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

• Main Authors: Bai, C. (Author), Casanova, R. (Author), Manini, T.M (Author), Mardini, M.T (Author), Saldana, S. (Author), Wanigatunga, A.A (Author)
• Format: Article
• Language: English
• Published: MDPI 2022
• Subjects: accelerometer; Accelerometers; Electric batteries; Energy; energy ex-penditure; Energy expenditure; Energy ex-penditure; eXtreme Gradient Boosting; Extreme gradient boosting; Gradient boosting; Machine learning; Machine learning models; Mean square error; physical activity; Physical activity; Physical performance; short physical performance battery; Short physical performance battery; Walking aids; wrist; Wrist
• Online Access: View Fulltext in Publisher

 Sufficient physical activity (PA) reduces the risk of a myriad of diseases and preserves physical capabilities in later life. While there have been significant achievements in mapping accelerations to real‐life movements using machine learning (ML), errors continue to be common, particularly for wrist‐worn devices. It remains unknown whether ML models are robust for estimating age‐related loss of physical function. In this study, we evaluated the performance of ML models (XGBoost and LASSO) to estimate the hallmark measures of PA in low physical performance (LPP) and high physical performance (HPP) groups. Our models were built to recognize PA types and intensities, identify each individual activity, and estimate energy expenditure (EE) using wrist-worn accelerometer data (33 activities per participant) from a large sample of participants (n = 247, 57% females, aged 60+ years). Results indicated that the ML models were accurate in recognizing PA by type and intensity while also estimating EE accurately. However, the models built to recognize individual activities were less robust. Across all tasks, XGBoost outperformed LASSO. XGBoost obtained F1‐Scores for sedentary (0.932 ± 0.005), locomotion (0.946 ± 0.003), lifestyle (0.927 ± 0.006), and strength flexibility exercise (0.915 ± 0.017) activity type recognition tasks. The F1‐Scores for recognizing low, light, and moderate activity intensity were (0.932 ± 0.005), (0.840 ± 0.004), and (0.869 ± 0.005), respectively. The root mean square error for EE estimation was 0.836 ± 0.059 METs. There was no evidence showing that splitting the participants into the LPP and HPP groups improved the models’ performance on estimating the hallmark measures of physical activities. In conclusion, using features derived from wrist‐worn accelerometer data, machine learning models can accurately recognize PA types and intensities and estimate EE for older adults with high and low physical function. © 2022 by the authors. Licensee MDPI, Basel, Switzerland.