An interpretable machine learning model for predicting 1-year all-cause mortality in patients with non-ischaemic acute myocardial injury: development and external validation
P Sun, X Qian, Z Zhang, L Liu, L Zhao, T Yang, X Xu, H Qin, Q Yang, X ZhouAbstract
Background
Non-ischaemic acute myocardial injury (NI-AMInj) carries high long-term mortality, yet no dedicated risk prediction tool exists. Existing scores developed for acute coronary syndromes inadequately capture the systemic pathophysiology of NI-AMInj.
Purpose
To develop and externally validate an interpretable machine learning model predicting 1-year all-cause mortality in NI-AMInj, and to identify key prognostic drivers.
Methods
NI-AMInj patients were identified from a regional multi-centre platform (development cohort, 2013–2022) and an independent critical care database (external validation, 2008–2022). NI-AMInj required troponin above the 99th percentile with >20% dynamic change, excluding acute myocardial infarction. The development cohort was split 8:2 for training and internal validation. Feature selection used a consensus strategy across LASSO, Boruta, and recursive feature elimination. Nine algorithms were compared. Performance was evaluated by area under the curve (AUC), calibration, and decision curve analysis. Shapley Additive exPlanations (SHAP) provided interpretability.
Results
78,666 patients were included (development: 67,773; external: 10,893; 1-year mortality: 19.3% vs 34.7%). Eleven predictors were selected: age, estimated glomerular filtration rate (eGFR), white blood cell count (WBC), pulmonary infection, heart failure, malignancy, intensive care unit admission, cardiology admission, surgical procedures, reperfusion therapy, and statin use. Gradient boosting machine (GBM) achieved optimal performance (internal AUC: 0.783; external AUC: 0.773), with good calibration and positive net clinical benefit. SHAP analysis identified eGFR, WBC, pulmonary infection, age, and heart failure as top prognostic drivers. Statin use and reperfusion showed protective associations. A web-based calculator was deployed for individualised assessment.
Conclusions
An interpretable GBM model using 11 routine clinical variables accurately predicts 1-year mortality in NI-AMInj with good cross-population generalisability. Renal function and inflammatory markers are dominant prognostic determinants. This tool may support early risk stratification in this high-risk population.SHAP