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

Area of Science:

Neuroscience
Developmental Pediatrics
Biomedical Engineering

Background:

General Movement Assessment (GMA) is crucial for early detection of neurodevelopmental disorders in infants.
Current automated GMA methods often use 2D data and composite indices, limiting detailed analysis of key parameters like complexity, variability, and fluidity.
Existing research lacks public datasets for preterm infant movements and parameter-specific quantification aligned with clinical criteria.

Purpose of the Study:

To address critical gaps in automated GMA by creating a public 3D dataset for preterm infants.
To develop and validate parameter-specific quantification methods for complexity, variability, and fluidity in infant movements.
To explore the utility of deep learning models for analyzing 3D skeletal data in automated GMA.

Main Methods:

Introduced the AGMA dataset: 264 3D movement trajectories from 126 preterm infants (<33 weeks gestational age) with expert annotations.
Developed novel handcrafted features for individual quantification of GMA parameters (complexity, variability, fluidity).
Utilized Random Forest classifiers for parameter classification and evaluated graph convolutional networks (STGCN, DeGCN) on 3D skeletal data.

Main Results:

Handcrafted features achieved AUCs of 0.75 (complexity), 0.77 (variability), and 0.86 (fluidity) using Random Forest.
Spectral entropy demonstrated high discriminative power for fluidity assessment (AUC=0.86, recall=0.83), a novel application in GMA.
Graph convolutional networks showed limited performance, likely due to dataset size, highlighting the need for larger datasets.

Conclusions:

The AGMA dataset and parameter-specific quantification methods provide a reproducible baseline for automated GMA research.
The developed methods enable more precise, clinically relevant automated analysis of infant movements for early neurodevelopmental disorder detection.
Public release of the AGMA dataset facilitates future research, validation, and clinical translation of automated GMA.