Machine Learning for Early Risk Stratification and Longitudinal Outcomes in Children with Cerebral Palsy in Low-Resource Settings: A Systematic Review

Abstract

Cerebral palsy is a major cause of physical disability in childhood and has a long-term impact on motor, functional, communication and developmental problems. Early detection, regular follow-up, timely intervention, rehabilitation planning and family support are all crucial but may not be available in low-resource communities. This systematic review aimed to summarizes evidence on machine learning methods for early risk stratification and longitudinal prediction of outcomes in children with cerebral palsy and to identify if these methods could be relevant to health systems with limited resources. Studies published between 2015 and 2026 were selected in PubMed, Scopus, Web of Science, IEEE Xplore, ScienceDirect and Google Scholar, and older foundational documents were included if necessary for the review, based on PRISMA. Eligible studies included those using machine learning, artificial intelligence, or predictive modelling to diagnose, predict and analyses gait, classify patients in terms of function, monitor rehabilitation, predict treatment response, or follow up patients over time. After screening, 18 studies were included. From the evidence reviewed, the use of deep learning, neural networks, random forest, XGBoost, video-based movement analysis, and sensor-based approaches were identified. Early cerebral palsy prediction, gait-event detection, gait-phase recognition, gross motor function prediction and functional outcome estimation showed promising performance results. The evidence base was, however, predominantly in high-resource contexts with some external validation and little firsthand evidence from lower-resource health systems. There was also low confidence in generalizability due to small datasets, specialized equipment, differing outcome measures, and limited implementation in real-world applications. The use of video-based approaches, simple clinical data, and lower-cost sensors could have more potential in low-resource settings but needs to be locally validated in these settings prior to their routine use. Machine learning should thus be seen as a clinical decision support tool for early referral, rehabilitation monitoring and long-term care planning, not to supplant clinical judgement.