Obstructive sleep apnea detection using optimized Bi-LSTM with random forest based exhaustive feature selector

The diagnosis of sleep disorders like obstructive sleep apnea (OSA) is one of the most common types of sleep disorder, which requires the identification of the phases of sleep that occur throughout sleep. Manual assessment of sleep phases, on the other hand, is not only time consuming but also subjective and expensive. In addition, the traditional computer aided methodologies of OSA failed to obtain acceptable percentage of accuracy for enhanced diagnosis system. Therefore, this work focuses on development of OSA detection network (OSAD-Net) using optimized bi-directional long short-term memory (OBi-LTSM) with random forest based exhaustive feature selector (RF-EFS). Initially, multi-layer convolution neural network (MLCNN) model applied to extract the deep features from electrocardiogram (ECG) based OSA dataset. Then RF-EFS is applied to extract the optimal features using multi-level decisions. Finally, OBi-LSTM is trained with the optimal RF-EFS features, which performs the detection of OSA. The simulations are conducted on publicly available Apnea-ECG and university college of Dublin database (UCDDB), and it shows that the proposed OSAD-Net resulted in superior performance. The proposed OSAD-Net improved accuracy by 4.92%, precision by 5.15%, recall by 5.21%, F1-score by 4.73%, sensitivity by 6.55%, and specificity by 4.99%, as compared to existing methods for Apnea-ECG dataset. In addition, the proposed OSAD-Net has increased accuracy by 1.73%, precision by 1.29%, recall by 1.20%, F1-score by 0.92%, sensitivity by 4.14%, and specificity by 0.41%, as compared to existing methods for UCDDB dataset.