0257 Topological Data Analysis of Electroencephalogram Signals for Pediatric Obstructive Sleep Apnea

Abstract

Abstract Introduction Obstructive sleep apnea (OSA) is difficult to diagnose and leads to significant complications, including poor rest and heart disease. Identification of OSA in patients typically requires a polysomnogram, which includes acquisition of electroencephalogram (EEG) signals in order to assist with identification of sleep stages. Recent advancements in signal processing technology, including topological data analysis (TDA), provide a novel new way to analyze EEG signals to assist with identification of OSA before disordered breathing occurs. Methods We leverage EEG signals from polysomnogram studies included in the Nationwide Children’s Hospital Sleep DataBank (NCHSDB) dataset for this study. This dataset includes 3,984 pediatric patients, of which we leverage 2,881 containing usable data. We analyze each sleep state (awake before sleeping, awake in between sleep, awake after study, N1, N2, N3, and REM sleep) individually. We first break the data into 30-second chunks, then compute a distance matrix based on the multivariate time series obtained as part of the polysomnogram for the delta, theta, alpha, beta, and gamma bands individually. We then apply TDA to the distance matrices, leading to a set of persistence landscapes for the 0th homology group. We then group persistence landscapes for OSA positive patients and OSA negative patients, and perform a permutation test for each frequency band, leading to a p-value indicating significance. Results Based on our permutation test, we find significant (p < 0.05) results in all bands for the N1 (p=0.0 for all bands), N2 (p=0.0 for all bands), N3 (p=0.0 for all bands), and REM sleep (p=0.023 for delta band, p=0.0 for all other bands) cases. In the cases when the patient is awake, we find significant differences in the delta (p=0.022), theta (p=0.007), and beta (p=0.002) bands in the awake before study case, the delta (p=0.003), theta (p=0.003), and beta (p=0.0) bands for the inter-sleep awake case, and no significant differences in the awake after study case. Conclusion Our results indicate that a statistically significant difference between OSA positive and OSA negative patients can be identified when patients are awake through topological data analysis. Support (if any)