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

Movement velocity is a key indicator for CP risk: The study found that velocity-driven features of the arms, hips, and legs have a "dominant influence" on the model's CP risk predictions, suggesting that how fast infants move certain body parts may be more important than the angles at which they move them for early detection.

Explainability methods show partial but incomplete agreement: The two explanation techniques tested (CAM and Grad-CAM) only showed "partial convergence" in identifying which movements matter for predictions, indicating that even when using explainability tools, clinicians should be cautious about relying on a single method's interpretation of what the AI model is detecting.

Clinical validation is still needed before deployment: The authors explicitly state their findings "warrant further clinical validation," meaning this research demonstrates a promising approach but should not yet be used in actual clinical settings without additional testing to confirm the identified movement patterns truly indicate CP risk.

Targeted testing can reveal what AI models actually use: By deliberately changing specific movements in videos and observing how predictions changed, researchers could verify whether the model was genuinely using the movements it appeared to focus on, rather than relying solely on what the explanation tools claimed—an important quality check for medical AI systems.

Early detection of cerebral palsy could improve outcomes: The research is motivated by the fact that early CP detection "can significantly improve treatment outcomes," making this work potentially impactful for children's health if the approach proves clinically valid.