Search

Cardiovascular disease remains the leading global cause of death, emphasizing the need for improved risk stratification beyond traditional tools such as Framingham, ASCVD, QRISK, and SCORE, which show limitations in diverse modern populations. Machine learning methods applied to electronic health records can enhance prediction by capturing complex, high-dimensional, and nonlinear relationships. This systematic review (2017–2022) evaluated machine learning models for cardiovascular risk prediction using EHR data, focusing on discrimination (AUROC, AUPRC), calibration, external validation, and reporting quality including TRIPOD adherence. A PRISMA-compliant search identified peer-reviewed studies applying machine learning to EHR-based cardiovascular risk prediction. Risk of bias was assessed using PROBAST, and narrative synthesis was conducted due to heterogeneity. Twenty-nine studies were included. XGBoost, random forest, and neural networks were the most common models and generally outperformed logistic regression and traditional risk scores in discrimination. However, calibration was infrequently reported, and external validation was limited, often showing reduced performance. Machine learning models demonstrate improved predictive discrimination over conventional risk scores, but limited calibration assessment and weak external validation constrain clinical applicability. Stronger validation frameworks are needed for clinical translation.

Hospital-acquired pressure injuries (HAPIs) are a common and largely preventable complication in ICU patients, affecting 5–15% of cases and contributing to increased morbidity and healthcare costs. Despite standardized nursing protocols, incidence remains high, highlighting the need for more effective predictive and preventive approaches. While traditional tools like the Braden Scale offer interpretability, they lack sufficient predictive accuracy in critically ill populations. In contrast, machine learning models such as XGBoost and random forests improve prediction but function as black boxes, limiting clinical trust and actionable insight. To address this gap, this work proposes an Explainable Boosting Machine (EBM) framework trained on electronic health record (EHR) data from over 50,000 ICU admissions (2017–2023). EBMs combine strong predictive performance with interpretability by modeling feature effects through shape functions and capturing pairwise interactions. This allows identification of both global and patient-specific risk factors while maintaining transparency. The framework emphasizes modifiable factors such as repositioning frequency, nutrition, and medical device management, revealing nonlinear thresholds and interaction effects often missed by conventional methods. Overall, the proposed approach integrates accurate prediction with clear, clinically interpretable insights, enabling real-time identification of actionable risk factors for HAPI prevention. By bridging predictive modeling and nursing decision-making, it supports more targeted interventions and improved patient outcomes in critical care settings.

Adverse drug reactions (ADRs) are a major global health issue, contributing to significant morbidity, mortality, and healthcare costs. Many ADRs are detected only after widespread drug use, reflecting limitations of pre-market trials in capturing real-world patient variability. Although electronic health records (EHRs) collected between 2017 and 2023 provide rich data for post-market surveillance, they remain underused for systematic ADR detection. Current pharmacovigilance methods rely heavily on spontaneous reporting systems, which suffer from underreporting and bias, while supervised machine learning approaches require labeled ADR data that are often unavailable for rare or novel events. This paper proposes a variational autoencoder (VAE)-based unsupervised framework to detect ADR signals from multimodal EHR data, including clinical notes and laboratory results. The model learns normal patient data distributions and identifies deviations as potential safety signals without requiring labeled ADR examples. A multimodal architecture combines natural language processing of clinical notes with structured laboratory encoders, forming a shared latent space for anomaly detection based on reconstruction error. The framework enables detection of unknown ADRs by flagging abnormal patterns in patient records across large datasets from 2017 to 2023. Its unsupervised nature makes it suitable for identifying rare or previously unrecognized drug safety issues. Overall, this approach offers a scalable, proactive pharmacovigilance strategy that shifts drug safety monitoring from reactive reporting to predictive detection using routine EHR data.

Hypertension affects about 1.4 billion adults globally and is a major modifiable risk factor for cardiovascular disease. Although several first-line antihypertensive drug classes exist, randomized controlled trials typically report only average treatment effects (ATEs), which mask important variability in individual patient responses. As a result, clinical guidelines often assume a homogeneous patient population, leading to trial-and-error prescribing, delayed blood pressure control, and avoidable adverse effects. I argue that causal forest models combined with double machine learning (DML) enable reliable estimation of heterogeneous treatment effects (HTEs) from observational electronic health record data. These methods can approximate randomized trial validity while capturing clinically meaningful variation in treatment response across patients. Compared with traditional approaches, they are computationally feasible and better suited for individualized treatment assessment. Therefore, comparative effectiveness research in hypertension should move beyond ATE-focused analyses toward routine HTE estimation using causal machine learning. This shift would support more precise, data-driven prescribing and improve patient outcomes.

Telemedicine expanded rapidly in the United States during the COVID-19 public health emergency as Medicare and state Medicaid programs relaxed coverage restrictions. Diabetes affects about 37 million Americans, and key outcomes such as HbA1c, blood pressure, and LDL cholesterol are routinely tracked in electronic health records. However, the causal impact of telemedicine expansion on these outcomes remains uncertain, as simple pre–post comparisons are confounded by concurrent trends such as the pandemic and seasonal variation. Randomized policy experiments are impractical, leaving a gap in high-quality causal evidence. We argue that Bayesian structural time series (BSTS) applied to state-level EHR aggregates provides a strong alternative. BSTS constructs a synthetic counterfactual from similar states, modeling trend and seasonality to estimate what outcomes would have been without telemedicine expansion. This allows clearer separation of policy effects from underlying time dynamics and produces interpretable estimates with uncertainty bounds. Unlike difference-in-differences, BSTS does not rely on parallel trends assumptions that may be violated in this context. It offers a transparent framework for causal inference using routinely available aggregated data. Policymakers should prioritize such causal methods when evaluating whether telemedicine expansions should become permanent rather than relying on descriptive before–after analyses.

Telemedicine expanded rapidly during COVID-19 as Medicare and states relaxed long-standing restrictions. Diabetes outcomes (HbA1c, blood pressure, LDL cholesterol) are routinely tracked in EHRs, yet causal evidence that telemedicine improves these outcomes remains limited. Pre-post analyses cannot separate telemedicine effects from confounding time trends such as seasonality and pandemic-related changes, while randomized state-level policy trials are infeasible. This leaves a key evidence gap for policy decisions. Bayesian structural time series (BSTS) using state-level EHR aggregates is the most suitable approach for causal inference, constructing counterfactual outcomes from similar donor states. BSTS accounts for trends, seasonality, and autocorrelation, and uses synthetic control principles to reduce confounding. It also provides uncertainty estimates and works with routinely collected aggregate data. Policymakers should require BSTS-based evidence before making telemedicine coverage permanent. Researchers should apply these methods to existing policy variation and share data and code. States should build routine EHR-based monitoring systems for diabetes outcomes. Causal evaluation of telemedicine is feasible now using existing data and methods. Relying on pre-post studies or waiting for randomized trials delays actionable evidence needed for policy decisions.

Suicidality and depression are major global health burdens, with over 700,000 suicide deaths annually and ~280 million people affected by major depressive disorder. Early risk prediction could support prevention, but traditional methods show limited accuracy. This PRISMA-compliant systematic review evaluated machine learning models for predicting suicidality and depression across electronic health records, social media, and wearable sensor data, focusing on performance, unimodal vs multimodal approaches, and ethical reporting. Searches of PubMed, PsycINFO, IEEE Xplore, arXiv, and ACM Digital Library identified eligible studies. EHR-based models showed AUROC 0.70–0.85 for suicide attempt prediction, social media models 0.70–0.80 for suicidal ideation, and wearable sensor models lower performance (0.65–0.75). Multimodal approaches improved performance by 5–10% over unimodal models. However, fewer than 20% of studies reported ethical considerations such as privacy, bias, or deployment safeguards. Overall, machine learning shows moderate-to-good predictive performance, with multimodal models performing best, but ethical reporting remains critically insufficient for clinical translation.

Clinicians often need rapid, evidence-based answers that integrate patient-specific electronic health records (EHRs) with clinical guidelines, but existing decision support tools are limited in real-time personalization. While large language models (LLMs) offer strong medical reasoning, they are prone to hallucinations and lack direct access to local EHR data, making them unsafe for standalone clinical use; meanwhile, traditional retrieval systems cannot synthesize coherent, context-aware responses. This paper proposes a retrieval-augmented generation (RAG) framework that combines dual-source retrieval from both institutional EHRs and clinical guideline databases. The system includes an EHR indexer, a guideline repository, a semantic retriever, an LLM-based generator, and a safety filter for hallucination mitigation. By grounding outputs in retrieved patient data and evidence-based recommendations, the model improves factual reliability, explainability, and clinical trustworthiness. Overall, the framework enables safe, real-time clinical question answering by integrating LLM reasoning with verified medical sources, with future validation planned on public EHR and guideline datasets.

Rare diseases collectively affect over 300 million people globally, yet individual conditions are often missed due to low clinician familiarity and non-specific presenting symptoms that mimic common disorders. Supervised machine learning requires large numbers of labeled examples for training, but rare diseases have too few diagnosed cases to develop condition-specific predictive models using traditional approaches. We propose a multimodal foundation model pretrained on 10 million de-identified electronic health records (EHRs) combining clinical notes and laboratory values for zero-shot rare disease diagnosis without requiring labeled training examples. The framework comprises four components: a clinical note encoder based on a large language model, a laboratory value encoder using a time-series transformer, a multimodal fusion module with cross-attention, and a zero-shot classifier that compares patient embeddings to disease descriptions. Pretraining on large-scale EHR data enables the model to learn general medical knowledge and disease patterns, allowing diagnosis of rare conditions by recognizing manifestations even when no labeled examples of that specific disease were used for training.

Long COVID (post-acute sequelae of SARS-CoV-2 infection, PASC) affects roughly 10–30% of COVID-19 survivors and is marked by persistent symptoms such as fatigue, cognitive dysfunction (“brain fog”), shortness of breath, loss of smell, and post-exertional malaise that can last for months or years, while its underlying biological mechanisms and validated diagnostic biomarkers remain unclear. The condition is highly heterogeneous, with patients showing different recovery patterns and no clearly defined clinical subtypes, and the scarcity of labeled datasets further limits the use of supervised machine learning methods for phenotyping. To address this, we propose a self-supervised contrastive multi-view learning framework that integrates three temporal data modalities—pre-infection electronic health records, acute-phase clinical and biomarker data (e.g., CRP, ferritin, D-dimer, lymphocyte counts), and post-acute symptom trajectories—using separate encoders and a shared latent space aligned through contrastive learning without requiring phenotype labels, followed by unsupervised clustering to identify potential subtypes. By exploiting the natural temporal linkage within each patient and contrasts across patients, this approach enables data-driven discovery of long COVID phenotypes, supports early prediction of subgroup membership, and may ultimately inform personalized treatment strategies, clinical trial design, and improved understanding of disease mechanisms.

Hypertension affects over 1.4 billion adults worldwide, and antihypertensive dose titration is a common but complex clinical decision. Although electronic health records contain longitudinal data on medication adjustments and blood pressure outcomes, determining optimal individualized dosing remains challenging due to confounding in observational data, where patients receiving higher doses often have worse baseline health. We propose a transformer-based model with causal attention masking to estimate counterfactual blood pressure outcomes under alternative dose regimens. The architecture ensures temporal validity by preventing information leakage from future events and encodes medication dose changes in a continuous representation. It includes a dose encoder, outcome predictor, and counterfactual contrastive loss to distinguish between competing treatment paths. This framework learns patient-specific dose–response relationships and enables personalized predictions for antihypertensive adjustments. While it supports individualized treatment planning from observational EHR data, prospective validation is still required before clinical deployment.

Sepsis continues to be a major contributor to morbidity and mortality among hospitalized patients globally, especially within intensive care and emergency departments, where rapid recognition is essential for improving survival through timely treatment. In recent years, machine learning approaches have gained attention for their ability to predict sepsis onset using routinely collected electronic health record data. This systematic review, conducted in accordance with PRISMA 2020 guidelines, synthesizes evidence from studies published between 2017 and 2025, focusing on model architectures, feature selection and engineering strategies, prediction time horizons, and validation methodologies. Searches across major biomedical and informatics databases identified 67 eligible studies. The included literature shows that logistic regression, ensemble tree-based algorithms, and deep learning models are most frequently applied for sepsis prediction tasks. However, the majority of studies rely on retrospective datasets with internal validation, while only a limited number incorporate prospective or real-world validation frameworks. Overall, although reported model performance is often strong in retrospective analyses, a consistent decline in accuracy is observed when models are evaluated in real clinical environments. These findings highlight that prospective validation and improved generalizability are still underdeveloped areas, underscoring the need for future research to emphasize real-time deployment and robust external validation before clinical integration.

Clinical trial recruitment is hindered by slow, costly, and labor-intensive processes, particularly due to the complexity of eligibility criteria often written in free text. This systematic review examines the use of large language models (LLMs) for matching clinical trial eligibility criteria to electronic health records (EHR). It evaluates zero-shot, few-shot, and fine-tuned LLM approaches, comparing their strengths, limitations, and deployment readiness in supporting patient-trial matching. Thirty-three studies published from 2017 to 2026 were included, with findings showing that zero-shot prompting is most adaptable for simple criteria, few-shot prompting offers consistent reasoning for ambiguous criteria, and fine-tuned models excel in task-specific performance but require labeled data and are less portable. The review concludes that no single approach is optimal for all trial screening tasks, and hybrid workflows combining various methods with human verification are most suitable for clinical use.