0582 Polysomnography Parameters Predict Mean Sleep Latency Test Results Using Machine Learning

Abstract

Abstract Introduction The mean sleep latency test (MSLT) objectively measures mean sleep latency (MSL) and sleep onset during REM periods (SOREMPs) in the absence of external alerting factors used for diagnoses of central hypersomnolence disorders. Polysomnography (PSG) is routinely performed the night before an MSLT. We explored the use of PSG variables at the individual level to predict MSL and SOREMPs assisted by machine learning methods. Methods 802 PSG/MSLTs from 796 patients (75% female, age 34 years[SD:12.1], BMI 27.9[SD:7.6], ESS 13.9[SD:4.9]) performed at Cleveland Clinic Sleep Disorder Center from 2012 to 2022 were included. We used 5-fold cross-validation experimental settings where five types of machine learning models (logistic regression, support vector machine, multilayer perceptron, random forest, XGBoost) were trained for two separated binary classification tasks: predict MSL< 8 minutes (343 out of 802) and SOREMP>=2 (234 out of 802). Within each fold, we performed a feature selection analysis on the training set in 39 PSG and demographic variables based on Analysis of Variance (ANOVA) with p < 0.05. A random search approach with 1000 combinations was used for machine learning hyperparameter tuning. The area under the ROC (AUC-ROC) score was used to rank the methods based on the test set of each fold (mean+SD). We retrained the best machine learning configuration in a new single train/test (80%/20%) split to evaluate the top 5 features that were more relevant to the best model by computing their Shapley Additive Explanations values (SHAP). Results XGBoost had the best average AUC-ROC for predicting MSL of 0.71 + 0.04. Its most important features out of 14 selected were total recording time, percentage of sleep time, total wake time after sleep onset, non-supine sleep time, and sleep efficiency. Random forest was the best for SOREMP prediction with AUC-ROC of 0.75 + 0.05, and the most important features were REM latency, age, and Epworth Sleepiness Scale score, percentage of time on REM, minimum oxygen saturation, out of 13 selected. Conclusion PSG the night prior to MSLT provides predictive information that can be leveraged to anticipate MSLT results on the individual level. This work provides the foundation to better identify patients with central hypersomnolence disorders. Support (if any)