Treatment Outcome Prediction Using Multi-Task Learning: Application to Botulinum Toxin in Gait Rehabilitation.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Neurological diseases often cause gait disorders due to spasticity.
• Botulinum toxin type A (BTX-A) is a common treatment for spasticity.
• Personalized treatment with optimal benefit-risk ratio is crucial.

Purpose of the Study:

• To develop and evaluate a framework for predicting knee and ankle kinematics post-BTX-A treatment.
• To optimize personalized treatment outcomes for patients with neurological gait disorders.
• To assess the impact of medical treatment data (MTD) and a gating mechanism on prediction accuracy.

Main Methods:

• A multi-task learning (MTL) regression strategy using LSTM models.
• Integration of medical treatment data (MTD) for context modeling.
• Implementation of a gating mechanism for efficient treatment interaction modeling.

Main Results:

• MTL models achieved the lowest RMSE for knee (5.60°) and ankle (3.77°) trajectories in TBI and CP patients, respectively.
• Overall best RMSE for MTL models ranged from 5.24° to 6.24°.
• MTL models significantly outperformed serial models, especially with MTD integration.

Conclusions:

• This is the first application of MTL for post-treatment gait trajectory prediction.
• The proposed MTL framework with a gating mechanism enhances prediction accuracy for BTX-A treatment outcomes.
• The approach facilitates more personalized and effective treatment strategies for neurological gait disorders.