Resting-state based prediction of task-related activation in patients with disorders of consciousness

Assessment of the level of awareness of people with disorders of consciousness (DOC) is clinically challenging, motivating several studies to combine brain imaging with machine learning to improve this process. While this work has shown promise, it has limited clinical utility, as misdiagnosis of DOC patients is relatively high. As machine learning algorithms rely on accurately labelled data, any error in diagnosis will be learned by the algorithm, resulting in an equally limited diagnostic tool. The goal of the present study is to overcome this problem by stratifying patients, not by diagnosis, but by their capacity to perform volitional tasks during functional magnetic resonance imaging (fMRI)scanning. A total of 71 patients were assessed for inclusion. They were excluded for the final analysis if they had large focal brain damage, excessive head motion during scanning, or suboptimal MRI preprocessing. Patients underwent both resting-state andtask-based fMRI scanning. Univariate fMRI analysis was performed to determine if an individual patient had brain activity consistent with having retained volitional capacity (VC). Differences in resting brain network connectivity between patients with VC and patients without volitional capacity (non-VC) were measured. Connectivity data was then entered as input to a deeplearning framework. We used ad eep graph convolutional neural network (DGCNN) on connectivity data to identify a specific brain network that most significantly differentiates patients. We included 30 patients in our final analysis. Univariate analysis revealed that 13 patients displayed signs of VC, while 17 did not. We found that resting-state connectivity between frontoparietal control and salience network was significantly different between VC and non-VC patients. Furthermore, we found that using frontoparietal control network connectivity as input to the DGCNN resulted in the best classification performance. We found that the DGCNN performed best at discriminating between patients with VC when using only the frontoparietal control network as input to the model. The use of this deep learning method is a significant advance since its inherent flexibility permits the inclusion of both whole-brain and network-specific properties as input, allowing us to classify patients as either having or not having VC. This inclusion of multi-scale inputs (e.g. whole-brain and network-level) facilitates model interpretability and increases our understanding of the neurobiology of DOC. The results propose that the integrity of frontoparietal control network, a brain network well known to play a key role in executive functions and cognitive control, is essential for volitional capacity preservation in patients with DOC. The study also lays groundwork for development of a biomarker to aid in the diagnosis of DOC patients.