MOE-ECG: Multi-objective ensemble fusion for robust atrial fibrillation detection using electrocardiograms

Atrial fibrillation (AFib) is the most common sustained cardiac arrhythmia and imposes a major clinical and economic burden worldwide. Early and reliable detection is essential for reducing adverse outcomes, particularly as ECG data are increasingly generated by clinical and wearable monitoring devices. However, manual inspection of large ECG volumes is time-consuming and may delay timely decision-making. This paper proposes MOE-ECG, a multi-objective ensemble selection and fusion framework for robust AFib detection. The framework uses multi-objective particle swarm optimization to construct ensembles that jointly optimize predictive performance and inter-model diversity. RR intervals were derived from reference beat annotations, segmented into 20-, 60-, and 120-heartbeat windows with 50% overlap, and represented using statistical and nonlinear RRI features. From 15 extracted features, ten were retained using a two-stage relevance–redundancy correlation-based procedure. The framework used the MIT-BIH Atrial Fibrillation Database (AFDB) for training, the Saitama Heart Database Atrial Fibrillation (SHDB-AF) for validation, and the Long-Term Atrial Fibrillation Database (LTAFDB) for independent testing, with five runs used to assess stability. The 60-beat configuration provided the most balanced performance, achieving mean accuracy of 84.7%, precision of 69.7%, sensitivity of 98.2%, F 1 -score of 81.5%, and AUC of approximately 0.95 on LTAFDB. MOE-ECG also showed higher mean accuracy, precision, and F 1 -score than the single-best classifier and unoptimized average ensemble. These findings demonstrate competitive cross-dataset AFib detection with particularly high sensitivity on the independent test set, while retaining a compact and interpretable RRI-based representation across heterogeneous recording conditions.