Improving Fall Classification Accuracy of Multi-Input Models Using Three-Axis Accelerometer and Heart Rate...

Area of Science:

• Biomedical Engineering
• Gerontology
• Artificial Intelligence

Background:

• Aging leads to reduced mobility and muscle strength, increasing fall risks and potential injuries.
• Current fall detection methods often rely on single data sources, limiting accuracy.
• Continuous monitoring for fall prevention in the elderly is a significant challenge.

Purpose of the Study:

• To develop and validate a multi-input deep learning model for accurate fall detection and movement classification in the elderly.
• To integrate data from three-axis acceleration sensors and Holter electrocardiographs for enhanced fall detection.
• To leverage heart rate variability (HRV) and baroreflex characteristics for improved classification accuracy.

Main Methods:

• Implemented a deep learning model (CNN-LSTM) to analyze acceleration sensor data for movement patterns.
• Utilized a wide learning model to analyze heart rate variability (HRV) data, incorporating baroreflex characteristics.
• Developed a multi-input wide and deep learning model combining acceleration and HRV data for fall classification.

Main Results:

• The multi-input model demonstrated significantly improved accuracy in fall classification compared to conventional methods.
• Achieved a precision, recall, and F1 score of 0.91 for fall detection across 15 different movements.
• Observed distinct HRV changes during falls and specific movements like standing from a chair, reflecting baroreflex activity.

Conclusions:

• The proposed multi-input model effectively enhances fall classification accuracy in the elderly by integrating kinematic and cardiac data.
• This approach shows promise for developing more reliable fall prevention systems.
• The findings highlight the utility of baroreflex characteristics derived from ECG and accelerometer data for fall detection.