<![CDATA[Electrocardiogram (ECG)-based arrhythmia detection remains a critical challenge due to nonlinear heartbeat variability, temporal instability, and dynamic cardiac regime transitions. Conventional Long Short-Term Memory (LSTM) networks are capable of learning sequential ECG patterns; however, they often struggle to simultaneously represent short-term and long-term cardiac temporal dynamics. This study proposes a Regime-Aware LSTM framework integrated with Adaptive Dual-Timescale Temporal Encoding (RADTTE) for cardiac dynamics modeling and arrhythmia detection. The proposed approach introduces exponentially weighted temporal representations operating at dual timescales to capture adaptive cardiac regime transitions from RR interval sequences. The framework extracts short-term and long-term temporal cardiac dynamics prior to sequential modeling using an LSTM architecture. Experiments were conducted using the MIT-BIH Arrhythmia Database consisting of annotated ECG recordings. The proposed framework was compared against a conventional LSTM baseline using accuracy, precision, recall, F1-score, specificity, and ROC-AUC metrics. Experimental results demonstrate that the proposed method achieved superior performance with improved anomaly sensitivity and reduced false positive rates. The proposed model achieved 98.1% accuracy and a 97.0% F1-score, outperforming the conventional LSTM baseline. The findings indicate that regime-aware temporal representation significantly enhances sequential cardiac anomaly modeling and improves arrhythmia transition detection.]]>
