Machine Learning Predicts Phenoconversion from Polysomnography in Isolated REM Sleep Behavior Disorder

Abstract

Isolated rapid eye movement (REM) sleep behavior disorder (iRBD) is a prodromal stage of alpha-synucleinopathies. This study aimed at developing a fully-automated machine learning framework for the prediction of phenoconversion in patients with iRBD by using data recorded during polysomnography (PSG). A total of 66 patients with iRBD were included, of whom 18 converted to an overt alpha-synucleinopathy within 2.7 ± 1.0 years. For each patient, a baseline PSG was available. Sleep stages were scored automatically, and time and frequency domain features were derived from electromyography (EMG) and electroencephalography (EEG) signals in REM and non-REM sleep. Random survival forest was employed to predict the time to phenoconversion, using a four-fold cross-validation scheme and by testing several combinations of features. The best test performances were obtained when considering EEG features in REM sleep only (Harrel’s C-index: 0.723 ± 0.113; Uno’s C-index: 0.741 ± 0.11; integrated Brier score: 0.174 ± 0.06). Features describing EEG slowing had high importance for the machine learning model. This is the first study employing machine learning applied to PSG to predict phenoconversion in patients with iRBD. If confirmed in larger cohorts, these findings might contribute to improving the design of clinical trials for neuroprotective treatments.