An ECG-based machine-learning approach for mortality risk assessment in a large European population.

Abstract

Aims: Through a simple machine learning approach, we aimed to assess the risk of all-cause mortality after 5 years in a European population, based on electrocardiogram (ECG) parameters, age, and sex.

Methods: The study included patients between 40 and 90 years old who underwent ECG recording between January 2008 and October 2022 in the metropolitan area of Modena, Italy. Exclusion criteria established a patient cohort without severe ECG abnormalities, namely, tachyarrhythmias, bradyarrhythmias, Wolff-Parkinson-White syndrome, second- or third- degree AV block, bundle-branch blocks, more than three premature beats, poor signal quality, and presence of pacemakers and implantable cardioverter- defibrillators. Mortality was assessed using a set of logistic regression models, differentiated by age group, to which the Akaike Information Criterion was applied. Model fitting was evaluated using confusion matrix-related performance metrics, the area under the receiver operating characteristic (ROC) curve (AUC), and the predictive significance against the no-information rate (NIR).

Results: 53692 patients were enrolled, of whom 14353 (26.73 %) died within 5 years of ECG registration. The logistic regression model distinguished between those who died and those who survived based on the predicted mortality probability for all age groups, obtaining a significant difference between the predicted mortality and the NIR in 14 of the 55 age groups. Good accuracy and performance metrics were observed, resulting in an average AUC of 0.779.

Conclusions: The proposed model showed a good predictive performance in patients without severe ECG abnormalities. Therefore, this study highlights the potential of ECGs as prognostic rather than diagnostic tools.