Prediction of in-hospital atrial fibrillation after acute myocardial infarction: a cross-cohort study in Italian and Finnish patients

Background: The development of atrial fibrillation (AF) during hospitalization for acute myocardial infarction (AMI) is common and associated with worse prognosis. We aimed to train and, for the first time, cross-nationally validate a machine learning model to predict AF developing during the cardiac intensive care unit (CICU) stay in post-AMI patients admitted in sinus rhythm, using only routine, non-electrocardiographic clinical parameters available at CICU admission. Methods: We developed and tested classification pipelines on two independent cohorts of post-AMI patients from Italy (ITA, Sinus Rhythm = 2,200, AF = 240) and Finland (FIN, Sinus Rhythm = 3,631, AF = 452). Models were first evaluated via nested cross-validation on held out samples of the training cohort (internal testing) and then cross-cohort (external testing). Results: An L2-regularized logistic regression model performed optimally in each case, achieving comparable performances in internal (AUROC: 0.740 ITA, 0.755 FIN) and external (AUROC: 0.735 ITA-on-FIN, 0.723 FIN-on-ITA) testing. A combined-cohort model achieved an AUROC of 0.749. A posteriori feature relevance analysis confirmed the predictive value of established risk factors including advanced age, lower left ventricular ejection fraction and hypertension in both cohorts. Ongoing aspirin therapy and non-ST-elevation myocardial infarction emerged as additional protective factors exclusive to cohort ITA, whereas a higher high-sensitivity C-reactive protein (risk factor) and the current smoking status (protective factor) were specific to cohort FIN. Conclusions: Our study demonstrates that machine learning models using routinely collected clinical parameters could help predict in-hospital AF in post-AMI patients across different populations and national healthcare systems. Notably, a simple and fully interpretable logistic regression model achieved robust performance in the first cross-national validation for this prediction task, suggesting that more complex, black-box approaches may offer limited additional benefit for this specific task. Further improvements are needed to reach real-world application in clinical contexts for early risk stratification.