An attention-enhanced CNN-BiLSTM framework with confidence-gated inference for arrhythmia detection using ECG signals

Abstract Automated arrhythmia detection from electrocardiogram (ECG) signals is essential for early diagnosis and continuous cardiac monitoring. However, accurate classification remains challenging due to complex temporal dependencies in ECG waveforms and severe class imbalance, where clinically significant arrhythmias are underrepresented. To address these challenges, we propose an Attention-Enhanced CNN–BiLSTM framework with a Confidence-Gated Inference mechanism for ECG-based arrhythmia detection. The proposed model integrates a one-dimensional convolutional neural network (1D-CNN) to capture local morphological features with a Bidirectional Long Short-Term Memory (BiLSTM) network to model long-range temporal dependencies in ECG signals. A temporal attention mechanism is employed to focus on diagnostically relevant segments of the signal. Additionally, a confidence-gated inference strategy is introduced to improve prediction reliability, particularly for minority classes. An extensive ablation study across 18 configurations is conducted to evaluate the contribution of temporal modeling, attention, and inference strategies. On the MIT-BIH Arrhythmia Database, the proposed model achieves an accuracy of 96.58% and a macro-F1 score of 86.05%. To ensure an unbiased evaluation free of test-set leakage, we further adopt a strict protocol in which all hyperparameters ( $$\gamma$$ , $$\tau$$ ) are selected exclusively on a held-out validation set; under this leakage-free protocol the model attains an accuracy of 96.74% and a macro-F1 score of 86.54% on the untouched test set. These findings demonstrate that integrating attention-driven temporal modeling with calibrated inference yields efficient arrhythmia classification on the MIT-BIH benchmark.