Spatiotemporal Deep Learning for early detection and differentiation in idiopathic REM Sleep Behavior Disorder and Parkinson's Disease using functional MRI data

Isolated REM sleep behavior disorder (iRBD) has garnered significant attention as a potential early indicator of neurodegenerative diseases, particularly Parkinson's disease (PD). The growing acknowledgment of iRBD as a potential prodromal phase of PD has intensified efforts to identify brain biomarkers that may facilitate early diagnosis and intervention. There is an urgent need for these biomarkers to facilitate neuroprotection and improve patient outcomes. In this context, functional magnetic resonance imaging (fMRI) has emerged as a powerful tool for exploring the neural correlates of iRBD and PD. By providing insights into the dynamic brain activity patterns associated with these disorders, fMRI holds promise for enhancing our understanding of the underlying mechanisms and enabling timely therapeutic strategies. This study aimed to develop a spatiotemporal deep neural network (stDNN) using fMRI data to identify brain biomarkers linked to iRBD. The research focused on differentiating between iRBD patients, PD patients, and healthy controls (HC), utilizing data from three cohorts: IRCCS San Raffaele Scientific Institute (HSR), Parkinson's Progression Markers Initiative (PPMI), and the Movement Disorders Unit at the University of Campania “Luigi Vanvitelli” (MDU). The study also employed explainable AI (XAI) techniques to enhance interpretability, identifying the brain areas most involved in the disease process. 44 patients with iRBD, 86 patients with PD, and 129 HC were recruited from HSR, with 34 HC and 26 PD patients scanned using a different MRI scanner. The PPMI cohort included 100 patients with iRBD, 247 patients with PD, and 37 HC. Additionally, the MDU cohort included 90 patients with PD and 38 HC. The stDNN aimed to capture complex patterns in fMRI data that could distinguish iRBD from both PD and HC. To improve model performance, we incorporated transfer learning, adapting a model trained to differentiate HC from PD to the iRBD context. Cross-validation was used to ensure the robustness and generalizability of our findings. The stDNN model achieved balanced accuracy between 70% and 74% when distinguishing HC from PD, and up to 82% in advanced PD cases. This improvement was achieved by considering the UPDRS III scores and including only those subjects with values above the median, allowing for a more focused analysis of the more severe cases. In addition, the model achieved high performance, exceeding 80%, when differentiating between HC and individuals with iRBD. However performance declined in distinguishing iRBD from PD, likely due to overlapping functional characteristics. The network performed better when testing iRBD against advanced PD, achieving a balanced accuracy of 67%. Analysis of significant brain regions using XAI confirmed findings consistent with existing literature, enhancing the validity of our results. This study demonstrates the potential of stDNN to differentiate between iRBD, PD, and HC using fMRI data. While the ability to classify iRBD and HC highlights the relevance of identified biomarkers, challenges remain in distinguishing iRBD from advanced PD. Further research is necessary to elucidate the functional characteristics of these conditions, ultimately advancing our understanding of the neural mechanisms underlying iRBD and informing strategies for early detection and management of neurodegenerative disorders.