Machine Learning Applications for Physical Activity and Behaviour in Early Childhood: A Systematic Review

This systematic review evaluated machine learning applications for analysing physical activity and behaviour in preschool children using accelerometer data. Following the PRISMA guidelines, we systematically searched PubMed, FECYT, and ProQuest Central databases. Fourteen studies implementing machine learning approaches for preschool accelerometry data were identified and assessed using the MINORS scale. Studies focused on two primary domains: physical activity analysis (n = 10) and sleep monitoring (n = 4). The ActiGraph GT3X+ was predominantly used, with placement varying between the hip and wrist. Random Forest algorithms proved most effective, achieving accuracy rates up to 86.4% in activity classification and 96.2% in sleep prediction. Sampling frequencies (0.25–100 Hz) and epoch lengths (1–60 s) varied considerably across studies. Machine learning applications show promising results for preschool physical activity assessment. However, small sample sizes and methodological inconsistencies limit generalizability. Future research should prioritise larger cohorts, explore multiple sensor integrations, and develop standardised protocols to enhance practical applications.