Neurological Institute Outcomes

Excessive daytime sleepiness (EDS) is the propensity to fall asleep during the daytime, leading to poor quality of life, difficulty concentrating, and increased car and work accidents. The clinical assessment for EDS relies on the Epworth Sleepiness Scale (ESS) score, a subjective questionnaire. The patient responds to 8 scenarios, answering how likely they would be to fall asleep on a scale of 0-3 for a minimum score of 0 and a maximum score of 24. A higher score indicates more daytime sleepiness.

The objective was to identify surrogate markers for EDS using polysomnography-derived electroencephalography (EEG) data. A computational pipeline was created for automated, rapid, high-throughput, objective analysis of EED data of patients at the Cleveland Clinic. There were N=31 patients with high ESS (≥20) and N=41 patients with low ESS (<5). Polysomnography epochs were extracted from the STARLIT Registry (Sleep Signals, Testing, and Reports Linked to Patient Traits). The most recent epochs of wakefulness were analyzed. Several machine learning algorithms were considered, and the k-nearest neighbors algorithm was chosen.

Significant differences were identified in the high ESS group compared to the low ESS group: an increase in delta power band spectral density (0.294±0.010 vs. 0.239±0.005, p<0.001) and in theta power band spectral density (0.208±0.005 vs. 0.181±0.004, p<0.001) and a decrease in alpha power band spectral density (0.260±0.011 vs. 0.318±0.007, p<0.001) and beta power band spectral density (0.087±0.003 vs. 0.112±0.007, p<0.001). The coherence in 6 individual EEG channels across each frequency band was compared and found to be different for the high ESS and low ESS groups and between coherence in the C4 and O2 channels. The machine learning algorithms had an accuracy of 80.2%, precision of 79.2%, recall of 73.8%, and specificity of 85.3%. The confounding effects of age, sex, and body mass index were considered and the machine learning predictors alone explained EDS.

The results showed that the average power spectrum across EEG channels in low ESS patients is significantly enhanced in the alpha and beta bands and attenuated in the delta and theta bands compared to high ESS subjects. EEG data contains information that could be utilized for the objective assessment of EDS using machine learning.