Advanced Classifiers and Feature Reduction for Accurate Insomnia Detection Using Multimodal Dataset

Abstract

Sleep deprivation is a significant contributor to various diseases, leading to poor cognitive function, decreased performance, and heart disorders. Insomnia, the most prevalent sleep disorder, requires more effective diagnosis and screening for proper treatment. Actigraphic data and its combination with physiological sensors like electroencephalogram (EEG), electrocardiogram (ECG), and body temperature have proven significant in predicting insomnia using machine learning methods. Studies focusing solely on actigraphic data achieved an accuracy of 84%, combining it with other wearable devices increased accuracy to 88%, and 2-channel EEG alone yielded an accuracy of 92%, but limits scalability and practicality in real-world settings. Here we show that using the hybrid approach of incorporating both recursive feature elimination (RFE) and principal component analysis (PCA) on sleep and heart data features yields outstanding results, with the multi-layer perception (MLP) achieving an accuracy of 95.83% and an F1 score of 0.93. The top-ranked features are predominantly sleep-related and time-domain RR interval. The dependent variables in our study have been extracted from the self-report Pittsburgh Sleep Quality Index questionnaire responses. Our findings emphasize the importance of tailoring feature sets and employing appropriate reduction techniques for optimal predictive modeling in sleep-related studies. Our results demonstrate that the ensemble classifiers generalize well on the dataset regardless of the feature count, while other algorithms are hindered by the curse of dimensionality.

Keywords

• Actigraphy
• classification
• feature reduction
• heart rate variability
• insomnia