GaitForeMer

GaitForeMer predicts gait impairment severity in Parkinson's disease (PD) patients by estimating Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) gait impairment scores from learned motion features derived via human motion forecasting.

Key Features:

• Transformer architecture: Uses a transformer-based model (Gait Forecasting and Impairment Estimation Transformer) for motion forecasting and impairment estimation.
• Self-supervised pre-training: Employs a self-supervised pre-training strategy inspired by GPT-3 to learn motion representations.
• Pre-training on public datasets: Pre-trains on extensive public human motion datasets to forecast gait movements and capture universal motion features.
• Transfer learning to clinical data: Applies learned motion features to clinical PD gait data to predict MDS-UPDRS gait impairment severity.
• Performance metrics: Reports evaluation metrics with F1 score 0.76, precision 0.79, and recall 0.75.
• Non-intrusive monitoring: Enables non-intrusive assessment of motor impairments using motion-forecasting-derived representations.

Scientific Applications:

• MDS-UPDRS gait severity prediction: Predicts MDS-UPDRS gait impairment severity in Parkinson's disease patients from motion data.
• Longitudinal motor monitoring: Supports non-intrusive longitudinal monitoring of motor impairments in PD.
• Transfer of public motion data: Leverages representations learned from public human movement datasets for clinical gait assessment.
• Improved clinical assessment accuracy: Enhances predictive accuracy relative to models trained solely on limited clinical PD gait data.

Methodology:

Uses a transformer-based model with self-supervised pre-training inspired by GPT-3 on public human motion datasets to perform human motion forecasting, then applies the learned motion features to clinical PD gait data to predict MDS-UPDRS gait impairment severity.