A Machine Learning Approach for Detecting Idiopathic REM Sleep Behavior Disorder

Background and purpose: Growing evidence suggests that Machine Learning (ML) models can assist the diagnosis of neurological disorders. However, little is known about the potential application of ML in diagnosing idiopathic REM sleep behavior disorder (iRBD), a parasomnia characterized by a high risk of phenoconversion to synucleinopathies. This study aimed to develop a model using ML algorithms to identify iRBD patients and test its accuracy. Methods: Data were acquired from 32 participants (20 iRBD patients and 12 controls). All subjects underwent a video-polysomnography. In all subjects, we measured the components of heart rate variability (HRV) during 24 h recordings and calculated night-to-day ratios (cardiac autonomic indices). Discriminating performances of single HRV features were assessed. ML models based on Logistic Regression (LR), Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) were trained on HRV data. The utility of HRV features and ML models for detecting iRBD was evaluated by area under the ROC curve (AUC), sensitivity, specificity and accuracy corresponding to optimal models. Results: Cardiac autonomic indices had low performances (accuracy 63–69%) in distinguishing iRBD from control subjects. By contrast, the RF model performed the best, with excellent accuracy (94%), sensitivity (95%) and specificity (92%), while XGBoost showed accuracy (91%), specificity (83%) and sensitivity (95%). The mean triangular index during wake (TIw) was the best discriminating feature between iRBD and HC, with 81% accuracy, reaching 84% accuracy when combined with VLF power during sleep using an LR model. Conclusions: Our findings demonstrated that ML algorithms can accurately identify iRBD patients. Our model could be used in clinical practice to facilitate the early detection of this form of RBD