Interpretable deep learning for the remote characterisation of ambulation in multiple sclerosis using smartphones

Area of Science:

• Digital Health
• Machine Learning in Medicine
• Neurology

Background:

• Smartphones offer potential for remote, continuous, and objective measurement of multiple sclerosis (MS) disability.
• Deep Convolutional Neural Networks (DCNNs) can extract richer ambulatory features from smartphone sensor data than traditional methods.
• Challenges in remote health data include small sample sizes, data sparsity, and heterogeneity.

Purpose of the Study:

• To develop and evaluate a transfer learning (TL) framework using DCNNs for remote MS disability assessment.
• To improve the accuracy of detecting MS-related ambulatory characteristics from smartphone sensor data.
• To enhance the interpretability of deep learning models in clinical applications.

Main Methods:

• A TL model was developed by leveraging knowledge from large-scale human activity recognition (HAR) datasets.
• DCNN models were fine-tuned for MS disease recognition using smartphone inertial sensor data.
• Layer-Wise Relevance Propagation (LRP) was employed to visualize DCNN decision-making processes.

Main Results:

• The TL DCNN HAR models outperformed SVM and end-to-end DCNN models by 8-15% in MS recognition tasks.
• LRP analysis identified distinct ambulatory patterns differentiating people with MS (PwMS) from healthy individuals.
• Key distinguishing features included cadence, gait speed, and signal perturbations during ambulation.

Conclusions:

• Transfer learning with DCNNs provides a robust and accurate method for remote MS disability assessment using smartphone data.
• Interpretable deep learning models enhance clinical acceptance and understanding of digital health assessments.
• These findings can improve disease management and therapeutic interventions for PwMS through objective, out-of-clinic data.