A Simple Decision Tree Model for Detection of Movement Behaviors in Non-Ambulatory Children with Cerebral Palsy

Background Children and adolescents with cerebral palsy (CP) who are non-ambulatory are particularly vulnerable to the negative health impacts of physical inactivity and sedentary behaviour. Yet, there is a paucity of validated methods to measure movement behaviours and interventions to reduce sedentary behaviour in this patient group. Accelerometer-based motion sensors have become the method of choice for measuring movement behaviours in both typically developing children as well as those with CP. However, interpretable and accessible physical activity classification models for children with more severe mobility limitations are lacking. The purpose of this study was to develop and test a simple decision tree (DT) model to classify five physical activity types in children with non-ambulant CP using a single thigh-mounted accelerometer. Methods Twenty-one children (mean age 9.4 ± 3.6 y) with CP classified at GMFCS level IV/V completed a series of data acquisition sessions designed to elicit a broad range of activities. During each session, participants wore a SENS motion accelerometer on the least affected thigh. Tri-axial accelerometer data (12.5 Hz) were segmented into 4-sec windows and annotated with ground truth labels derived from a bespoke direct observation scheme. Four time-domain features served as inputs to a handcrafted DT model. Feature thresholds were identified through visual inspection of probability density plots and ROC curve analyses in the training sample and evaluated for accuracy in the testing sample. Results Within the testing sample, the DT exhibited substantial agreement with ground truth activity with F1 statistics exceeding 0.70. In comparison, the SENS proprietary model exhibited only moderate agreement. When applied to data from an independent sample performing a different set of activities, the DT demonstrated excellent recognition of sit/lie, stepping, and cycling. However, seated throwing and catching and manual wheelchair propulsion were predominantly misclassified as sedentary. Conclusions This study provides a significant advance toward accurate, interpretable, and accessible device-based measurement of physical activity and sedentary behaviors in children with CP who are non-ambulatory. The DT model can be seamlessly integrated into smartphone applications or other technology platforms, empowering children with CP, their parents and carers, and support workers to monitor movement behaviors in real time.