3D machine learning-based complexity variability and fluidity quantification of preterm and writhing general movements

General Movement Assessment (GMA) is a reliable non-invasive method for the early detection of neurodevelopmental disorders in infants, based on the qualitative evaluation of three key parameters: complexity, variability, and fluidity. Current automated approaches primarily use composite indices on two-dimensional data, limiting parameter-specific evaluations and failing to fully exploit three-dimensional kinematic information. This study addresses three critical gaps in automated GMA research: (1) the lack of publicly available datasets for preterm and writhing movements, (2) the need for parameter-specific quantification methods aligned with clinical assessment criteria, and (3) the limited exploration of deep learning approaches for 3D skeletal data analysis. We introduce the AGMA dataset containing 264 three-dimensional movement trajectories from 126 preterm infants (born < 33 weeks gestational age), with expert annotations for complexity, variability, and fluidity. We propose novel handcrafted features for individually quantifying each GMA parameter, achieving area under the curve (AUC) values of 0.75, 0.77, and 0.86 for complexity, variability, and fluidity classification, respectively, using Random Forest classifiers. Notably, spectral entropy emerged as a highly discriminative feature for fluidity assessment (AUC = 0.86, recall = 0.83), representing the first application of this measure in the GMA context. We also evaluate graph convolutional networks (STGCN and DeGCN) for movement classification, though their performance was limited by the dataset size. The publicly released AGMA dataset, containing 264 trajectories from 126 infants, along with our parameter-specific classification methods, establishes a reproducible baseline for future automated GMA research and clinical validation studies.