ECG-based NSR, CHF, and ARR classification: Tradeoff between classical feature engineering and TCN-based deep features

Accurate and efficient automatic classification of electrocardiogram (ECG) signals is essential for scalable cardiac monitoring and timely clinical decision making. Herein, we focus on classifying Arrhythmia (ARR), Congestive Heart Failure (CHF), and Normal Sinus Rhythm (NSR). In particular, this study compares a classical feature-based pipeline that includes time-domain, frequency-domain, and time-scale features with a modern deep feature extractor employing the Temporal Convolutional Network (TCN). For this, MATLAB simulations were conducted on ECG signals from the MIT-BIH database to evaluate the classification performance and quantify practical trade-offs by reporting wall-clock training times, inference latency, and hardware requirements. The results showed that the Support Vector Machine (SVM) achieved classification accuracies of 88%, 91%, and 100% for ARR vs NSR, ARR vs CHF, and NSR vs CHF, respectively, when trained and tested on the classical features. On the other hand, the TCN achieved an overall accuracy of 99.4% to distinguish between the three classes. Furthermore, the TCN yielded superior raw predictive performance but offered reduced interpretability compared with the classical feature-based pipeline, at the expense of requiring substantially greater computational resources. These results demonstrated that well-engineered classical descriptors remain competitive for low-resource deployment, while data-driven TCNs deliver higher accuracy when computational budgets permit.