Background: Literature related to objective measurement of habitual physical activity (PA) disproportionately over represents children with Cerebral Palsy (CP) who are ambulant. Consequently, it is unknown if methods used to examine PA, such as machine learning models built on accelerometer data, are able to accurately detect PA in children with CP who use mobility aids for ambulation. Objective: To develop and test machine learning models used for the automatic detection and classification of PA type in children with CP who use mobility aids for ambulation. Methods: Eleven children and adolescents with CP, age 11±3yrs (range 6-16yrs); six females; Gross Motor Function Classification System (GMFCS) III: n=5 and IV: n=6 participated. Participants completed six PA trials of increasing intensity while wearing an ActiGraph GT3X+ accelerometer on the wrist, hip and thigh. PA trials included: supine rest, seated colouring, seated ball throwing, overground walking with a mobility aid, wheelchair propulsion and riding on a modified tricycle. Decision Tree (DT), Support Vector Machine (SVM) and Random Forest (RF) classifiers were trained on 40 features in the vector magnitude of raw acceleration signal using 5s non-overlapping windows. Performance was evaluated using leave-one-subject-out cross validation. Comparisons of performance were subsequently made between all single placement models, all combinations of two placement models, and models trained on data from all three placements. Results: The best performing single-placement model was a RF classifier trained on wrist features, yielding an overall prediction accuracy of 79%. The best performing model built on a combination of two placements was a RF classifier trained on wrist and hip features, yielding an overall prediction accuracy of 92%. The combinations of multiple accelerometer placements were significantly more accurate than a single monitor alone. Models based on the combination of two placements were more accurate than those based on a combination of three placements; however, this difference was not significant. Limitations: The PA protocol consisted of structured activity trials performed in a controlled, clinical environment. Thus, the performance of the models under free living conditions require further investigation. The sample size used may limit the generalisability and robustness of the findings given the variability in movement patterns of the population of interest. Conclusions: Machine learning techniques afford robust and accurate classification of PA in children with CP who use mobility aids for ambulation (GMFCS III & IV) within a laboratory setting. This is significant, as it is the first study to develop methods for objectively measuring habitual PA in this population. Future research should investigate performance of the methods utilised in the current project in children engaged in free living conditions.