Wearable airbag technology and machine learned models to mitigate falls after stroke

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Falls are a significant complication post-stroke, impacting mobility and health.
• Current wearable airbag fall detection systems lack validation for stroke survivors.
• Stroke-related motor impairments may hinder the accuracy of existing fall detection technologies.

Purpose of the Study:

• To investigate the necessity of population-specific training data and modeling parameters for pre-fall detection in chronic stroke patients.
• To address the gap in fall detection technology for individuals with stroke-related motor impairments.

Main Methods:

• Collected inertial measurement unit (IMU) data from wearable airbags of stroke (n=20) and control (n=15) groups.
• Recorded 842 falls and 961 non-falls in a controlled laboratory setting.
• Utilized a leave-one-subject-out cross-validation with adaptive boosting classifiers trained on cohort-specific data (stroke vs. control).

Main Results:

• The stroke-trained model demonstrated statistically significant higher recall (0.905) compared to the control-trained model (0.800) (P=0.0035).
• Stroke-trained models showed a 35% higher F1-score for anterior-posterior falls (P=0.019).
• Using activities of daily living as non-fall data improved anterior-posterior fall classification (AUC, P<0.04), with potential greater benefit for individuals with severe stroke.

Conclusions:

• Population-specific data, appropriate non-fall data, and optimal lead time are crucial for effective machine-learned pre-impact fall detection in stroke survivors.
• Existing fall mitigation technologies require validation with data from neurologically impaired individuals.
• Development of fall detection systems should consider diverse high-risk populations.