Brain tissue damage is closely linked to disability in multiple sclerosis (MS). The localization of white matter (WM) lesions influences the course of the disease.
However, the interrelation between lesions topography and cortical atrophy distribution for predicting the clinical disability remains unclear. Use a deep learning neural network framework with the purpose to identify critical co-varying patterns for individualized disease prediction.
Clinical disability was measured using the Expanded Disability Status Score at baseline and at a one-year follow-up in a cohort of 119 patients with early relapsing-remitting MS and in a replication cohort of 81 patients. Co-varying patterns of cortical atrophy and baseline lesion distribution were extracted by parallel ICA and used as features for constructing a deep learning convolutional neural network. The prediction was conducted for each identified lesion pattern separately using 50% as training cohort and 50% as testing cohort.
In the study cohort, we identified three distinct distribution types of WM lesions (“cerebellar”, “bihemispheric” and “left-lateralized”). The “cerebellar” and “left-lateralized” patterns were reproducibly detected in the second cohort. Each of the patterns predicted to different extents, short-term disability progression, while the “cerebellar” pattern predicting individual disability progression with an 10-fold cross-validation accuracy of above 90% for the Study cohort (95% CI: 88%-94%) and above 85% for the replication cohort (95% CI: 81%-88%) respectively.
These findings highlight that role of distinct spatial distribution of cortical atrophy and WM lesions predicting disability. The cerebellar involvement is shown as a key feature in the CNN framework for prediction of rapid clinical deterioration.