Cross-modal deep learning to predict multiple sclerosis cervical spinal cord damage with digital monitoring of motor function

Motivation: People with multiple sclerosis (PwMS) might suffer from cervical spinal cord atrophy, which can only be assessed through regular MRI follow-up. Goal(s): Our goal was to explore using deep learning if the smartphone sensor signal can capture structural changes on MRI in PwMS' cervical spinal cord. Approach: MPRAGE-based atrophy measures and ten-second signal segments of 166 PwMS were derived. PwMS were split for five-fold cross-validation. A neural network was adapted from a ResNet pretrained on the UK-Biobank sensor dataset. Segment predictions were grouped for the prediction on PwMS. Results: For the PwMS prediction, the NN achieved sensitivity=0.74, specificity=0.81, and F1=0.80. Impact: The performance of the proposed cross-modal neural network indicates that smartphone walking tests correlate with structural changes in cervical spinal cord and might assist clinical assessments in multiple sclerosis.