0715 Machine Learning Model to Predict Isolated REM Sleep Behavior Disorder Phenoconversion Time and Subtype using EEG

Abstract Introduction More than 80% of patients of isolated rapid eye movement (REM) sleep behavior disorder (iRBD), a prodromal disease of α-synucleinopathies, progress to neurological disease like Parkinson's disease (PD), dementia with Lewy bodies (DLB), and multiple system atrophy (MSA). Resting-state EEG measurements taken at baseline have been related to the phenoconversion. The timing of the conversion and the disease to which it will convert are crucial issues in iRBD. This work used baseline EEG in iRBD to create a prediction model for the phenoconversion time and subtype of α-synucleinopathy. Methods Resting-state EEG and neurological assessments were performed at baseline on patients with iRBD. EEG spectral power, Shannon entropy and weighted phase lag index were employed as features. Four models were used to predict subtypes for the PD-MSA and DLB groups, and three models were used to predict survival. External validation was also performed. Results 29 patients out of 143 who were followed up to nine years (mean 3.4 years) later developed α-synucleinopathies (14 PD, 9 DLB, 6 MSA). With a concordance index of 0.8130 and an integrated Brier score of 0.0921, the random survival forest was the best model for predicting survival. For the subtype prediction analysis, the model with the highest accuracy, extreme gradient boosting, had an accuracy of 86.52%. Both models indicated a high importance on EEG slowing related features. Conclusion It is possible to predict the timing and subtype of phenoconversion in iRBD using machine learning models of using EEG biomarkers. To confirm our findings, further study is required, including large sample data from various countries. Support (if any)