An interpretable machine learning model using XGBoost with 108-dimensional clinical features effectively predicted acute myocardial infarction risk, achieving accuracy of 0.864 on test data and 0.932 on external validation, with SHAP analysis identifying Hs-cTnI as the primary predictor alongside nine additional clinical features.
The weighted XGBoost model achieved the best overall performance among all ML algorithms tested for AMI risk prediction.
The model was externally validated on an independent cohort collected from January 2025 to April 2025, demonstrating strong generalizability.
SHAP analysis identified Hs-cTnI (high-sensitivity cardiac troponin I) as the primary predictor of AMI risk.
The study enrolled 7,939 patients from a single hospital using a retrospective cohort design with 108-dimensional clinical features.
An interactive web server embedding the optimal model was developed and made publicly accessible to facilitate clinical use.
Traditional AMI risk assessment tools were identified as limited due to reliance on a restricted number of variables and static thresholds.
This research suggests that a machine learning model can predict the risk of acute myocardial infarction (heart attack) more effectively than traditional methods by analyzing a wide range of patient data. Using records from nearly 8,000 hospital patients, researchers trained several types of machine learning algorithms and found that an XGBoost-based model performed best, correctly classifying patients about 86% of the time on test data and 93% of the time when tested on an entirely new group of patients collected later. The model analyzed 108 different clinical measurements, including blood tests and patient characteristics, to make its predictions. To make the model understandable to clinicians, the researchers used a technique called SHAP analysis, which explains which factors most influenced each prediction. The ten most important predictors identified were: a highly sensitive heart enzyme test (Hs-cTnI), a heart failure marker (NT-proBNP), bad cholesterol (LDL-C), a kidney function measure (CG), a clotting marker (D-dimer), a liver enzyme (AST), platelet count, blood glucose, female sex, and BMI. This combination of factors reflects how heart attack risk involves interacting metabolic, clotting, and demographic factors in complex, non-linear ways that simple scoring systems may miss. The researchers also built a publicly available website (https://www.mips.net.cn) where clinicians can enter patient data and receive a risk prediction from the model. This research suggests that integrating a broad range of routine clinical data into an interpretable machine learning tool could help healthcare providers identify patients at risk for heart attacks earlier, potentially enabling faster and more targeted interventions. However, as this is a single-center retrospective study, further validation across diverse populations and healthcare settings would be important before widespread adoption.
Cui S, Gao L, Zhang N, Zhang H, Gong N. (2026). Development and validation of an interpretable machine learning model for early risk prediction of acute myocardial infarction.. International journal of medical informatics. https://doi.org/10.1016/j.ijmedinf.2026.106489