Augmenting a ResNet + BiLSTM Deep Learning Model with Clinical Mobility Data Helps Outperform a Heuristic...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning in Healthcare

Background:

• Wearable sensors are crucial for gait assessment in various environments.
• Detecting walking in free-living settings, particularly for clinical groups, remains difficult.
• Existing machine learning models often lack generalizability due to training on healthy subjects.

Purpose of the Study:

• To enhance a machine learning model for robust gait identification using diverse datasets.
• To evaluate the model's performance against a heuristic frequency-based algorithm.
• To improve walking detection for clinical populations with impaired gait.

Main Methods:

• An existing machine learning model was retrained using progressively diverse datasets, including healthy and osteoarthritis populations.
• The model's walking identification performance was assessed using a frequency-based gait detection algorithm.
• Model accuracy and recall were evaluated on held-out test data.

Main Results:

• The model trained on all datasets achieved 96% accuracy in activity classification.
• Performance was comparable to the heuristic method for general walking bout identification.
• The machine learning model demonstrated superior recall (>0.89) for slow gait speeds (<0.8 m/s) where the heuristic method failed (recall as low as 0.38).

Conclusions:

• Training machine learning models with diverse datasets significantly enhances their robustness and generalizability.
• This approach is vital for developing accurate gait analysis tools for clinical applications.
• The enhanced model provides reliable walking detection, especially for individuals with slower gait.