FORECAST-RBD: Forecasting Phenoconversion Risks and Its Clinical Phenotype in Patients with Isolated RBD Using Machine Learning and Explainable AI

Abstract

AbstractBackground and ObjectivesIsolated Rapid Eye Movement (REM) sleep behavior disorder (iRBD) is a sleep disorder associated with neurodegenerative diseases such as Parkinson’s disease and dementia with Lewy bodies. Predicting which iRBD patients will phenoconvert to neurodegenerative diseases is crucial for prognosis and management. The objective of this study is to develop a machine learning model using clinical markers to predict phenoconversion in patients with RBD.MethodsAnalyzing a cohort of 178 iRBD patients over a median follow-up period of 3.6 years, during which 30 patients converted to neurodegenerative conditions, we leveraged an comprehensive dataset encompassing demographics, medication history, cognitive assessments, sleep quality, autonomic symptoms, and parkinsonian signs. We explored a variety of feature selection methods and survival models. Additionally, separate models to predict the subtype of photoconversion—whether motor-first or cognition-first—were developed.ResultsThe extreme gradient boosting survival embeddings-Kaplan neighbors (XGBSE-KN) model demonstrated the best performance, achieving a concordance index of 0.823 and integrated Brier score of 0.123 on 10-fold cross-validation. Explainable AI methods provided insights into prediction rationales and key risk factors including age, RBDQ-KR factor 2 (behavioral factors), weight, antidepressant, coffee use, and UPDRS III excluding tremor score. For subtype classification, the RandomForestClassifier utilizing three features (PSQI-TST, MoCA, and age), emerged as the most effective, achieving a Matthews Correlation Coefficient of 0.697 in 100 repeated 5-fold cross-validations. These models have been deployed on a server for physician access.DiscussionThese models can aid prognosis and enable personalized management in RBD patients, potentially improving patient care and outcomes. While these findings are promising, further external validation of the models is necessary to confirm their efficacy and reliability in clinical settings. Future research should focus on incorporating additional biomarkers and exploring the models’ performance in larger, diverse cohorts.