Can Wearable Devices and Machine Learning Techniques Be Used for Recognizing and Segmenting Modified Physical...

Area of Science:

• Gerontology
• Biomedical Engineering
• Rehabilitation Science

Background:

• Assessing physical performance is crucial for predicting frailty in older adults.
• The modified Physical Performance Test (mPPT) is a clinical tool evaluating nine distinct activities.
• Current mPPT assessment relies on manual duration measurement, limiting frequent monitoring.

Purpose of the Study:

• To develop and evaluate a wearable sensor system for automatic recognition and duration prediction of mPPT activities.
• To explore the impact of machine learning models, sensor placement, and sampling frequencies on system performance.
• To enable frequent and objective monitoring of physical performance for timely interventions.

Main Methods:

• Five wearable devices (accelerometers, gyroscopes) were placed on the waist, wrists, and ankles of eight participants.
• A non-causal six-stage temporal convolutional network was employed for activity recognition and duration prediction.
• System performance was optimized by testing various machine learning models, sensor placements, and sampling frequencies (6.25 Hz).

Main Results:

• The optimal system configuration utilized signals from the left wrist and right ankle at 6.25 Hz.
• Duration prediction errors varied across mPPT items, ranging from 0.63±0.29 s for turning 360° to 8.21±16.41 s for walking.
• The proposed system demonstrated potential for automatic recognition and segmentation of mPPT activities.

Conclusions:

• Wearable sensor technology can accurately automate the assessment of physical performance tests like the mPPT.
• The developed system facilitates objective and frequent monitoring, supporting personalized frailty interventions.
• Future research will focus on enhancing recognition for specific tasks (e.g., lifting a book) and integrating frailty score prediction.