A Performance-Driven Hybrid CNN-RF Model for Arrhythmia Detection: Insights Approach

Cardiovascular disease treatment requires arrhythmia detection from electrocardiogram (ECG) signals to be successful. This paper develops a deep learning system that combines Convolutional Neural Networks for feature extraction functions with Random Forest for classification purposes to enhance analysis of arrhythmia signals. The data from the MIT-BIH Arrhythmia Database showed the hybrid model produced 95.2% accuracy which outperformed the CNN-only model at 92.5% and the RF-only model at 87.3%. The model delivered precision and recall and F1-scores of 94.6% along with 94.1% and 94.3%, respectively. The AUC-ROC analysis confirmed that the model exhibited 97.1% average discriminating power for detecting the five arrhythmia classes. The performance evaluation for Normal heartbeats showed almost flawless accuracy (Prec: 99.1%, Rec: 99.2%) whereas Ventricular heartbeats demonstrated high detection rates (Prec: 95.2%, Rec: 96.1%). Between Fusion and Unknown classes there were only 3.2% misrepresented samples because of their similar morphological characteristics according to the confusion matrix and heatmap visualizations. Training-validation curves exhibited stable convergence following 100 epochs until both losses reached 0.09 for training and 0.11 for validation respectively. The t-SNE scatter map produced clusters which corresponded to different arrhythmia types because the model identified relevant clinical wave characteristics. The proposed CNN-RF hybrid model stands as a strong interpretable and easy-to-implement system for real-time ECG-based arrhythmia detection suitable for clinical cardiac monitoring.