Predicting postoperative myocardial infarction in noncardiac surgery patients using explainable artificial intelligence

Background Postoperative myocardial infarction (PMI) is a major contributor to morbidity and mortality following noncardiac surgery. Current strategies emphasize preoperative risk stratification, early detection, and timely treatment, yet no single biomarker reliably predicts PMI. As a result, there is a need for novel predictive tools to improve clinical decision-making and patient outcomes. Purpose This study aimed to develop and validate an ensemble algorithm to predict PMI using preoperative electrocardiogram (ECG) data and explainable artificial intelligence. Methods We included patients aged 60 years and older who underwent noncardiac surgery at a large Dutch academic hospital between 2011 and 2012 with routine post-operative troponin I monitoring. Patients without a baseline ECG within 60 days before surgery were excluded. PMI was defined according to the Third Universal Definition of Myocardial Infarction. Patients were randomly split into a training and testing set. The ensemble model combined FactorECG, a previously developed variational auto-encoder that compresses median beat ECGs into 32 latent variables, with a logistic regression classifier trained on these latent variables to identify patients at risk of PMI. Results A total of 2,383 patients were included, of which 1,115 (46.8%) were female and 72 (3.0%) developed PMI. The model was trained on 1,906 ECGs and tested on 477 ECGs. It achieved an area under the receiver operating characteristic curve (AUROC) of 0.76 (95% confidence interval [CI]: 0.70–0.82), with a sensitivity of 0.68 (95% CI: 0.57–0.79) and specificity of 0.74 (95% CI: 0.72–0.76). The latent variables with the largest absolute coefficients correspond with inferolateral T-wave orientation (latent 5), left bundle branch delay (latent 26), and ventricular frequency (latent 10). Conclusions We developed an explainable ensemble model that is able to predict PMI with an AUROC of 0.76. With further prospective validation, this model could enhance preoperative risk stratification and guide clinical decision-making in perioperative care.Latent variable visualisation