Data-Driven Early Prediction of Cerebral Palsy Using AutoML and interpretable kinematic features.

Area of Science:

• Computational Neuroscience
• Developmental Pediatrics
• Machine Learning in Healthcare

Background:

• Early identification of cerebral palsy (CP) is crucial but challenging due to time-intensive expert assessments.
• Current machine learning approaches often predict CP scores, not direct clinical risk.
• There is a need for scalable, data-driven methods for early neurodevelopmental risk assessment.

Purpose of the Study:

• To develop and validate a machine learning pipeline for predicting CP risk using infant movement data.
• To assess the generalizability of early motor representations to long-term neurodevelopmental outcomes.
• To leverage AutoML for robust and scalable CP risk prediction from video motion tracking.

Main Methods:

• Extracted movement features from 3- to 4-month-old infant videos using motion tracking.
• Employed AutoML (AutoSklearn) for data-driven feature extraction and model selection to minimize overfitting.
• Utilized pre-registered lock-box validation for rigorous performance evaluation.
• Integrated pre-trained vision transformers for enhanced feature extraction.

Main Results:

• The model achieved an ROC-AUC of 0.78 in predicting the General Movements Assessment (GMA) clinical score.
• The same model, without retraining, predicted CP risk at later follow-ups with an ROC-AUC of 0.74.
• Kinematic movement features effectively captured clinically relevant variability and generalized to long-term neurodevelopmental risk.

Conclusions:

• AutoML-powered movement analytics offer a scalable and interpretable approach for early neurodevelopmental screening.
• Video-derived movement features can provide an accurate and generalizable method for assessing long-term CP risk.
• This data-driven approach enhances early risk detection for neurodevelopmental disorders using readily available infant videos.