Support vector machines classifiers of physical activities in preschoolers

Area of Science:

• Biomedical Engineering
• Kinesiology
• Machine Learning

Background:

• Accurate classification of physical activity in preschool children is crucial for health.
• Accelerometer data offers objective measurement but requires robust analytical methods.
• Distinguishing various activities, including sedentary periods, is challenging.

Purpose of the Study:

• To develop, test, and compare multinomial logistic regression (MLR) and support vector machines (SVM) for classifying physical activity in young children.
• To evaluate the performance of MLR and SVM using accelerometer data.
• To determine the optimal classification model for preschool physical activity.

Main Methods:

• Sixty-nine children (3-5 years) wore triaxial accelerometers during supervised activities.
• K-means clustering identified natural data groupings.
• MLR and SVM models were trained and compared using 10-fold cross-validation, with and without sleep data.
• Direct observation served as the criterion method for validation.

Main Results:

• SVM classifiers generally exhibited lower 10-fold cross-validation error rates than MLR.
• The best performing SVM model, including sleep, had a 10-fold CV error rate of 24.70%.
• An SVM model without sleep achieved a 10-fold CV error rate of 20.16% and an overall error rate of 15.56%.

Conclusions:

• Support Vector Machines (SVM) are superior to multinomial logistic regression (MLR) for classifying physical activities in preschool-aged children.
• Accelerometer data, analyzed with SVM, can accurately classify physical activities with an acceptable error rate.
• SVM provides a reliable method for objective physical activity assessment in early childhood.