Development of a machine learning model to detect toddlers' physical activity and sedentary time using...

Toddler movement patterns challenge current accelerometer-based detection of physical activity (PA) and sedentary time (SED). The objectives of this study were to: (1) develop a novel machine learning (ML) model to detect toddlers' PA and SED; and (2) compare this ML model to existing cut-point methods to analyze toddlers' PA (independent sample cross-validation of existing methods).

We recruited 111 toddlers (21 ± 7 months; 51% female) to two 1-hour semi-structured visits wearing a waist-worn ActiGraph wGT3X-BT accelerometer. Video recordings were manually annotated using a modified Children's Activity Rating Scale to determine a ground truth. We extracted 40 time and frequency domain features from raw accelerations and trained 4 gradient boosted tree ML models (distinguishing SED, total PA (TPA), light PA (LPA), moderate-to-vigorous PA (MVPA), and non-volitional movement (NVM)). Models were assessed using accuracy, F1 scores, and confusion matrices. For the validation of 11 existing methods, we calculated accuracy, F1, and mean absolute differences in TPA and MVPA estimation.

ML models classifying NVM/SED/TPA and NVM/SED/LPA/MVPA reached 82% and 74% accuracy with mean absolute differences of 3.0 and 3.2 min/h, respectively. Independent sample cross-validation found accuracies from 33 to 74% and mean absolute differences from 7.6 to 18.6 min/h in TPA and 10.7 to 25.8 min/h in MVPA.

We recommend the NVM/SED/TPA or NVM/SED/LPA/MVPA models' given alignment with toddler TPA guidelines and MVPA link to health outcomes, respectively. We additionally present an open-access interface for using these ML models that does not require coding knowledge. This presents a substantial step forward in the measurement of toddlers' physical activity.