Poincaré plot analysis for sleep-wake classification of unseen patients using a single EEG channel

Abstract

This study explores automated sleep-wake classification using Poincaré plots derived from a single EEG channel. In order to quantify the Poincaré plots and utilize them for the distinction of sleep and wake states of the healthy individuals and patients with sleep disorders, various descriptors are computed. The most commonly used standard descriptors are SD1 and SD2, which determine the width and length of Poincaré plot. Along with SD1 and SD2, the ratio of SD1 to SD2, area of the Poincaré plots, energy of the slopes, and offsets obtained by linear fits to Poincaré plots with distinct lags, standard deviation, and complex correlation measure are also computed. Random undersampling with boosting technique (RUSBoost) is adopted to deal with the class imbalance problem. The performance of the method is evaluated on three different publicly available datasets by using 50%-holdout and 10-fold crossvalidation techniques. We achieved crossvalidation accuracies of 98.2%, 96.0%, and 94.4% for Sleep-EDF, DREAMS-Subjects and DREAMS-Patients datasets, respectively, by utilizing only eight features, and a single EEG channel. Furthermore, for the patient population with various sleep disorders such as mixed apnea, periodic leg movement syndrome, sleep apnea-hypopnea syndrome, and dyssomnia, we obtained average sensitivity of 96.8%, precision of 95.6%, and F1-score of 96.2%, for the sleep state; and 88.3%, 91.3%, and 89.8%, respectively for the wake state. Our results are comparable to or better than the existing studies in the literature. Further, the classification accuracies for the patients with a model trained only on the healthy population are quite impressive. Thus, the model is effective and generalizes well for the patient population.