Enhanced Arrhythmia Detection Employing CNN-LSTM Hybrid Architectures and Advanced Signal Processing Techniques

Cardiac arrhythmias are abnormal heart rhythms that can result in severe health complications, including stroke, heart failure, and sudden cardiac arrest. Early detection of arrhythmias is essential to reduce the risk of these life-threatening events. Electrocardiogram (ECG) signals are a widely accepted tool for diagnosing heart conditions by capturing the electrical activity of the heart, but accurate classification remains challenging due to signal noise and inter-patient variability. This research proposes a novel framework for the classification of ECG signals into multiple arrhythmia categories using an integrated signal processing and deep learning pipeline. In the preprocessing stage, noise is removed using Discrete Wavelet Transformation (DWT), followed by accurate R-peak detection to localize the QRS complex. Feature extraction is performed using a combination of Principal Component Analysis (PCA) and Wavelet Packet Decomposition (WPD), which reduces dimensionality while preserving discriminative time–frequency characteristics. These features are then classified using a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model that captures both spatial morphology and temporal dependencies in ECG beats. The proposed method is evaluated on the MIT-BIH Arrhythmia Database, achieving an overall classification accuracy of 98.32%, with precision 98.5%, recall 98.55%, F1-score 98.52%, and specificity 99%. Cross-database validation using a patient-separated split (DS1/DS2) and the European ST-T Database further confirms the generalization ability of the framework, with accuracies of 97.8% (DS2) and 96.5% (cross-database) and AUC-ROC values above 0.90 across classes. These results demonstrate that the proposed CNN–LSTM framework offers an efficient and accurate solution for real-time arrhythmia detection.