Automatic detection of non-apneic sleep arousal regions from polysomnographic recordings

A signal processing/machine learning (ML), data-driven approach for classifying targeted sleep arousal regions of polysomnography (PSG) signals is presented focusing on feature subset selection and consensus methods, deploying ensemble techniques. The targeted regions are the regions where RERA and Non-RERA-Non-Apnea events are present. The sensor independent and sensor-based features in time and frequency domain were derived from the PSG signals. To reduce the feature space dimension, a combination of feature selection strategies and a method of rank aggregation was applied to rank the features. Aiming to find a feature set, which conveys the most discriminative information of detection in designated learning models, the Non-Dominated Sorting Genetic Algorithm was used as the optimization algorithm. In order to capture the relation between feature vectors across time, a composition of feature vectors was formed. To tackle the unbalanced data problem, several techniques were used and a data fusion strategy stood out. Also, considering a more robust classifier, a metaclassifier was generated using different features, datasets, and classifiers. Finally, the predictions of models generated by bagging techniques and boosting methods were compared. The presented method was developed, validated and tested on the PhysioNet Challenge 2018 training dataset consisting of 994 subjects. The highest performance on 192 test subjects based on the area under precision-recall curve (AUPRC) and the area under receiver operating characteristic (AUROC) curve were 0.465 and 0.927, respectively. This study suggests that automatic detection of RERA and Non-RERA-Non-Apnea sleep arousal regions from biosignals is possible and can be a suitable substitution for PSG.

Automated summary

This study presents a signal processing/machine learning approach for classifying targeted sleep arousal regions in PSG signals, focusing on feature subset selection and consensus methods. By utilizing optimization algorithms and feature combination, the most discriminative feature set was identified to achieve automatic detection of sleep arousal regions.