Deep Learning for Multiple Sclerosis Differentiation Using Multi-Stride Dynamics in Gait.

Area of Science:

• Neurology
• Biomedical Engineering
• Computer Science

Background:

• Multiple sclerosis (MS) is a chronic central nervous system disorder impacting mobility.
• Gait abnormalities are early and frequent indicators of MS.
• Accurate differentiation of MS gait is crucial for diagnosis and management.

Purpose of the Study:

• To evaluate a deep learning (DL) framework, DeepMS2G, for classifying gait in persons with MS (PwMS) versus healthy controls (HC).
• To assess the framework's generalizability across different walking tasks and unseen participants.
• To explore the potential for automating MS diagnosis through gait analysis.

Main Methods:

• Collected gait data from 20 HC and 20 PwMS using an instrumented treadmill during single-task and dual-task walking.
• Utilized domain knowledge-based spatiotemporal and kinetic gait features with regression-based normalization.
• Compared 16 machine learning and DL algorithms, focusing on residual neural network (ResNet) models.

Main Results:

• ResNet-based models with regression-based normalization achieved the highest performance in classifying MS gait.
• A multi-scale ResNet attained perfect accuracy (1.0) when generalizing from single-task to dual-task walking.
• ResNet models achieved high accuracy (0.83) and F1-scores (0.81) when generalizing to unseen participants.

Conclusions:

• Advanced DL and gait dynamics analysis successfully classified MS gait across varied conditions and participants.
• The DeepMS2G framework demonstrates high accuracy in differentiating MS from healthy gait.
• These DL algorithms offer a potential pathway for automated MS diagnosis.