Key result
LightGBM model predicts post-revascularization MI or repeat revascularization with strong accuracy at ~85%.
Why the study?
Patients after coronary revascularization face elevated risk of repeat revascularization or MI, yet risk communication and education are often difficult for them to understand and act on. AI-based tools were proposed to improve engagement by delivering personalized and actionable explanations of individual risk factors.
Does an EHR-based machine learning model integrated with an LLM-powered chatbot accurately predict repeat revascularization or myocardial infarction and provide individualized risk education in patients with coronary artery disease after coronary revascularization?
Observational (n=26,746)
Does an EHR-based machine learning model integrated with an LLM-powered chatbot accurately predict repeat revascularization or myocardial infarction and provide individualized risk education in patients with coronary artery disease after coronary revascularization?
Answers are grounded in this paper and cite their sources.
An EHR-trained LightGBM model (AUC 0.862) combined with SHAP and an LLM-powered chatbot is a feasible and scalable approach to quantify post-revascularization risk and deliver personalized education to CAD patients.
Background Coronary artery disease (CAD) remains a major health concern, and patients after coronary revascularization face elevated risk of repeat revascularization or myocardial infarction (MI)1. However, risk communication and education are often difficult for patients to understand and act on2. AI-based tools may improve engagement by delivering personalized, clear, and actionable explanations of individual risk factors3. Purpose To develop and validate an AI-powered chatbot that provides individualized risk communication and patient education for CAD patients after revascularization, leveraging an Electronic Health Records (EHR)-based prediction model to explain personal risk and support informed care decisions. Methods We analyzed EHR records for 26,746 patients who underwent coronary revascularization between 2021 and 2025; 6,727 had repeat revascularization or diagnosed MI during follow-up. Median age was 69.2 years (IQR: 62.3–76.2), and 77.3% were male. Predictors included demographics, key clinical history (e.g., prior MI / angina / heart failure / stroke / hypertension / diabetes / chronic kidney disease), and laboratory markers (e.g., glucose, HDL, LDL) summarized in pre- and post-operative time windows (e.g., 0–14, 14–30 days). We trained multiple machine-learning models (logistic regression, random forest, XGBoost, SVM, LightGBM). SHapley Additive exPlanations (SHAP) provided patient-level interpretability8, and a large language model (LLM)-powered chatbot generated patient-friendly explanations by integrating model outputs with EHR context in a web interface. Results The system produced personalized risk explanations that highlighted patient-specific drivers and offered actionable lifestyle modification suggestions. The LightGBM model showed strong discrimination (AUC 0.862, accuracy 84.7%, F1 56.73%) and SHAP waterfall plots improved transparency by visualizing contributions from features such as age, comorbidities, and post-operative biomarkers. For instance, chronic ischemic heart disease and prior MI substantially increased predicted risk, while post-operative changes in white blood cell count and hemoglobin showed time-window–dependent effects. The LLM-powered chatbot translated these insights into understandable narratives, supporting patient comprehension and engagement. The web-based interface enabled intuitive interaction for both patients and clinicians. Conclusions An AI-powered chatbot for post-revascularization CAD risk communication is feasible and scalable. An EHR-trained predictive model can quantify risk, while SHAP improves interpretability and trust. The LLM-powered chatbot provides personalized education that may empower patients to make informed decisions. Future work will prioritize clinical validation, user experience testing, and clinical trials assessing long-term outcomes to establish effectiveness in routine care.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) conducted an observational in Coronary artery disease (CAD) post-revascularization (n=26,746). EHR-based machine learning prediction model (LightGBM) with SHAP interpretability and LLM-powered chatbot vs. Logistic regression, random forest, XGBoost, and SVM models was evaluated on Repeat revascularization or diagnosed myocardial infarction (MI) (AUC 0.862). The LightGBM model predicted post-revascularization repeat revascularization or MI with strong discrimination (AUC 0.862, accuracy 84.7%, F1 56.73%) to power an educational LLM chatbot.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: