, myocardial infarction, stroke, and all-cause mortality. A random-effects model calculated pooled hazard ratios and 95% CIs. Only databases meeting all diagnostic criteria were included in the meta-analysis. RESULTS: Large-scale propensity score matching identified 166 005 patient pairs. During the 365-day follow-up, the risk of major adverse cardiovascular events did not differ significantly between patients receiving clopidogrel plus strong CYP2C19-inhibiting PPIs and those receiving clopidogrel plus weak or non-CYP2C19–inhibiting PPIs (17.63 per 1000 person-years versus 16.82 per 1000 person-years; calibrated hazard ratio, 1.00 [95% CI, 0.79–1.26]). No significant difference was observed in the risk of secondary outcomes (calibrated hazard ratio, cardiovascular mortality 1.10 [95% CI, 0.87–1.39], myocardial infarction 0.98 [95% CI, 0.81–1.19], stroke 1.05 [95% CI, 0.87–1.27], and all-cause mortality 1.18 [95% CI, 0.93–1.50]). CONCLUSIONS: Concomitant use of clopidogrel and strong CYP2C19-inhibiting PPIs was not associated with a higher cardiovascular risk compared with concomitant use of clopidogrel and weak or non-CYP2C19–inhibiting PPIs. This large-scale study does not support the clinical significance of potential interactions between PPIs and clopidogrel.

Summary: Advances in multimodal longitudinal data and artificial intelligence (AI) create new opportunities for cancer etiology research. We envision an AI-powered discovery workflow integrating an interoperable epidemiologic data ecosystem and causal inference frameworks to accelerate the identification of both cancer causes and the converging biological states for prevention.

Importance Women have been under-represented in clinical trials of type 2 diabetes mellitus (T2D), and evidence on sex differences in effectiveness of T2D treatments remains limited. Objective To assess sex differences in comparative effectiveness and safety of four second-line antidiabetic agents: glucagon-like peptide-1 receptor agonists (GLP-1RA), sodium-glucose cotransporter-2 inhibitors (SGLT2i), dipeptidyl peptidase-4 inhibitors (DPP4i), and sulfonylureas (SU). Design Retrospective cohort study using an active-comparator new-user design, following each participant till treatment discontinuation or end of data. Setting Multinational study across ten real-world databases from the Observational Health Data Sciences and Informatics (OHDSI) network in the United States, United Kingdom, Germany, and Spain. Participants 5.15 million adults with T2D who initiated one of the four second-line therapies following metformin during 1992-2021. Exposures GLP-1RA, SGLT2i, DPP4i, or SU. Main Outcomes and Measures Cardiovascular effectiveness as measured through 7 outcomes (major adverse cardiovascular events and glycemic control) and safety through 18 outcomes as highlighted by ADA guideline. Hazard ratios (HRs) are estimated separately for women and men using propensity score-stratified Cox models with empirical calibration. Sex differences were tested using Z-tests on log-HR differences. Results Drug initiation rates differed by sex with 9.28% of women initiating on GLP-1RA, 11.91% SGLT2i, 27.81% DPP4i, and 50.99% SU; the rates among the men were 5.41%, 12.84%, 24.64%, and 57.10%. No significant sex differences were observed for cardiovascular effectiveness outcomes. Several safety outcomes showed significant sex differences that are consistent across drug comparisons. Focusing on GLP-1RA compared to SGLT2i for brevity, GLP-1RA users experienced the following comparative benefits and risks: higher risk of acute pancreatitis among women (HR 1.39 [1.13, 1.70]) while non-differential risk among men (HR 0.91 [0.74, 1.12]) with p = 0.005 for the test of difference; non-differential risk of hypotension among women (HR 1.08 [0.98, 1.19]) while lower risk among men (HR 0.87 [0.78, 0.96]) with p = 0.003. Where no sex differences were found, our findings were consistent with existing evidence. Conclusions and Relevance This large-scale multinational study on antidiabetic agents identified clinically relevant sex differences, which are biologically plausible but previously lacked clinical evidence. Our findings reinforce the importance of tailoring T2D management according to sex.

Propensity score adjustment is commonly used in observational research to address confounding. Controversy persists about how to select covariates as possible confounders to generate the propensity model. A desire to include all possible confounders is offset by a concern that more covariates will augment bias or increase variance. Much of concern is over instruments, which are variables that affect the treatment but not the outcome. Adjusting for an instrument has been shown to increase bias due to unadjusted confounding and to increase the variance of the effect estimate. Large-scale propensity score (LSPS) adjustment includes most available pre-treatment covariates in its propensity model. It addresses instruments with a pair of diagnostics, ceasing the analysis if any covariate exceeds a correlation coefficient of 0.5 with the treatment and checking for an aggregation of instruments with equipoise reported as a preference score. Our simulation assesses the impact of adjusting for instruments in the context of LSPS’s diagnostics. In our simulation, even when the variance of the treatment contributed by the adjusted instrument(s) exceeds an unadjusted confounder by over twenty-fold, when the correlation between the instrument(s) and the treatment was less than 0.5 and the equipoise was greater than 0.5, the additional shift in the effect estimate due to adjusting for the instrument(s) was less than the shift due to confounding by itself. Therefore, we find in this simulation that adjusting for instruments contributed a minor amount of bias to the effect estimate. This simulation aligns well with a previous assessment of the impact of adjusting for instruments and with separate empirical evidence that adjusting for many covariates surpasses attempts to identify a limited set of confounders.