Personalised Gait Recognition for People with Neurological Conditions.

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Wearable inertial sensors facilitate free-living gait monitoring in neurological conditions.
• Accurate gait instance recognition is crucial for gait assessment in uncontrolled environments.
• Existing methods like wavelet transforms and general machine learning models have limitations for diverse or impaired gait patterns.

Purpose of the Study:

• To propose and evaluate a lightweight, personalized machine learning approach for gait recognition in individuals with severe motor impairments.
• To address the unsuitability of general models for distinct gait patterns in neurological conditions.
• To improve the accuracy of gait instance recognition for outlier cases.

Main Methods:

• Developed a personalized machine learning approach by fine-tuning a general model with individual patient data.
• Conducted a comparative evaluation against general neural network (NN) and convolutional neural network (CNN) models.
• Assessed performance based on overall accuracy and recognition improvement for underrepresented participants.

Main Results:

• Personalized models improved overall accuracy by 3.5% (NN) and 5.3% (CNN) compared to general models.
• For participants with extreme gait patterns, personalized approaches boosted recognition accuracy by up to 16.9% (NN) and 20.5% (CNN).
• Demonstrated superior performance of personalized models, particularly for individuals with highly individual motor patterns.

Conclusions:

• Personalized machine learning offers a more effective alternative to general models for gait recognition in neurological conditions with severe motor impairments.
• The findings highlight the need for individualized approaches that account for unique gait characteristics.
• Encourages further research into personalized methods for gait analysis in clinical populations.