Fall Detection in Individuals With Lower Limb Amputations Using Mobile Phones: Machine Learning Enhances Robustness...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning

Background:

• Automatic fall detection using mobile phones offers rapid injury response for at-risk populations like lower limb amputees.
• Previous research primarily focused on able-bodied individuals in lab settings, limiting real-world applicability for amputees.

Purpose of the Study:

• To develop a fall detection classifier using data from able-bodied individuals for use in individuals with lower limb amputations.
• To evaluate the classifier's performance during free-living conditions with the mobile phone in various locations.

Main Methods:

• Collected fall and daily living activity data from 10 able-bodied and 6 lower limb amputee participants.
• Utilized mobile phone accelerometer and gyroscope data, extracting 40 features to train stacked machine learning classifiers.
• Compared population-specific models against a model trained on able-bodied data and tested on amputees, benchmarking against a threshold-based classifier.

Main Results:

• A model trained on able-bodied data achieved comparable accuracy (sensitivity: 0.989, specificity: 0.968) in detecting falls in amputees as a model trained on amputees (sensitivity: 0.984, specificity: 0.965).
• The machine learning model generated significantly fewer false alarms (mean 2.2/day) compared to a threshold-based classifier (mean 122.1/day).
• Fall detection performance was robust across various mobile phone carrying locations and orientations.

Conclusions:

• Mobile phone-based fall detection is feasible using data from non-amputee individuals to identify falls in amputees using prostheses.
• The system demonstrates effectiveness across diverse carrying locations and orientations, approaching real-world applicability.
• Low false alarm rates suggest practical utility for long-term deployment in daily life.