Learning Developmental Age from 3D Infant Kinetics Using Adaptive Graph Neural Networks

TL;DR

KA predicts infant corrected age from spontaneous motor patterns using pose-estimated 3D data and an adaptive graph neural network that models spatio-temporal dependencies. The approach integrates multi-stream features (joint coordinates, bone directions, velocity, acceleration) within a single graph, and leverages ST-GCN/AAGCN blocks with per-sample attention to optimize age estimation. Experiments show 3D data and adaptive graph structures outperform 2D baselines and hand-crafted feature indices, enabling KA and KA-gap to distinguish typical and at-risk infants. The work provides a publicly released 3D infant kinetics dataset and growth-chart style KA predictions, underscoring potential for early screening while calling for larger, longitudinal validation.

Abstract

Reliable methods for the neurodevelopmental assessment of infants are essential for early detection of problems that may need prompt interventions. Spontaneous motor activity, or 'kinetics', is shown to provide a powerful surrogate measure of upcoming neurodevelopment. However, its assessment is by and large qualitative and subjective, focusing on visually identified, age-specific gestures. In this work, we introduce Kinetic Age (KA), a novel data-driven metric that quantifies neurodevelopmental maturity by predicting an infant's age based on their movement patterns. KA offers an interpretable and generalizable proxy for motor development. Our method leverages 3D video recordings of infants, processed with pose estimation to extract spatio-temporal series of anatomical landmarks, which are released as a new openly available dataset. These data are modeled using adaptive graph convolutional networks, able to capture the spatio-temporal dependencies in infant movements. We also show that our data-driven approach achieves improvement over traditional machine learning baselines based on manually engineered features.

Figures (6)

• Figure 1: Overview of the study pipeline. The pipeline begins with data collection, where 3D video recordings of spontaneous infant movements are captured using a standardized setup. Pose estimation is applied to generate skeletal videos, extracting spatio-temporal time series of 18 anatomical landmarks. The data is preprocessed through centering and rotation normalization. Finally, Kinetic Age (KA) is predicted using an adaptive graph neural network (GNN) that models spatio-temporal dependencies in the movement patterns.
• Figure 2: Preprocessing strategy for infant skeletal graphs. The panels shows the average original graph of one recorded segment at each rotation step, with the spine vector indicated in red and the backline vector (average of hipline and shoulder line) in green.
• Figure 3: Three different graph initializations for the AAGCN model. The edge weights are colored according to the normalized version of the centrifugal group in the graph partitioning strategy.
• Figure 4: Illustration of the network architecture. The diagram shows the ordering of layers in the STGCN/AAGCN models. The dimensions of the input tensor (including batch size), are highlighted underneath those layers that affect them.
• Figure 5: Evaluation of the STGCN model performance as a function of temporal kernel size. The left plot shows the RMSE against $K_t$ the middle plot depicts $R^2$ scores, and the right plot illustrates how the number of model parameters increases linearly for larger convolution kernels.