Ting Ye

BACKGROUND: Ischemic stroke preventive therapies have been slow to emerge due to a limited understanding of the underlying molecular mechanisms. We aimed to discover new protein risk factors that are potential etiologic factors and may be candidates for therapeutic intervention. METHODS: We included ARIC study (Atherosclerosis Risk in Communities) and CHS (Cardiovascular Health Study) participants without prevalent stroke at the time of proteomic profiling (1990–1992 in ARIC and 1992–1993 in CHS). Relative concentrations of proteins in blood were assessed with aptamers (n=4954). Ischemic stroke events were adjudicated through 2019 in ARIC and 2015 in CHS. Study-specific Cox regression adjusting for ischemic stroke risk factors was used to estimate associations between proteins and incident ischemic stroke. Study-specific estimates were combined using fixed-effects meta-analysis. Replication was conducted in 2 independent cohorts (AGES-Reykjavik, UK Biobank). Proteins with a false discovery rate q value P −5 ), of which 26 are novel. In 2 independent cohorts, 37 (84%) associations were directionally concordant, and 26 (59%) met criteria for replication. Mendelian randomization analyses suggested MMP12 (matrix metallopeptidase-12) and CNTN1 (contactin-1) may be causally related to ischemic stroke risk. In multivariable Mendelian randomization analyses, higher MMP12 levels had a direct protective effect on ischemic stroke risk, while CNTN1 had a direct harmful effect. CONCLUSIONS: This proteomic and genomic investigation of ischemic stroke identified novel risk biomarkers and 2 potential etiologic CNTN1 factors that are regulated in response to atherosclerosis and obesity. If future studies confirm our findings, these results may have implications for the development of new preventive therapies.

Electronic health records (EHRs) are a primary target for foundation models because they capture longitudinal, multimodal, and large-scale clinical trajectories. However, routinely collected clinical data remain difficult to model due to extreme sparsity, irregular temporal sampling, heterogeneous representations, and informative missingness. In this Perspective, we provide an EHR-centered synthesis of foundation models across the translational pipeline, spanning data resources, model architectures, and the requirements for reliable clinical integration. We first examine the data characteristics and real-world barriers that shape EHR modeling, including coding variation and distribution shift. We then review the three dominant model families, including structured-sequence, clinical language, and multimodal models, with a focus on architectural adaptations for longitudinal clinical reasoning. Next, we summarize emerging clinical use cases, from outcome prediction and phenotyping to assistive documentation, digital-twin simulation, and causal inference. Finally, we delineate the engineering and governance conditions necessary for clinical translation, including evaluation under distribution shift, workflow-integrated monitoring, and regulatory oversight. By linking technical advances to the practical constraints of healthcare delivery, this Perspective outlines a roadmap for transitioning EHR foundation models from research benchmarks to safe, impactful clinical technologies.

Abnormal cardiac atrial structure and function, termed atrial cardiopathy, typically precedes atrial fibrillation and downstream cardiovascular complications, yet detection is limited by the cost and accessibility of high-quality cardiac imaging. Here we show that a deep learning model trained on 12-lead electrocardiograms paired with 21,749 cardiac magnetic resonance scans from the UK Biobank predicts left atrial structure and function. Model-derived measures of atrial cardiopathy are strongly associated with new-onset atrial fibrillation, heart failure, and ischemic stroke after adjustment for clinical risk factors and biomarkers in two external cohorts, with magnitudes comparable to or greater than those for direct imaging measures and clinical risk factors. The risk of cardioembolic stroke, the hallmark complication of atrial fibrillation, increases 66% per standard deviation of left atrial volume. In exploratory analyses, model measures predict cardiac monitor-detected atrial fibrillation more accurately than a clinical risk prediction tool and NT-proBNP levels. This model is an inexpensive, accessible tool that identifies individuals at high-risk for atrial fibrillation and related complications.

We apply covariate adjustment to the Wilcoxon two sample statistic and Wilcoxon–Mann–Whitney test in comparing two treatments. The covariate adjustment through calibration not only improves efficiency in estimation/inference but also widens the application scope of the Wilcoxon two sample statistic and Wilcoxon–Mann–Whitney test to situations where covariate‐adaptive randomization is used. We motivate how to adjust covariates to reduce variance, establish the asymptotic distribution of adjusted Wilcoxon two sample statistic, and provide explicitly the guaranteed efficiency gain. The asymptotic distribution of adjusted Wilcoxon two sample statistic is invariant to all commonly used covariate‐adaptive randomization schemes so that a unified formula can be used in inference regardless of which covariate‐adaptive randomization is applied.

Covariate adjustment uses baseline prognostic variables to improve the precision of treatment effect estimates. Recent Food and Drug Administration guidance and scientific consensus emphasize three principles for its use, namely estimand-focused analyses, assumption-lean robustness, and fit-for-purpose variance estimation. Despite substantial methodological progress, practical guidance for trial practitioners remains fragmented. We review covariate adjustment strategies for continuous, discrete, and time-to-event endpoints in randomized trials that adhere to these three principles. We show how unadjusted estimators, as well as linear and non-linear adjusted estimators, can be viewed as special cases of the general augmented inverse probability weighting framework. For time-to-event endpoints, we describe how covariate adjustment can be applied to Kaplan-Meier estimators, log-rank tests, and estimation of the unconditional hazard ratio without altering the estimand or introducing additional assumptions. We also synthesize recent developments in multi-arm trials, covariate-adaptive randomization, data-adaptive covariate selection, and covariate adjustment in interim analyses, and we provide practical insights for implementation. Covariate-adjusted estimators target the same marginal estimands as unadjusted analyses but typically achieve greater efficiency. Linear adjustment with Analysis of Heterogeneous Covariance guarantees asymptotic efficiency gains under minimal assumptions. Augmented inverse probability weighting generalizes covariate adjustment to flexible modeling frameworks and remains consistent even under model misspecification. For survival analysis, covariate-adjusted versions of the log-rank test and Cox model improve power without altering the estimand or requiring additional assumptions. Properly accounting for covariate-adaptive randomization is essential for valid inference. The reviewed methods are implemented in the RobinCar family of R packages: RobinCar and RobinCar2. Covariate adjustment is a principled and practical approach for improving trial efficiency, aligned with current regulatory guidance. By adhering to the principles of estimand-focus, assumption-lean robustness, and fit-for-purpose variance estimation, practitioners can apply covariate adjustment with confidence across diverse trial settings. Further work on evaluating finite-sample performance and re-analyses of completed trials will deepen understanding of covariate adjustment in practice.

We investigate adaptive designs that can be used to take advantage of increases in efficiency from the use of the covariate‐adjusted log‐rank test in trials with time‐to‐event endpoints. These adaptive designs are intended to address a key practical challenge in taking advantage of efficiency gains from this test, which is that the actual efficiency gain attained in a trial may differ from estimates of the efficiency gain at the design stage. We evaluate information‐based interim monitoring and blinded event target adjustment (BETA) as tools for improving efficiency, focusing on statistical and operational trade‐offs between these approaches. We show using two different data‐generating processes that regression coefficients used in the construction of the covariate‐adjusted log‐rank test may increase over time. As a result, variance reductions from adjustment with the covariate‐adjusted log‐rank test may also increase with additional follow‐up in the trial. This means that BETA, which estimates variance reductions at an interim time‐point in order to decide the timing of interim and final analyses, may not fully take advantage of the attainable efficiency gains with covariate adjustment. Simulations that incorporate repeated testing compare trial designs in terms of power, duration, and sample size reduction. While information‐based monitoring enables faster analyses when covariates are highly prognostic, it may be operationally burdensome. BETA offers logistical simplicity but may not fully realize the potential efficiency gains from covariate adjustment. Reducing sample size with either adaptive design is potentially risky, as losses from a longer trial duration arising from over‐optimistic estimates of efficiency gains at the design stage may be much greater than savings from a smaller trial.

Adjusting for prognostic baseline covariates in the design and analysis of randomized controlled trials can increase statistical power and improve the precision of treatment effect estimates by reducing outcome variability. In this paper, we describe how the covariate-adjusted log-rank test can be used to enhance the power of late-phase oncology studies with primary time-to-event endpoints. Using a case study of a metastatic castration-resistant prostate cancer trial with overall survival as the primary endpoint, we demonstrate how to estimate variance reductions achievable through different covariate adjustment strategies, including adjustment for a prognostic score, based on historical data. We also compare these strategies with respect to Type I error rate control. We further describe a method that can be used to prevent Type I error rate inflation that can arise when naively embedding covariate adjustment within a group sequential design. A simulation study illustrates the performance of this method across scenarios with different prognostic covariate strength and different numbers of events, randomization strata, and interim analyses. Finally, we present a re-analysis of the overall survival endpoint from a trial in metastatic colorectal cancer using covariate adjustment in a group sequential design.

Atrial fibrillation and heart failure impose substantial health burdens worldwide, yet existing prediction models lack sufficient accuracy and generalizability. We developed CARDIAC-FM, a multimodal foundation model that learns joint representations of 12-lead electrocardiogram (ECG) and cardiac magnetic resonance imaging (MRI) through contrastive learning. We trained CARDIAC-FM on 57,609 paired ECG-cardiac MRI samples from UK Biobank and evaluated it in two external cohorts: the Cardiovascular Health Study (CHS) and the Multi-Ethnic Study of Atherosclerosis (MESA). CARDIAC-FM consistently outperformed unimodal models across all cohorts, and jointly incorporating ECG features with established clinical risk scores yielded additive gains in discrimination, indicating that ECG and traditional risk factors capture complementary dimensions of cardiovascular risk. The learned representations improved prediction across a range of cardiovascular outcomes with minimal task-specific fine-tuning, reflecting real-world settings where many diseases have limited positive samples and lack dedicated risk models. Although trained on paired ECG and MRI data, CARDIAC-FM generates predictions using ECG alone or ECG combined with established risk scores, enabling broad clinical deployment without MRI. These findings demonstrate the promise of multimodal pre-training for generalizable cardiovascular risk prediction.

BACKGROUND: Atrial cardiopathy often precedes atrial fibrillation (AF) and has emerged as an independent risk factor for cardiovascular outcomes. However, previous studies have been limited in size and have overlooked the right atrium. METHODS: In 51 693 UK Biobank participants without prevalent AF, we assessed biatrial volumes and emptying fraction from cardiac magnetic resonance imaging using deep learning segmentation. We evaluated associations with new-onset AF, ischemic stroke, heart failure, and dementia, conducted a genome-wide association study, and evaluated causal associations using Mendelian randomization. RESULTS: Among 51 693 adults, the mean (SD) age was 65 (7.7) years, and 24 584 (48%) were male. During the 4-year follow-up, 964 (1.9%) developed AF, 266 (0.5%) developed ischemic stroke, 365 (0.7%) developed heart failure, and 72 (0.1%) developed dementia. After adjustment for clinical risk factors, both left and right atrial measures were associated with new-onset AF (left atrial minimal volume; hazard ratio, 1.55 [95% CI, 1.48–1.62]), ischemic stroke, and heart failure, with stronger associations in women. Left atrial minimal volume was also associated with dementia. Our genome-wide association study identified 51 (27 novel) genetic associations with atrial measures, many of which do not overlap with established AF loci. Genetic correlations revealed that each atrium had varying correlations with cardiometabolic risk factors, and Mendelian randomization demonstrated that left atrial measures had direct causal effects on AF and stroke risk. However, the stroke associations were attenuated after accounting for AF variants. CONCLUSIONS: In this largest assessment of biatrial structure and function to date, both left and right atrial cardiopathies were independently associated with increased risk of adverse cardiovascular events. We identified several novel genetic loci for atrial traits and observed unique genetic correlations between left and right atrial traits and cardiovascular phenotypes, providing insight into chamber-specific remodeling. Several of these measures are likely to be causal determinants of cardiovascular complications previously attributed to AF.

Generative modeling builds on and substantially advances the classical idea of simulating synthetic data from observed samples. This paper shows that this principle is not only natural but also theoretically well-founded for bootstrap inference: it yields statistically valid confidence intervals that apply simultaneously to both regular and irregular estimators, including settings in which Efron's bootstrap fails. In this sense, the generative modeling-based bootstrap can be viewed as a modern version of the smoothed bootstrap: it could mitigate the curse of dimensionality and remain effective in challenging regimes where estimators may lack root-$n$ consistency or a Gaussian limit.