ENHANCING DIAGNOSTIC PRECISION IN EMERGENCY CARDIAC MONITORING: A LIGHTWEIGHT HYBRID CNN-LSTM APPROACH FOR ARRHYTHMIA DETECTION

Abstract

The prevalence of cardiovascular diseases necessitates the development of reliable automated diagnostic tools to assist clinicians in interpreting complex electrocardiogram (ECG) signals, where subtle irregularities often signal life-threatening conditions. While traditional diagnostic methods are limited by human fatigue and the complex non-linear nature of cardiac rhythms, this study introduces a robust hybrid deep learning framework designed to enhance arrhythmia detection precision. Utilizing the MIT-BIH Arrhythmia Database, we implemented a data-balancing strat- egy through targeted undersampling to overcome the inherent bias toward normal sinus rhythms, resulting in a refined training set of 24,729 samples. Our proposed architecture synergistically inte- grates 1D-Convolutional Neural Networks (CNN) for spatial feature extraction, Long Short-Term Memory (LSTM) units for capturing temporal rhythmic dependencies, and a Global Attention Mechanism to highlight clinically significant segments within the heartbeat. The experimental results demonstrate that the proposed hybrid model achieves a superior test accuracy of (97.15%), outperforming baseline architectures such as standalone CNNs (96.99%) and traditional machine learning models. An extensive ablation study further validates that the integration of attention- based weighting significantly improves the classification of minority classes, such as Fusion and Supraventricular beats. By offering high diagnostic sensitivity and generalizability, this framework provides a viable foundation for real-time cardiac monitoring and clinical decision support systems in emergency medicine.