Towards Biomarker Discovery for Early Cerebral Palsy Detection: Evaluating Explanations Through Kinematic Perturbations

TL;DR

The paper tackles early cerebral palsy detection from infant movement by combining a skeleton-based GCN CP predictor with a biomechanically-informed perturbation framework to evaluate explainable AI methods. It compares CAM and Grad-CAM explanations, identifies velocity-driven limb features as dominant contributors to CP risk, and demonstrates how controlled perturbations can reveal potential movement-based biomarkers. The work advances clinically relevant, explainable AI for pediatric motor outcome prediction and provides a foundation for biomarker discovery that warrants prospective clinical validation. Limitations include the absence of ground-truth biomechanical labels and recognition of head-movement complexity, with future work proposed to align perturbations with qualitative movement assessments like MOS and to refine head perturbations for improved interpretability.

Abstract

Cerebral Palsy (CP) is a prevalent motor disability in children, for which early detection can significantly improve treatment outcomes. While skeleton-based Graph Convolutional Network (GCN) models have shown promise in automatically predicting CP risk from infant videos, their "black-box" nature raises concerns about clinical explainability. To address this, we introduce a perturbation framework tailored for infant movement features and use it to compare two explainable AI (XAI) methods: Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM). First, we identify significant and non-significant body keypoints in very low- and very high-risk infant video snippets based on the XAI attribution scores. We then conduct targeted velocity and angular perturbations, both individually and in combination, on these keypoints to assess how the GCN model's risk predictions change. Our results indicate that velocity-driven features of the arms, hips, and legs have a dominant influence on CP risk predictions, while angular perturbations have a more modest impact. Furthermore, CAM and Grad-CAM show partial convergence in their explanations for both low- and high-risk CP groups. Our findings demonstrate the use of XAI-driven movement analysis for early CP prediction and offer insights into potential movement-based biomarker discovery that warrant further clinical validation.

Figures (9)

• Figure 1: Overview of CP prediction ensemble pipeline, showing (\ref{['fig:flow']}) where the CP risk result and XAI attribution scores are obtained, and where the data perturbation is added for this study. The figure in (\ref{['fig:ensemble']}) shows the flow of data in each model in the CP prediction ensemble, and where the individual XAI attributions and CP risk result are obtained. The attribution scores are calculated from Equations \ref{['eq:cam']} and \ref{['eq:gradcam']}.
• Figure 2: (left) Sample visualization of attribution scores from the ensemble model's XAI methods tested on the same video snippet. Green indicates very low attribution scores, yellow for low scores, orange for high scores, and red (not present in this example) for very high scores relative to the defined threshold scores (about 0.29 for CAM and 0.17 for Grad-CAM). (right) Numerical values of attribution scores, with $k$ denoting the body keypoint number.
• Figure 3: The set of top-k and non-top-k joints identified using (\ref{['fig:cam_top-k']}) CAM, (\ref{['fig:gcam_top-k']}) Grad-CAM. Joints selected as significant by both the Kneedle Method and K-Means Clustering are classified as top-k, while the remaining joints are categorized as non-top-k. Although CAM and Grad-CAM differ in the specific joints they flag as significant, a consistent pattern emerges: in low-risk windows, the extremities (arms and legs) seem to be most influential, whereas in high-risk windows, the head, neck, and shoulders seem to carry more significance. These sets of top-k and non-top-k joints are used throughout the experiment.
• Figure 4: Definition of infant skeletal segments applied in velocity perturbation. If one keypoint in a segment is selected for perturbation, the entire segment is perturbed together to preserve bone length and maintain anatomically realistic movement.
• Figure 5: Velocity perturbation analysis under \ref{['fig:cam_velocity']}) CAM and \ref{['fig:gcam_velocity']}) Grad-CAM. Each sub-plot shows how scaling the velocity of either top-k (highlighted in teal in the skeletal diagram on the right) or non-top-k joints affects the model’s predicted CP risk for low and high risk windows, with the median prediction (solid line) and interquartile range (blue shaded region). The horizontal dashed line represents the model’s CP risk threshold. $min$ refers to the 5th-percentile velocity, while $max$ refers to the 95th-percentile velocity.