Validation of accuracy of SVM-based fall detection system using real-world fall and non-fall datasets

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Falls are a significant cause of injury and mortality in older adults, often resulting in prolonged immobility and secondary complications.
• Current wearable fall detection systems show high accuracy in lab settings but lack validation in real-world environments.
• Automatic fall detection aims to reduce the time individuals spend on the floor after a fall, mitigating associated risks.

Purpose of the Study:

• To evaluate the accuracy and reliability of a wearable fall detection system in a real-world setting using actual fall and non-fall data.
• To assess the system's performance in detecting falls among older adults during their daily activities.
• To compare the system's performance, including fall detection rate and false alarm rate, against existing fall detection technologies.

Main Methods:

• Utilized tri-axial accelerometers worn by 19 older adults and 5 young adults during daily activities.
• Collected approximately 400 hours of activity data, including 10 unanticipated falls experienced by older adults.
• Employed a Support Vector Machine (SVM) machine learning classifier to analyze accelerometer data and distinguish between fall and non-fall events.

Main Results:

• The system successfully detected 8 out of 10 falls (80% accuracy) in older adults using a single accelerometer (waist or sternum).
• No false alarms were recorded during approximately 28.5 hours of data from young adults.
• Older adults experienced a false positive rate ranging from 0 to 0.3 alarms per hour.

Conclusions:

• The developed fall detection system demonstrates promising real-world accuracy and a reduced false positive rate compared to existing systems.
• Further validation studies with larger real-world datasets are necessary to optimize fall detection systems for the target population.
• Continuous data collection in real-world environments is crucial for enhancing the reliability and effectiveness of wearable fall detection technology.