Early Detection of Cardiac Arrhythmia Using Bi-LSTM Networks on ECG Signals

Abstract

Cardiovascular diseases (CVDs) are a top global cause of death. Early detection of cardiac abnormalities using non-invasive techniques, such as electrocardiogram (ECG) signals, plays a significant role in improving patient outcomes and reducing mortality. Traditional methods rely on handcrafted feature extraction, which limits scalability and robustness. Therefore, this paper presents the implementation of a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network for binary arrhythmia classification using the MIT-BIH Arrhythmia Database. A structured preprocessing pipeline, which consists of bandpass filtering, Z-score normalization, and R-peak-centered segmentation, is applied to extract 112,572 labeled heartbeat segments. The implemented Bi-LSTM model is systematically evaluated against three classical machine learning baselines: Support Vector Machine (SVM), Logistic Regression, and Random Forest. Accuracy, precision, recall, F1-score, specificity, False Positive Rate (FPR), and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) are used to evaluate performance. The Bi-LSTM gets 96.41% overall accuracy, while under the same conditions, conventional models like Random Forest obtain 98.76% accuracy, showing that the benefits of deep learning depend on the quantity and balance of the data. The effect of class imbalance (90,603 normal vs. 21,969 abnormal beats) on model sensitivity is also shown in the paper. The results highlight the significance of domain-specific evaluation and balanced training procedures for ECG classification systems.