Introduction
Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide. Early and accurate detection of heart arrhythmias is critical for effective intervention, yet manual analysis of long-term Electrocardiogram (ECG) recordings is labor-intensive and prone to human error. While Computer-Aided Diagnosis (CAD) systems have been developed, they often struggle with a trade-off between computational efficiency and diagnostic accuracy. This research introduces a hybrid "multi-tier" deep learning model designed to detect five distinct types of arrhythmias from single-lead ECG data with high precision and low latency.
Methodology
The proposed system utilizes a four-tier architecture that integrates deep learning for feature extraction with evolutionary algorithms for optimization. This approach aims to bridge the gap between simple manual feature engineering and computationally heavy deep learning models.
1. Local Feature Learning (ResNet)
The first tier employs a Residual Convolutional Neural Network (ResNet) to extract morphological features from segmented ECG signals. By focusing on the specific shapes and spikes (P-QRS-T complexes) within short time windows, the model captures the essential physical characteristics of different heartbeats.
2. Global Feature Learning (LSTM)
To account for the temporal nature of cardiac rhythms, the second tier uses Long Short-Term Memory (LSTM) layers. These layers summarize how local morphological features evolve over time, effectively learning the "rhythm" of the heart.
3. Feature Optimization (Genetic Algorithm)
A key innovation of this approach is the use of a Genetic Algorithm (GA) to prune the high-dimensional feature vectors generated by the deep learning layers. By "evolving" a subset of the most relevant features, the system reduces noise and computational overhead without sacrificing predictive power. The researchers found that this pruning step significantly boosted accuracy compared to using raw deep learning outputs alone.
4. Classification
The optimized feature set is finally passed to a classifier. While several methods were evaluated, including Support Vector Machines (SVM) and Multi-layer Perceptrons (MLP), the k-Nearest Neighbor (k-NN) algorithm provided the best performance when paired with GA-optimized features.
Experimental Setup and Results
The model was validated using the MIT-BIH Arrhythmia Database, a benchmark containing 48 half-hour ECG recordings. Signals were classified into five categories following AAMI standards: Normal (N), Supraventricular ectopic (S), Ventricular ectopic (V), Fusion (F), and Unknown (Q).
Performance Metrics
The system demonstrated exceptional performance in intra-patient testing, achieving an accuracy of 98.0% and a sensitivity of 99.7%. In the more challenging inter-patient scenario—where the model is tested on individuals not seen during training—it maintained an accuracy of 87.2%, surpassing many existing complex models.
Furthermore, the model is highly efficient, requiring only 10ms to process a single recording. This low computational cost makes it suitable for real-time monitoring on wearable e-health devices, where battery life and processing speed are critical.
Discussion and Conclusion
The results highlight that "more data" is not always superior; the strategic pruning of features via Genetic Algorithms significantly enhanced model performance. By operating on a single-lead ECG, the system offers a practical solution for remote and low-resource medical environments.
The primary limitation remains the performance drop during inter-patient classification, suggesting natural variability between individuals is still a challenge. Future work will focus on improving inter-patient generalization and exploring the integration of Bidirectional LSTM (BiLSTM) components to further refine diagnostic accuracy while maintaining real-time feasibility.
Summary updated