Braverman, Vladimir2025-01-172025-01-172024-122024-11-21December 2https://hdl.handle.net/1911/118219Arrhythmia detection is challenging due to the imbalance between normal and arrhythmia heartbeats, compounded by environmental noise in wearable devices compared to clinical settings. We propose a novel hierarchical model using CNN+BiLSTM with Attention for arrhythmia detection, featuring a binary classification for normal vs. arrhythmia beats and a multi-label classification for various arrhythmia types. We evaluated our model against several baselines on a proprietary dataset. Our model achieved 95% accuracy, 0.838 F1-score, and 0.906 AUC for binary classification, and 88% accuracy, 0.736 F1-score, and 0.875 AUC for multi-label classification, outperforming existing methods.application/pdfenMachine learningDeep learningElectrocardiogramWearable deviceMulti-label classificationThesis2025-01-17