Are Machine Learning Models on Wrist Accelerometry Robust against Differences in Physical Performance among Older...

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Physical activity (PA) is crucial for health and preserving function in aging adults.
• Wrist-worn accelerometers combined with machine learning (ML) show promise for PA assessment, but accuracy challenges persist.
• Robust ML models are needed to estimate age-related physical function and PA accurately.

Purpose of the Study:

• To evaluate ML models (XGBoost, LASSO) for estimating PA measures in older adults.
• To assess model performance in recognizing PA types, intensities, and estimating energy expenditure (EE).
• To determine if physical performance groups (low vs. high) impact model accuracy.

Main Methods:

• Utilized wrist-worn accelerometer data from 247 older adults (57% female, aged 60+).
• Trained ML models (XGBoost, LASSO) to recognize 33 distinct PA types and intensities.
• Evaluated model accuracy in classifying PA and estimating EE, comparing performance between LPP and HPP groups.

Main Results:

• ML models accurately recognized PA types (F1-scores > 0.91) and intensities (F1-scores > 0.84).
• Energy expenditure (EE) estimation yielded a root mean square error of 0.836 ± 0.059 METs.
• XGBoost outperformed LASSO; grouping participants by physical performance did not enhance model accuracy.

Conclusions:

• Wrist-worn accelerometers and ML can accurately recognize PA types/intensities and estimate EE in older adults.
• ML models demonstrate robustness for PA assessment across varying physical function levels in older populations.
• Future research should focus on improving individual activity recognition for enhanced PA monitoring.