2026-10-12

Background: Coronary heart disease (CHD) with concomitant tumors is increasingly prevalent and is associated with a higher mortality. We aimed to develop 2 predictive models to enable early identification of tumor risk in patients with CHD and mortality risk in those with CHD and cancer. Methods: We analyzed 134 141 patients with CHD treated at Tongji Hospital (Huazhong University of Science and Technology, Wuhan, Hubei, China) during 2001 to 2023. Data collected from the First Hospital of Peking University (Beijing, China) were used for external validation. We developed and evaluated 6 machine learning models (extreme gradient boosting, light gradient boosting machine, random forest, neural network, decision tree, and logistic regression). Model explainability was improved using Shapley additive explanation, and an online prediction platform was subsequently implemented. Results: For the model of predicting tumor risk in patients with CHD, extreme gradient boosting showed the best performance in the test set (area under the receiver operating characteristic curve =0.935), achieving the lowest Brier score (0.080). Its sensitivity, accuracy, and F1 score were 0.740, 0.882, and 0.751, respectively. Meanwhile, light gradient boosting machine performed well for in-hospital mortality prediction in the test set (area under the receiver operating characteristic curve =0.965), achieving the lowest Brier score (0.019), and the highest sensitivity (0.677), and F1 score (0.599). The 2 models also showed good performance in the external validation, suggesting a degree of generalizability. Conclusions: We constructed 2 predictive models to assess the risk of tumor and cancer-associated mortality in patients with CHD. The web platform assists clinicians in decision-making and patient management.

Background And Aims: Refeeding syndrome (RFS) is a common and serious metabolic reaction during nutritional support in critically ill patients, associated with substantially increased mortality risk and poor prognosis. Existing studies mainly focus on predicting its occurrence, while predictive models for 28-day mortality risk remain lacking. This study aimed to develop and validate an interpretable machine learning model to support short-term prognostic assessment in clinical practice. Methods: Data for critically ill patients with RFS were extracted from the Medical Information Mart for Intensive Care IV (MIMIC-IV, 2008-2019) as the internal cohort. A total of 8,705 patients were randomly assigned to training and test sets at a 7:3 ratio. Thirty-eight candidate variables were initially included, and feature selection was performed using least absolute shrinkage and selection operator (LASSO) logistic regression. Three machine learning models were subsequently constructed, and the optimal model was validated using two independent datasets. Model interpretability was enhanced using Shapley additive explanations (SHAP). Kaplan-Meier curves were used to compare survival between the high- and low-risk, and nomograms were constructed to facilitate clinical application. Results: Nine variables were ultimately selected to construct the logistic regression model. The model demonstrated excellent discriminative performance and generalizability, with area under the receiver operating characteristic curve (AUC) values of 0.821, 0.809, and 0.731 in the internal test set, external validation set I (MIMIC-III), and external validation set II (the Affiliated Hospital of Yangzhou University), respectively. SHAP analysis identified partial pressure of oxygen, mean blood glucose level, and age as the key prognostic factors and further revealed nonlinear effects among the variables. Kaplan-Meier survival curves confirmed that the model significantly distinguished between the high- and low-risk groups across all cohorts (all p < 0.05). Decision curve analysis showed that the model provided favorable clinical net benefit in both external validation cohorts. Conclusion: The interpretable machine learning model developed in this study effectively predicts the 28-day mortality risk of RFS in critically ill patients, provides an objective reference for individualized clinical decision-making, and may help improve patient prognosis.

Wide variation in outcomes is observed for patients with hormone-receptor-positive metastatic breast cancer (MBC) receiving first-line CDK4/6 inhibitors plus endocrine therapy (CDK4/6i+ET). Although front-line escalation strategies are beginning to enter clinical practice, assigning patients to intensified or de-intensified approaches remains challenging due to several mechanisms of resistance spanning tumor-intrinsic and host factors. To address this, we developed CDKPredict, an interpretable machine-learning model integrating baseline clinicopathologic and genomic features to predict outcomes from first-line CDK4/6i+ET. The model was trained on a densely annotated cohort of 1,264 patients and validated in a large multi-institutional external cohort (n=564, time-dependent AUC=0.74), stratifying patients into low-, intermediate-, and high-risk groups (median progression-free survival: 29.8, 14.7, and 8.0 months, respectively; hazard ratio [high vs low]=3.85, P=1.5x10-17). The model retained robust prognostication across clinically relevant subgroups and biomarker-selected populations and outperformed the current escalation criterion of treatment-free interval ≤12 months. Notably, CDKPredict demonstrated strong clinical and biological plausibility, with high-risk tumors featuring genomic and transcriptomic programs mechanistically linked to CDK4/6i+ET resistance, and higher-risk strata progressively enriched for actionable alterations. Finally, we developed CDKPredict-S, a pragmatic and scalable model requiring a restricted number of features intended for resource-variable settings, which was validated across two independent external cohorts totaling 1,019 patients and made publicly available. Collectively, CDKPredict provides a foundation for the prospective evaluation of risk-adapted, biomarker-informed therapeutic strategies.

Mitral regurgitation (MR) is among the most prevalent valvular heart diseases. We developed and externally validated a machine learning (ML)-based mortality prediction model for patients with significant MR. From the ChinaValvular Heart Disease (China-VHD) registry, 6323 patients with moderate or greater MR were randomly split 7:3 into training and internal validation sets. Variables selected concordantly by Cox stepwise regression, Least Absolute Shrinkage and Selection Operator (LASSO)-penalized Cox regression, and Boruta were retained. Seventeen ML and traditional models were compared. The final model was tested in the internal set and an external cohort of 4658 patients from the China Elderly Valve Disease (China-DVD) study. SHapley Additive exPlanations (SHAP) analysis assessed interpretability. Fifteen predictors were identified. CatBoost performed best, with AUCs of 0.783 (95% confidence interval [CI] 0.751-0.815) internally and 0.703 (0.678-0.728) externally, outperforming Cox (0.699), MitraScore (0.620), and EuroSCORE II (0.679). Age, NYHA class, and left ventricular ejection fraction were the top predictors. The model effectively distinguished high-risk from low-risk patients internally (adjusted hazard ratio [HR] 5.49, p < 0.001) and externally (adjusted HR 3.05, p < 0.001). In conclusion, we developed and externally validated an ML-based mortality prediction model for MR irrespective of etiology or treatment strategy, with potential to improve risk stratification and clinical decision-making.

Objectives: Polypharmacy rates continue to rise which can lead to increased risk of adverse drug events. Deprescribing is essential but time-consuming, requiring extensive clinical note and chart review. We developed RxReduce, a large language model (LLM)-powered framework as a proof of concept that can assist with medication deprescribing. Materials And Methods: RxReduce utilizes LLMs and retrieval-augmented generation (RAG) to evaluate patient data. Focusing on proton pump inhibitors, the system analyzes diagnoses and clinical notes to generate recommendations (Continue, Deprescribe, or Stop). To optimize accuracy, clinical notes were preprocessed into summaries before RAG retrieval. Results: The optimized system achieved an F1 score of 0.84, significantly outperforming the baseline (0.70). Conclusion: RxReduce demonstrates the viability of artificial intelligence-assisted deprescribing. By extracting key contextual insights, the framework reduces clinician burden and can be adapted for various medication classes to improve patient safety.