Linying Zhang

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.

Objective: To determine how cohort versus nested case-control (NCC) design affects EHR-based prediction of a rare outcome. Materials and Methods: We compared two cohort and five NCC designs for predicting early-onset gastrointestinal cancer using L1-regularized logistic regression (LASSO), XGBoost, and multilayer perceptron (MLP). Models underwent nested five-fold cross-validation with model-specific hyperparameter tuning and temporal evaluation in a held-out prospective cohort representing intended deployment. Model discrimination was assessed using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). Model calibration was assessed using Brier scores and calibration plots before and after recalibration with isotonic regression. Feature-ranking similarity was quantified using rank-biased overlap. Results: The cohort development datasets included 256,854-415,529 patients, with 2-yr cumulative incidence of 0.12%-0.17%, compared with approximately 9% in the 1:10 matched NCC datasets. NCC designs generally produced higher internal discrimination, but temporal performance varied by algorithm and index-date definition. Cohort designs achieved the highest temporal AUPRC for LASSO (0.068), whereas diagnosis-anchored NCC achieved the highest temporal AUPRC for XGBoost (0.086) and MLP (0.070). Temporal AUPRC generally decreased as the required lead time before diagnosis increased. Raw predicted probabilities substantially overestimated observed risks in the temporal cohort. Isotonic recalibration improved probability scaling while largely preserving discrimination, although residual miscalibration remained. Cohort-based feature rankings were more consistent across algorithms, whereas NCC-based rankings depended more strongly on algorithm and index-date anchoring. Conclusion: Study design should be selected and evaluated according to the intended clinical prediction task and deployment setting.

Objectives: To investigate the longitudinal effects of an ambient artificial intelligence (AI) scribe use in primary care on outcomes related to documentation time, note writing patterns, and clinician financial productivity. Materials and Methods: We conducted a quasi-experimental study with an interrupted time-series design to ascertain the effect of AI scribe use on note writing time, time outside of scheduled hours (TOSH), pajama time (5:30PM-7:00aM), note length, note closure <24 h, and work relative value units (wRVU). For each primary care clinician (physician, advanced practice provider), date of initial access was day 0; each measure was calculated for 150 days prior (pre-period) and after (post-period) day zero. Results: Two hundred and twenty clinicians (mean age = 43.7, 70.9% females; 56.4% physicians) from 36 clinics, conducting 314 845 patient encounters were included. Compared to the pre-period, there was 7% decrease in note writing time at day 0 (incidence rate ratio [IRR] 0.93, 95% credible interval [CI] [0.89, 0.98]), increasing to 16% decrease by day 150 (IRR 0.84, 95% CI [0.82, 0.87]). There was no change in pajama time, TOSH, or wRVU at day 0; however, at day 150, there was a 19% decrease in pajama time (0.81, 95% CI [0.73, 0.90]), 13% decrease in TOSH (0.87, 95% CI [0.77, 0.98]), and a 2% increase in wRVU (1.02, 95% CI [1.01, 1.04]). At day 0, there was a 5% increase in note closures (1.05, 95% CI [1.02, 1.08]), with no difference to pre-period levels by day 150. There were no changes in note length at day 0 or day 150. Discussion and Conclusion: There were persistent changes in AI scribe use measures over time as clinicians situated these tools within their workflows.

Timely initiation of medications for opioid use disorder (MOUD) remains a critical gap in the health-system response to the opioid crisis. Although social determinants of health (SDoH) influence treatment access, their impact may vary across care settings owing to differences in clinical workflows, regulatory constraints, and system-level processes. In this retrospective cohort study, we linked electronic health records from an integrated midwestern U.S. health system with neighborhood-level SDoH and area deprivation data to examine time to MOUD initiation among patients diagnosed with opioid use disorder between 2020 and 2024. Using inverse-probability-of-censoring-weighted Cox proportional hazards models with care-setting–specific effects, we found that SDoH associations with MOUD initiation were concentrated in the outpatient setting and differed significantly across settings: a greater neighborhood share of racial or ethnic minority residents was associated with slower outpatient initiation, whereas inpatient initiation was driven primarily by clinical complexity and emergency-department initiation showed limited sensitivity to neighborhood context. These patterns were robust to adjustment for informative censoring and to follow-up length. By linking geospatial social context to routine EHR data, this approach enables scalable, care-setting–specific surveillance of treatment inequities and can inform resource allocation to accelerate equitable delivery of addiction treatment.

Purpose: To evaluate whether fluorescence lifetime imaging ophthalmoscopy (FLIO) combined with deep learning can detect metabolic signatures for classification of type 2 diabetes mellitus (T2DM). Design: Cross-sectional analysis of participants included AI-READI dataset (version 3) with FLIO imaging and and hemoglobin A1c (HbA1c) measurement. Subjects: 1,783 participants from the AI-READI dataset (version 3) with HbA1c measurements and FLIO imaging scans (6,912 total): 671 normoglycemic, 726 prediabetic, and 386 diabetic. Methods: Mean fluorescence lifetime maps were generated using a center-of-mass approach and used as inputs to AI models. We trained convolutional neural networks (CNNs), ResNet-18, and XGBoost under three-class (normal, prediabetic, diabetic) and two binary (normal vs. impaired; normal vs. diabetic) classification schemes, using nested 5-fold cross-validation with participant-level grouping. Main Outcome Measures: Macro-averaged accuracy, F1 score, area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and positive predictive value (PPV). Results: Group-averaged lifetime maps demonstrated consistent spatial differences across glycemic groups, with progressively longer lifetimes from normal to diabetic participants. The CNN achieved the best overall performance in the 3-class classification (accuracy 0.41 ± 0.03, F1 score 0.39 ± 0.02, AUROC 0.58 ± 0.02), compared to the random classifier for 3-class classification (AUROC = 0.50; accuracy = F1 = 0.33). ResNet-18 and XGBoost showed similar performance (AUROC 0.53–0.58). Confusion matrices revealed substantial overlap between classes, with frequent misclassification toward the prediabetes group. Binary reformulation (normal vs. diabetic) improved performance substantially, with the CNN resulting in AUROC 0.63 ± 0.02 and XGBoost 0.67 ± 0.07. Conclusions: FLIO-derived lifetime maps capture metabolic signals associated with glycemic status but yield modest classification performance with current AI models. These findings highlight both the potential and the challenges of using FLIO for early metabolic screening and monitoring, informing future development of clinically applicable imaging biomarkers.

Background: Large-scale propensity score (LSPS) models are increasingly used to control confounding in observational studies, but their reliability in small-sample settings is unclear. Small samples can limit a study’s ability to estimate propensity scores accurately and achieve adequate covariate balance, raising concerns about insufficient confounding adjustment. Resulting in small datasets being excluded, despite potentially containing valid information. Methods: We evaluated the balancing and bias reduction performance of LSPS models using six target-comparator pairs. To emulate a distributed data network analysis setting, we partitioned two large real-world data sources into smaller subsets. Within each subset, LSPS models were fit locally and evaluated on covariate balance, treatment effect bias and precision after empirical calibration, using real negative controls and synthetic positive controls. Performance was compared to a global LSPS model trained on the full dataset. Results: Across most target-comparator pairs, LSPS models substantially reduced bias and, when applying empirical calibration, improved precision. PS matching was more resilient to small sample sizes than PS stratification, achieving adequate balance even at n = 500, while stratification sometimes failed at n = 4000. A slight bias increase was observed at the smallest sample sizes, though not universally. Standard balance diagnostics consistently failed below n = 20 000, while a recently proposed diagnostic accounting for chance imbalance did not. Conclusions: LSPS models generally provide reliable bias reduction in small-sample settings, supporting their use in federated analyses. However, standard balance diagnostics may be misleading in small samples, and alternatives should be considered, such as significance checking. When LSPS fails to reduce bias adequately, additional adjustment strategies are required.

Retinal fluids, detectable through optical coherence tomography (OCT), are key biomarkers for retinal diseases such as diabetic macular edema and age-related macular degeneration, guiding treatment decisions and monitoring response to therapy. Automated segmentation of retinal fluids could support large-scale clinical research and the development of clinical decision support tools. Recent ophthalmic foundation models trained on massive retinal imaging datasets show promise across many downstream tasks, including disease risk prediction and biomarker segmentation, but their performance relative to task-specific models for specialized clinical tasks remains unclear. We compared a task-specific segmentation model (RetiFluidNet) and an ophthalmic foundation model (VisionFM) using a standard benchmarking dataset containing 4,248 OCT images from 48 patients with three retinal diseases. Models were evaluated using three-fold cross-validation and assessed for pixel-level segmentation accuracy and patient-level fluid burden estimation. The task-specific model achieved higher segmentation performance and more consistent fluid quantification across devices. These findings suggest that, for retinal fluid segmentation, specialized task-specific models currently remain more reliable than general-purpose foundation models, highlighting the need for targeted adaptation before clinical deployment.

Estimating heterogeneous treatment effects (HTE) from observational health data is essential for precision medicine, yet existing methods often struggle with high-dimensional covariates and time-to-event outcomes common in electronic health records (EHRs). We propose SurvTEDVAE, a disentangled variational autoencoder designed for causal survival analysis. The model learns latent representations corresponding to instrumental factors, confounders, and outcome-dependent risk factors, and integrates a survival likelihood to model time-to-event outcomes with censoring. The learned representations are used to estimate conditional average treatment effects using downstream causal estimators. We evaluated SurvTEDVAE using a semi-synthetic ACTG dataset and a high-dimensional EHR-based hypertension cohort with over 20,000 covariates. Across both datasets, SurvTEDVAE achieved lower estimation error for heterogeneous treatment effects compared with meta-learning and causal survival forest approaches. These results demonstrate that disentangled representation learning can improve causal effect estimation for survival outcomes in high-dimensional real-world health data.