Diego A. Salazar

Objectives: Social determinants of health (SDOH) are incompletely captured in structured electronic health records (EHRs) but are frequently documented in unstructured clinical notes. We evaluated large language models (LLMs) for extracting SDOH from clinical text. Materials and Methods: We constructed an adult sepsis cohort from the Medical Information Mart for Intensive Care-IV (Sequential Organ Failure Assessment ≥2) and analyzed clinical notes from 1 year prior to 30 days following suspected infection. Three instruction‑tuned, decoder‑only LLMs (Mistral‑Instruct‑7B-v0.2, DeepSeek‑R1‑Distill‑Qwen‑14B, and GPT‑oss‑20B) were evaluated using structured prompts with predefined label schemas and few‑shot examples. Performance was benchmarked against a clinically validated annotated dataset and compared with a fine‑tuned encoder‑decoder baseline. Macro‑F1 scores were reported. A gold‑standard Intensive Care Unit (ICU) sepsis subset was independently annotated by 3 reviewers to assess domain‑level performance and ensemble strategies. Results: Decoder‑only models outperformed the fine‑tuned encoder‑decoder baseline across SDOH domains. GPT‑oss achieved the highest macro‑F1 score (0.79) compared with Flan‑T5‑XXL (0.57). Prompt refinement substantially improved extraction accuracy. Ensemble majority voting increased robustness across domains, while unanimous agreement yielded high precision but limited coverage. In a subsequent mortality analysis, extracted SDOH did not independently predict 30-day mortality, which was instead associated with established clinical and demographic risk factors. Discussion: Instruction‑tuned decoder‑only LLMs can reliably extract multiclass SDOH from unstructured clinical notes without task‑specific fine‑tuning. Ensemble and agreement‑based strategies provide practical operating points for high‑precision clinical deployment. Conclusion: These findings support the feasibility of leveraging LLMs to enrich EHRs with structured SDOH data, providing a scalable approach for incorporating social context into downstream risk stratification and health outcome prediction.

This tutorial discusses a methodology for causal inference using longitudinal modified treatment policies. This method facilitates the mathematical formalization, identification, and estimation of many novel parameters and mathematically generalizes many commonly used parameters, such as the average treatment effect. Longitudinal modified treatment policies apply to a wide variety of exposures, including binary, multivariate, and continuous, and can accommodate time-varying treatments and confounders, competing risks, loss to follow-up, as well as survival, binary, or continuous outcomes. Longitudinal modified treatment policies can be seen as an extension of static and dynamic interventions to involve the natural value of treatment and, like dynamic interventions, can be used to define alternative estimands with a positivity assumption that is more likely to be satisfied than estimands corresponding to static interventions. This tutorial aims to illustrate several practical uses of the longitudinal modified treatment policy methodology, including describing different estimation strategies and their corresponding advantages and disadvantages. We provide numerous examples of types of research questions that can be answered using longitudinal modified treatment policies. We go into more depth with one of these examples, specifically, estimating the effect of delaying intubation on critically ill COVID-19 patients’ mortality. We demonstrate the use of the open-source R package lmtp to estimate the effects, and we provide code on https://github.com/kathoffman/lmtp-tutorial.

Multi-omic data integration analyzes a vast amount of biological data and contributes to understanding the biological processes underlying organisms. Multiple machine learning techniques have been proposed to solve this task, including extensions of the joint Non-negative Matrix Factorization (jNMF) method, such as the Multi-project and Multi-profile jNMF (M&M-jNMF). This method jointly factorizes input matrices from two projects into low-rank matrices which have clustering properties. However, the M&M-jNMF method does not capture the non-linear patterns of the data. This paper proposes an extension of the M&M-jNMF approach using projections into high-dimensional spaces through kernel functions; therefore, we propose the M&M-KjNMF method. We compared the standard M&M-jNMF and M&M-KjNMF methods using three different omic profiles of the lung adenocarcinoma data. As M&M-jNMF, we used data from experimental and observational data source. We evaluated the performance of both methods by comparing the cophenetic coefficient, AUC, and biological score. We found that M&M-KjNMF outperforms M&M-jNMF. The new proposed method enables the identification of molecule co-modules enriched in pathways tightly related to lung cancer emergence and progression.

The integration of multi-omic data using machine learning methods has been focused on solving relevant tasks such as predicting sensitivity to a drug or subtyping patients. Recent integration methods, such as joint Non-negative Matrix Factorization, have allowed researchers to exploit the information in the data to unravel the biological processes of multi-omic datasets. We present a novel method called Multi-project and Multi-profile joint Non-negative Matrix Factorization capable of integrating data from different sources, such as experimental and observational multi-omic data. The method can generate co-clusters between observations, predict profiles and relate latent variables. We applied the method to integrate low-grade glioma omic profiles from The Cancer Genome Atlas (TCGA) and Cancer Cell Line Encyclopedia projects. The method allowed us to find gene clusters mainly enriched in cancer-associated terms. We identified groups of patients and cell lines similar to each other by comparing biological processes. We predicted the drug profile for patients, and we identified genetic signatures for resistant and sensitive tumors to a specific drug. Source code repository is publicly available at https:/bitbucket.org/dsalazarb/mmjnmf/-Zenodo DOI: 10.5281/zenodo.5150920. Supplementary data are available at Bioinformatics online.