Spatiotemporal deep learning for early detection of isolated REM sleep behavior disorder and Parkinson's disease using functional MRI data.
Authors
Affiliations (15)
Affiliations (15)
- Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milan, Italy.
- Neurotech Hub, Vita-Salute San Raffaele University, Milan, Italy.
- Neurology Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy.
- Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics and Maternal Child Health, University of Genoa, Genoa, Italy.
- Biolab, PoliTo(BIO)Med Lab, Department of Electronics and Telecommunications, Politecnico di Torino, Torino, Italy.
- Vita-Salute San Raffaele University, Milan, Italy.
- Sleep Disorders Center, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milan, Italy.
- Neurorehabilitation Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy.
- Department of Advanced Medical and Surgical Sciences, University of Campania "Luigi Vanvitelli", Napoli, Italy.
- Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milan, Italy. [email protected].
- Neurotech Hub, Vita-Salute San Raffaele University, Milan, Italy. [email protected].
- Neurology Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy. [email protected].
- Vita-Salute San Raffaele University, Milan, Italy. [email protected].
- Neurorehabilitation Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy. [email protected].
- Neurophysiology Service, IRCCS San Raffaele Scientific Institute, Milan, Italy. [email protected].
Abstract
This study aimed to develop and evaluate a spatiotemporal deep-neural-network (stDNN) using resting-state fMRI (rs-fMRI) data to identify brain biomarkers associated with isolated REM sleep behavior disorder (iRBD) and Parkinson's disease (PD) and to differentiate these conditions from controls. The final sample included 771 subjects, comprising 423 patients with PD, 144 with iRBD, and 204 healthy controls. stDNN model was applied to mean timeseries extracted for each subject from rs-fMRI data. By integrating spatio-temporal features, the network classified subjects based on distinct neural patterns. Model generalizability was assessed using subject-wise k-fold cross-validation. Explainable artificial intelligence (XAI) methods were applied. stDNN achieved balanced accuracy rates of 71.0% in distinguishing controls from PD and up to 71.9% in middle-stage PD cases. It also demonstrated over 80% accuracy in differentiating healthy controls from iRBD. XAI analysis highlighted the involvement of fronto-parietal and temporal regions, including the dorsolateral prefrontal cortex, and anterior temporal gyri, in distinguishing controls from PD. In the comparison with iRBD, key contributing areas included the bilateral superior and medial frontal gyri, dorsolateral prefrontal cortex, parietal and occipital regions (lingual gyri and cuneus). This study demonstrates the potential of stDNN to differentiate between iRBD, PD, and controls using rs-fMRI data.