Jue Hou

BACKGROUND Renal cell carcinoma (RCC) recurrence after definitive therapy remains difficult to predict using conventional clinicopathologic risk models. Computed tomography (CT), electronic health record data, and radiology text may contain complementary prognostic signals, but many multimodal models are difficult to interpret. OBJECTIVE This study aimed to develop and evaluate an interpretable multimodal model that uses residual regression to identify imaging information not captured by known semantic features and to predict RCC recurrence after surgery. METHODS We conducted a retrospective cohort study using Mass General Brigham data from January 1, 2000, through September 1, 2022, with the public DeepLesion data set used for pretraining. The final RCC imaging cohort included 1615 patients. A CT contrastive language-image pretraining model aligned imaging and text embeddings; residual regression isolated an imaging component orthogonal to concept unique identifier–based text features. Baseline clinical, radiomic, natural language processing, and residual features were combined in XGBoost Cox models. Models were trained using algorithmically derived recurrence labels and evaluated using an expert-annotated subset. Discrimination was assessed at 12, 60, and 120 months, and predicted risk groups were compared using time-to-event analyses. RESULTS The ensemble model achieved area under the receiver operating characteristic curve values of 0.86 (95% CI 0.85-0.87) at 12 months, 0.78 (95% CI 0.77-0.80) at 60 months, and 0.82 (95% CI 0.81-0.83) at 120 months. The corresponding baseline-only values were 0.68 (95% CI 0.66-0.71), 0.73 (95% CI 0.72-0.75), and 0.79 (95% CI 0.78-0.81). At 60 months, ensemble-defined high- and low-risk groups had a hazard ratio of 6.34 (95% CI 2.70-14.91; P<.001). Residual features mapped to interpretable imaging phenotypes, including necrosis, infiltrative growth, and vascular characteristics. CONCLUSIONS Residual regression–derived imaging features added prognostic information to clinical, radiomic, and text-derived features while retaining an interpretable link to imaging phenotypes. External and prospective validation is needed before clinical use.

The multimodal data assembled through All of Us Research Program with enriched recruitment of populations underrepresented in biomedical research have provided a rich resource for equitable health research. The complexity of the All of Us research platform, however, is a common hurdle at the initiation of studies. This tutorial aims to demonstrate to researchers a practical, transparent and reproducible workflow for streamlining data curation and analysis. This tutorial covers modules: (1) design of All of Us studies; (2) selecting study cohort; (3) curating multimodal data (electronic health records, surveys, genomics, and wearable devices); (4) harmonizing disease onset outcomes in risk models. We showcased the proposed workflow using a case study for genetic risk prediction of rheumatoid arthritis for minority race/ethnicity groups. Step-by-step guides are provided in the supplementary materials and the online repository with links to SQL/R/Python example code and All of Us user interface operations.

Background: Patients with type 2 diabetes mellitus (T2D) have a higher risk of cardiovascular disease, including heart failure (HF), leading to health care burden including hospitalization and mortality. Among multiple T2D therapies, there are inadequate head-to-head comparisons of their effects on HF in the real-world patient population. Objective: This study aims to compare the time-to-HF among patients treated with different T2D drugs following metformin. Methods: We conducted a retrospective data analysis on electronic health records of 5000 patients with T2D. The inclusion criteria were previous treatment with metformin and initiation of glucagon-like peptide-1 receptor agonists (GLP1 RAs), dipeptidyl peptidase-4 inhibitors (DPP4i), sulfonylureas, or insulin. We grouped patients by the mechanism of their subsequent therapies and focused on 2 pairs of comparisons classified by insulin resistance: sulfonylureas versus insulin (increased resistance) and GLP1 RA versus DPP4i (decreased resistance). The outcomes were 5-year HF status and the HF-free survival time, which was verified manually by examining clinical notes. We applied doubly robust causal estimation and accounted for confounding by adjusting for coded and natural language processing electronic health record features identified through medical knowledge networks. Results: The study included 939 patients, of whom 204 (21.7%) received insulin, 482 (51.3%) received sulfonylureas, 90 (9.6%) received GLP1 RA, and 163 (17.4%) received DPP4i. Patients in the sulfonylureas group had a significantly higher 5-year HF-free survival compared to the insulin group (survival ratio of insulin/sulfonylureas 0.902, 95% CI 0.840‐0.976; P= .01). There was no significant difference between the DPP4i versus GLP1 RA group in 5-year HF-free survival (survival ratio of GLP1 RA/DPP4i was 0.953, 95% CI 0.849‐1.067; P= .40). For the occurrence of a HF-related hospitalization within 5 years, there were no significant differences between the sulfonylureas and insulin groups (risk difference 0.057, 95% CI –0.011 to 0.132; P= .11), and between the GLP1 RA and DPP4i groups (risk difference 0.010, 95% CI –0.096 to 0.129). Conclusions: We evaluated real-world evidence on 2 head-to-head comparisons of second-line T2D therapies on 5-year HF outcomes. Patients on sulfonylureas were associated with lower 5-year HF risks than those treated with insulin when measured by risk ratio, but no significant difference was detected when measured by the risk difference. Limitations of this study included potentially inadequate adjustment of confounding in the observational study and a limited sample size with validated HF outcomes.

Managing chronic diseases requires ongoing monitoring of disease activity and therapeutic responses to optimize treatment plans. With the growing availability of disease-modifying therapies, it is crucial to investigate comparative effectiveness and long-term outcomes beyond those available from randomized clinical trials. We introduce a comprehensive pipeline for generating reproducible and generalizable real-world evidence on disease outcomes by leveraging electronic health record data. The pipeline first generates scalable disease outcomes by linking electronic health record data with registry data containing a small sample of labeled outcomes. It then applies causal analysis using these scalable outcomes to evaluate therapies for chronic diseases. The implementation of the pipeline is illustrated in a case study based on multiple sclerosis. Our approach addresses challenges in real-world evidence generation for disease activity of chronic conditions, specifically the lack of direct observations on key outcomes and biases arising from imperfect or incomplete data. We present advanced machine learning techniques such as semisupervised and ensemble methods to impute missing outcome data, further incorporating steps for calibrated causal analyses and bias correction.

Federated learning of causal estimands may greatly improve estimation efficiency by leveraging data from multiple study sites, but robustness to heterogeneity and model misspecifications is vital for ensuring validity. We develop a Federated Adaptive Causal Estimation (FACE) framework to incorporate heterogeneous data from multiple sites to provide treatment effect estimation and inference for a flexibly specified target population of interest. FACE accounts for site-level heterogeneity in the distribution of covariates through density ratio weighting. To safely incorporate source sites and avoid negative transfer, we introduce an adaptive weighting procedure via a penalized regression, which achieves both consistency and optimal efficiency. Our strategy is communication-efficient and privacy-preserving, allowing participating sites to share summary statistics only once with other sites. We conduct both theoretical and numerical evaluations of FACE and apply it to conduct a comparative effectiveness study of BNT162b2 (Pfizer) and mRNA-1273 (Moderna) vaccines on COVID-19 outcomes in U.S. veterans using electronic health records from five VA regional sites. We show that compared to traditional methods, FACE meaningfully increases the precision of treatment effect estimates, with reductions in standard errors ranging from 26% to 67%.

UNSTRUCTURED Managing chronic diseases requires ongoing monitoring of disease activity and therapeutic responses to optimize treatment plans. With the growing availability of disease-modifying therapies, it is crucial to investigate comparative effectiveness and long-term outcomes beyond those available from randomized clinical trials. We introduce a comprehensive pipeline for generating reproducible and generalizable real-world evidence on disease outcomes by leveraging electronic health record data. The pipeline first generates scalable disease outcomes by linking electronic health record data with registry data containing a small sample of labeled outcomes. It then applies causal analysis using these scalable outcomes to evaluate therapies for chronic diseases. The implementation of the pipeline is illustrated in a case study based on multiple sclerosis. Our approach addresses challenges in real-world evidence generation for disease activity of chronic conditions, specifically the lack of direct observations on key outcomes and biases arising from imperfect or incomplete data. We present advanced machine learning techniques such as semisupervised and ensemble methods to impute missing outcome data, further incorporating steps for calibrated causal analyses and bias correction.

Background: Type 2 diabetes mellitus (T2DM) is a common health issue, with heart failure (HF) being a common and lethal long-term complication. Although insulin is widely used for the treatment of T2DM, evidence regarding the efficacy of insulin compared to noninsulin therapies on incident HF risk is missing among randomized controlled trials. Real-world evidence on insulin’s effect on long-term HF risk may supplement existing guidelines on the management of T2DM. Objective: This study aimed to compare insulin therapy against other medications on HF risk among patients with T2DM using real-world data extracted from insurance claims. Methods: A retrospective, observational study was conducted based on insurance claims data from a single health care network. The study period was from January 1, 2016, to August 11, 2021. The cohort was defined as patients having a T2DM diagnosis code. The inclusion criteria were patients who had at least 1 record of a glycated hemoglobin laboratory test result; full insurance for at least 1 year (either commercial or Medicare Part D); and received glucose-lowering therapy belonging to 1 of the following groups: insulin, glucagon-like peptide 1 receptor agonists (GLP-1 RAs), dipeptidyl peptidase-4 inhibitors (DPP-4Is), or sodium-glucose cotransporter-2 inhibitors (SGLT2Is). The main outcome was the 5-year incident HF rate. Baseline covariates, including demographic characteristics, comorbidities, and laboratory test results, were adjusted to correct for confounding. Results: After adjusting for a broad list of confounders, patients receiving insulin were found to be associated with an 11.8% (95% CI 11.0%‐12.7%), 12.0% (95% CI 11.5%‐12.4%), and 15.1% (95% CI 14.3%‐16.0%) higher 5-year HF rate compared to those using GLP-1 RAs, DPP-4Is, and SGLT2Is, respectively. Subgroup analysis showed that insulin’s effect of a higher HF rate was significant in the subgroup with high HF risk but not significant in the subgroup with low HF risk. Conclusions: This study generated real-world evidence on the association of insulin therapy with a higher 5-year HF rate compared to GLP-1 RAs, DPP-4Is, and SGLT2Is based on insurance claims data. These findings also demonstrated the value of real-world data for comparative effectiveness studies to complement established guidelines. On the other hand, the study shares the common limitations of observational studies. Even though high-dimensional confounders are adjusted, remaining confounding may exist and induce bias in the analysis.

This study leverages the rich diversity of the All of Us Research Program (All of Us)’s dataset to devise a predictive model for cardiovascular disease (CVD) in breast cancer (BC) survivors. Central to this endeavor is the creation of a robust data integration pipeline that synthesizes electronic health records (EHRs), patient surveys, and genomic data, while upholding fairness across demographic variables. We have developed a universal data wrangling pipeline to process and merge heterogeneous data sources of the All of Us dataset, address missingness and variance in data, and align disparate data modalities into a coherent framework for analysis. Utilizing a composite feature set including EHR, lifestyle, and social determinants of health (SDoH) data, we then employed Adaptive Lasso and Random Forest regression models to predict 6 CVD outcomes. The models were evaluated using the c-index and time-dependent Area Under the Receiver Operating Characteristic Curve over a 10-year period. The Adaptive Lasso model showed consistent performance across most CVD outcomes, while the Random Forest model excelled particularly in predicting outcomes like transient ischemic attack when incorporating the full multi-model feature set. Feature importance analysis revealed age and previous coronary events as dominant predictors across CVD outcomes, with SDoH clustering labels highlighting the nuanced impact of social factors. The development of both Cox-based predictive model and Random Forest Regression model represents the extensive application of the All of Us, in integrating EHR and patient surveys to enhance precision medicine. And the inclusion of SDoH clustering labels revealed the significant impact of sociobehavioral factors on patient outcomes, emphasizing the importance of comprehensive health determinants in predictive models. Despite these advancements, limitations include the exclusion of genetic data, broad categorization of CVD conditions, and the need for fairness analyses to ensure equitable model performance across diverse populations. Future work should refine clinical and social variable measurements, incorporate advanced imputation techniques, and explore additional predictive algorithms to enhance model precision and fairness. This study demonstrates the liability of the All of Us’s diverse dataset in developing a multi-modality predictive model for CVD in BC survivors risk stratification in oncological survivorship. The data integration pipeline and subsequent predictive models establish a methodological foundation for future research into personalized healthcare.