A hybrid multiscale feature fusion model for enhanced cardiovascular arrhythmia detection

Cardiovascular arrhythmia, characterized by irregular heart rhythms, can lead to severe complications such as stroke and heart failure if not detected promptly. Traditional arrhythmia classification methods often struggle with class imbalance and fail to capture critical multiscale temporal and spatial features, leading to suboptimal diagnostic performance. To address these challenges, this study proposes a Hybrid Multiscale Feature Fusion (HMFF) strategy that combines ConvNeXt-X models with advanced data balancing techniques. Specifically, we explore two fusion approaches, Hybrid Feature Fusion (HFF) and HMFF, leveraging ConvNeXtTiny, ConvNeXtSmall, and ConvNeXtBase architectures. To mitigate class imbalance, we integrate SMOTE-Tomek Link (STL) and Random Oversampling (RO) strategies, ensuring robust representation of minority classes. Using the MIT-BIHA Dataset, our HMFF approach demonstrated substantial improvements over baseline models, which achieved an accuracy of 97.35% without feature fusion or data balancing. The HFF and HMFF models improved accuracies to 98.44% and 98.49%, respectively, even without balancing techniques. When combined with RO, accuracies increased to 99.11% and 99.25%, and further improved with STL to 99.15% and 99.30%. These results highlight the significant advancements provided by the HMFF strategy, offering an effective and scalable solution for enhancing arrhythmia detection and diagnosis in clinical settings. • Developed HFF and HMFF strategies with ConvNeXt-X models for arrhythmia classification. • Achieved 99.30% accuracy using HMFF with STL data balancing on the MIT-BIH database. • Demonstrated robust improvement by integrating multiscale fusion with data balancing.