Automating Infant General Movements Assessment via Multi-Dataset Spatio-Temporal Graph Convolutional Networks

BackgroundThe General Movements Assessment (GMA) is a strong, non-invasive predictor of neurodevelopmental risk in early infancy, but automation studies often rely on small, heterogeneous public datasets, fostering overfitting and limited generalization.MethodsWe propose a unified preprocessing pipeline that standardizes 2D pose time-series from three public cohorts (MINI-RGBD, RVI-38, PMI-GMA; total n=1170) into histogram-encoded joint features. Sequences are represented as skeleton graphs and classified with an adapted multi-stage spatio-temporal graph convolutional network (MS-STGCN). Key hyperparameters (learning rate, weight decay, feature maps, bins, pooling, threshold) are tuned via systematic random search in single-dataset and pooled-dataset experiments.ResultsIn single-dataset evaluations, sensitivity reached up to 96% (with reduced specificity). On the combined cohort, the model achieved 62% accuracy, 83% sensitivity, and 44% specificity, indicating viable cross-dataset generalization with a sensitivity/specificity trade-off.ConclusionStandardizing preprocessing and rigorously tuning a graph-based model across multiple public datasets advances scalable, automated GMA while underscoring persistent specificity challenges in small, heterogeneous infant-movement data.