Automatic model of sleep apnea detection using optimized weighted fusion process of hybrid convolution (1D/2D) efficient attention network from EEG signals

Sleep apnea (SA) is a sleep disorder characterized by breathing interruptions, and it causes significant health risks, including cardiovascular diseases, stroke, and secondary issues such as daytime accidents. The complex nature of SA necessitates accurate and timely diagnosis. The SA detection is often performed using polysomnography (PSG). However, PSG generates extensive data, making manual analysis labor-intensive and inefficient. In this work, a hybrid deep learning framework for automated SA detection combines advanced feature extraction and efficient classification techniques. Initially, the required EEG signals are collected from the standard data resource. The collected signals undergo for pre-processing stage, which is further decomposed into five signals delta, theta, alpha, beta, and gamma. Further, spectrogram images are generated from these decompositions, which are fused using optimal weights to produce image-based features F1. Simultaneously, the deep restricted Boltzmann machine (DRBM) features are extracted from the five decomposed signals. Then, the resultant features are fused with the optimized weight and it is designated as F2. Here the weight in DRBM is optimized using the improved Eurasian oystercatcher optimizer (IEOO). Finally, the resultant two feature formats are subjected to the developed hybrid convolution (1D/2D)–based efficient attention network (HCEAN) for SA detection. Extensive experimentation is conducted with the given dataset, by comparing the proposed technique with other standard approaches. The proposed model achieves an accuracy of 95.9%, a sensitivity of 95.86%, and a specificity of 95.93% that outperforms standard methods such as DNN, DiDTCN-REsLSTM, and CNN-LSTM. The obtained outcome confirms that the developed technique is a more effective method for detecting apnea events.