0679 Enhancing Prediction of Sleepiness, Insomnia, and Cognitive Impairment Using Machine Learning on Polysomnography

Abstract Introduction Obstructive sleep apnea (OSA) is a prevalent disorder affecting approximately a billion people globally. Traditional diagnostic metrics from polysomnography (PSG) such as the apnea hypopnea index (AHI) are limited in predicting OSA-related impacts. This study aims to leverage machine learning (ML) methods to enhance PSG prediction of three outcomes: daytime sleepiness (defined as Epworth Sleepiness Scale >10), insomnia (Insomnia Severity Index >15) and cognitive impairment (Montreal Cognitive Assessment < 25) in patients with moderate to severe OSA. Methods This pilot study utilized a subset of Canadian Sleep and Circadian Network participants (patients with suspected OSA recruited from academic sleep centres). Advanced signal processing was used to derive 790 features from PSG European Data Format (edf) files including: respiratory events, heart rate variations, limb movements, and EEG based metrics from leads C3/C4 (e.g., sleep stages, spindle metrics, power frequencies, sleep depth, arousal intensity). 541 participants with moderate to severe OSA (AHI>15/hr) were included. Python 3.11.6 was used for analysis. For each outcome, dimensionality reduction was done using Lasso Regression, Principal Component Analysis (PCA), and Recursive Feature Elimination (RFE). Binary classification models with Logistic Regression (LR), Support Vector Classification (SVC), Random Forest (RF), Multi-Layer Perceptron (MLP), Gaussian Naïve bayes, and XGBoost were then used (18 models tested). Model accuracy was assessed using 5-fold cross validation. Results Models that employed RFE for feature selection and LR for classification (based on 50-70 features depending on outcome) yielded the best performance. The average accuracy/F1 values across all target outcomes were 0.71/0.71 for sleepiness, 0.68/0.61 for insomnia, and 0.62/0.62 for cognitive impairment. However, accuracy of all the ML models surpassed the predictive accuracy of AHI alone (accuracy of 0.55, 0.60, 0.53 using logistic regression). Conclusion This preliminary study supports the concept of applying advanced signal processing and ML techniques to PSG to help predict OSA-related outcomes. Future research with larger sample sizes, more diverse patients, more refined ML methodologies, and better feature engineering should further improve accuracy of these models. Support (if any) CIHR, BC Lung Association

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.