Akademik Veri Yönetim Sistemi
Journal of Medical and Biological Engineering, cilt.45, 2025 (SCI-Expanded)
Purpose: The study aimed to develop an interpretable machine learning system for detecting specific arrhythmias—namely ventricular ectopic beats, supraventricular ectopic beats, and atrial fibrillation—by integrating SHapley Additive exPlanations (SHAP) to enhance model transparency and clinical applicability. Methods: A retrospective analysis was performed using ECG data from the MIT-BIH Arrhythmia Database and a 3-channel Holter ECG dataset from the Dokuz Eylül University Heart Rhythm Management Centre (DEUHRMC). Two models were implemented: a LightGBM classifier utilizing 12 features derived from heart rate variability and ECG morphology over 30-minute intervals, and a 1D Convolutional Neural Network (CNN) processing raw ECG segments centered on R peaks. Data were split into training (80%) and testing (20%) sets with 10-fold cross-validation, and class imbalance was addressed using oversampling techniques. SHAP analysis was applied to elucidate feature contributions in the LightGBM model. Results: The LightGBM model achieved an AUROC of 0.993 (95% CI: 0.990–0.996), precision between 90.10% and 99.90%, and an overall accuracy of 98.39%, outperforming the 1D CNN, which achieved an AUROC of 0.967 (95% CI: 0.960–0.974) and an accuracy of 97.06%. SHAP analysis identified heart rate variability and QRS morphology as key features, aligning with clinical expectations. Conclusion: Integrating SHAP with the LightGBM model substantially enhances the interpretability and potential clinical utility of arrhythmia detection systems. Future studies should validate these findings across diverse populations and further refine model transparency for routine clinical use.