Whale Optimization Algorithm-Based Triple Attention Guided Hybrid Deep Learning Framework for ECG-Based Heart Failure Detection
Year: 2026
Type: article
Abstract: Heart failure (HF) is a life-threatening syndrome affecting over 64 million individuals globally and contributing disproportionately to cardiovascular morbidity and mortality. Electrocardiogram (ECG) signals offer a clinically viable, non-invasive modality for automated HF screening; however, existing deep learning frameworks suffer from inadequate preprocessing, suboptimal feature representation, limited attention granularity, and poor optimizer convergence in high-dimensional search spaces. This paper proposes a novel Whale Optimization Algorithm-based Triple Attention BiLSTM-CNN (WOA-TABiCNN) framework that addresses these limitations through five key contributions: (1) a dual-stage preprocessing pipeline integrating Discrete Wavelet Transform (DWT) with Empirical Mode Decomposition (EMD) and Z-score normalization for superior non-stationary noise suppression; (2) Hybrid Deep Statistical Invariant (HDSI) feature extraction combining seven complementary feature domains — Heart Rate Variability (HRV), Autoregressive (AR), Wavelet Packet Decomposition (WPD), VGG-16 deep features, Scale-Invariant Feature Transform (SIFT), hybrid statistical features, and frequency domain features — yielding a 1×1024 dimensional representation; (3) WOA-based joint feature selection and hyperparameter optimization converging to the global optimum faster than PSO, GA, GWO, COA, BOA, GOA, and BfSA; (4) a novel Triple Attention Mechanism (TAM) integrating Channel, Spatial, and Temporal Attention streams for selective amplification of diagnostically discriminative ECG patterns; and (5) a BiLSTM-CNN hybrid architecture with a hybrid ReLU-Tanh reconstruction module for simultaneous exploitation of temporal dependencies and spatial feature hierarchies. Experiments on the MIT-BIH Arrhythmia and St. Petersburg INCART 12-Lead databases at 80% training partitioning yield 97.61% accuracy, 98.94% sensitivity, 96.43% specificity, 97.08% precision, 98.01% F1-score, and 2.39% error rate, outperforming state-of-the-art methods including BfS-FSTReCN, HStA-CN, GOA-FSTReCN, and nine other baselines. The macro-average AUC of 0.9959 and lowest memory consumption (612.34 KB) and inference time (53.87 s) further confirm the framework's clinical viability
Authors: Amit M. Sahu, Dr.Jayant P. Mehare
Institution: G H Raisoni University
Language: English
Topic: ECG Monitoring and Analysis
Subfield: Cardiology and Cardiovascular Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W7159784381