Machine Learning Analysis of Physical Activity Data to Classify Postural Dysfunction.

Area of Science:

• Biomedical Engineering
• Gerontology
• Data Science

Background:

• Machine learning (ML) analysis of biometric data in uncontrolled environments remains underexplored.
• Postural dysfunction is a significant concern in middle-aged and older individuals.

Purpose of the Study:

• To evaluate the efficacy of ML analysis of physical activity data in classifying postural dysfunction.
• To assess the feasibility of using ML for health monitoring in real-world settings.

Main Methods:

• Utilized a waist-worn uni-axial accelerometer to measure physical activity over one week from the National Health and Nutrition Examination Survey (NHANES) dataset.
• Employed ML models, including Support Vector Machine (SVM) and XGBoost, to predict postural dysfunction based on physical activity features and demographics.
• Evaluated model performance using AUC-ROC, balanced accuracy, and F1-score.

Main Results:

• The study included 1625 participants aged 40 and above, with 47% diagnosed with postural dysfunction.
• ML models achieved AUC values ranging from 0.67 to 0.73, with SVM and XGBoost performing best (AUC=0.73).
• Key predictors included age and accelerometer-derived physical activity metrics.

Conclusions:

• ML analysis of accelerometer-derived physical activity data is a feasible method for classifying postural dysfunction in older adults.
• This approach can be applied in real-world environments, such as participants' homes, for health assessments.