Multi-Level Adversarial Spatio-Temporal Learning for Footstep Pressure Based FoG Detection.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Freezing of gait (FoG) is a debilitating symptom of Parkinson's disease, significantly impacting patient mobility and quality of life.
• Current methods for FoG assessment often lack objectivity and scalability, necessitating advanced technological solutions.
• Pressure-sensitive gait mats offer a non-invasive approach to capture detailed footstep dynamics for gait analysis.

Purpose of the Study:

• To develop and validate a novel deep learning model for the accurate detection and quantification of freezing of gait (FoG) using footstep pressure data.
• To create a subject-independent model that generalizes well to unseen individuals, overcoming inter-subject variability.
• To establish a foundation for computer-aided tools that can aid in the clinical and home-based evaluation of FoG.

Main Methods:

• Formulated FoG detection as a sequential modeling task.
• Proposed a novel deep learning architecture, the Adversarial Spatio-temporal Network (ASTN).
• Implemented an adversarial training scheme with a multi-level subject discriminator to achieve subject-independent feature learning.

Main Results:

• The ASTN model demonstrated robust performance in detecting FoG from footstep pressure sequences.
• Achieved an Area Under the Curve (AUC) of 0.85 in experiments involving 21 subjects and 393 trials.
• Significantly outperformed conventional machine learning methods in FoG detection accuracy.

Conclusions:

• The Adversarial Spatio-temporal Network (ASTN) represents a significant advancement in footstep pressure-based freezing of gait detection.
• The proposed subject-independent representation learning approach effectively mitigates overfitting and enhances generalizability.
• This study provides a promising deep learning framework for objective and scalable FoG assessment in Parkinson's disease.