Metaheuristic optimization for atrial fibrillation detection: feature extraction, selection, and hyperparameter tuning

Abstract

Atrial fibrillation (AF) detection from electrocardiogram (ECG) signals is crucial for early diagnosis and intervention. This study presents a multi-objective optimization approach for AF detection, focusing on feature extraction, selection, and neural network hyperparameter tuning. The methodology uses cross-validation during the training of the two concatenated ECG dataset features and simultaneously minimizes the error rate on the separate validation folds of each dataset and reduces the number of selected features, enhancing model generalization and efficiency. Particle swarm optimization (PSO), grey wolf optimization (GWO), and differential evolution (DE) algorithms were implemented to navigate this multi-objective space. While all three algorithms were explored, the final solution, demonstrating a superior trade-off between accuracy and feature reduction, was obtained using DE. This approach effectively identifies optimal feature subsets and neural network configurations, yielding a robust and compact AF detection model. The proposed approach has shown promising results, with the model achieving accuracies of 96.38% and 90.69%, and corresponding area under the curve (AUC) values of 0.99 and 0.96, for the first and second datasets, respectively, using 10 optimally selected features.