A computer vision and machine learning system for classifying gross motor performance in children with Down syndrome

Background Children with Down syndrome (DS) often experience delays in gross motor development due to hypotonia and motor coordination deficits, which can limit their participation in physical activities. Traditional assessments of motor performance are mainly subjective and may lack consistency across evaluators. Integrating artificial intelligence and computer vision into educational settings presents an opportunity to enhance the objectivity, precision, and scalability of such assessments. This study presents the development and evaluation of an automated system that utilizes pose estimation and machine learning to classify gross motor performance in children with Down syndrome, to support inclusive physical education practices. Methods We recruited 23 children with DS (aged 4–10) from educational institutions affiliated with Cáritas Arequipa, Peru. Participants performed four standardized gross motor exercises: two-footed jumps, ball toss, crawling, and sit-to-stand transitions. High-resolution Canon cameras recorded the movements, and the MediaPipe Holistic framework was used to extract 33 key body landmarks per frame. From these data, spatiotemporal features were computed and reduced using Principal Component Analysis (PCA) to minimize dimensionality while preserving variance. Expert evaluators labeled each exercise attempt into one of three performance categories: high, moderate, or low. Five supervised machine learning algorithms—Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), k-Nearest Neighbors (k-NN), and Logistic Regression—were trained and compared using a stratified 80/20 train-test split. Model performance was evaluated using precision, recall, F1-score, accuracy, and Cohen’s Kappa coefficient. Results All classifiers achieved high accuracy (>93%). The Random Forest model yielded the most balanced and robust results, with an overall accuracy of 95.55%, macro-average F1-score of 0.96, and a Cohen’s Kappa of 0.933, indicating near-perfect agreement with expert labels. Logistic Regression showed slightly higher accuracy (95.79%) but was less effective at modeling nonlinear patterns. Notably, a one-way analysis of variance (ANOVA) on the first principal component (PC1) revealed significant differences between performance classes (F(2, 9,123) = 62.98, p = 7.00 × 10 −28 , η² = 0.0143), confirming PC1 as a discriminative metric for motor performance. Misclassifications occurred primarily between adjacent categories ( e.g ., moderate vs . high), with minimal confusion between high and low classes. Conclusions The proposed system demonstrates strong potential as an objective and scalable solution for monitoring gross motor skills in children with DS using low-cost hardware and open-source tools. Its integration into educational settings may help personalize instruction and support inclusive physical education by providing real-time, data-informed feedback to educators.