Bidirectional LSTM With Feature Extraction For Sleep Arousal Detection In PSG Signal

Frequent arousals are correlated with a wide array of negative repercussions on health,such as neurocognitive impairment,mental health problems and cardiovascular disease.Polysomnography is the gold standard for the diagnosis of sleep disorders,which can monitor the changes of physiological signals of patients during sleep.In the real clinical situations,sleep specialists have to look through all the polysomnographic signals of the patients to identify the potential sleep disorders.The process of manual annotation is time-consuming.Therefore,it is necessary to establish an automatic annotation model to replace the manual annotation process.Based on the annotated database of 994 polysomnographic signals from the datasets of Physionet,we constructed a Bi-LSTM model employing the technique of the feature extraction to dectect the target arousal events.With the knowledge of clinical medicine and feature selection technology,this paper extracted 62 features containing biological information from the 13 signal channels of the polysomnography.These features included power spectral density and entropy extracted from EEG and EMG signals,the respiratory disturbance variables extracted from the airflow,chest and abdomen signals,the heart rate variability extracted from ECG signals,the time-domain characteristics of the EOG waveform,and hypoxic burden extracted from the blood oxygen channel,etc.We combined these features and normalized them before entering the classifier.The classifier was composed of three parts.The first layer was a Bi-LSTM network,Bi-LSTM network can connect the past and future signal information to the current input segment.The second layer used a fully connected network composed of two parallel 50 neurons.One network is used to output arousal judgment results,and the other is used to output sleep staging results.The final layer was softmax to output the prediction result.During the training process,we first used the smote algorithm to deal with the problem of the imbalanced datasets.Secondly,the cross-entropy auxiliary loss function was applied.We took the arousal detection as the main task and the sleep stage classification as the additional training task.The detection abilities of different channel combinations were discussed.The result showed that the combination of the breath,chest and ABD channels may has the strongest detection ability,followed by the combination of EEG and EMG channels.After five times of cross-validation,the AUPRC and AUROC of the integrated model were 0.481 and 0.878,respectively.The results show that the proposed method has a certain reference value in the judgment of patients’ arousal symptoms.The results show that the proposed method can help doctors read PSG images faster,and can avoid the subjective errors of the sleep specialists.