Facilitate Robust Early Screening of Cerebral Palsy via General Movements Assessment With Multi-Modality Co-Learning

Area of Science:

• Neurology
• Artificial Intelligence
• Infant Development

Background:

• General Movement Assessment (GMA) is crucial for early cerebral palsy (CP) detection in infants.
• Traditional GMA relies on subjective physician judgment, limiting accessibility and scalability.
• Existing AI methods for GMA often lack detailed body information, relying solely on motion skeletons.

Purpose of the Study:

• To introduce CoGMA, a multi-modality co-learning framework for enhanced General Movement Assessment.
• To improve the accuracy and efficiency of neuromotor behavior evaluation in infants.
• To address limitations of traditional GMA and current AI approaches.

Main Methods:

• Developed CoGMA, a novel framework integrating skeleton data, clinical information, RGB video, and text descriptions using a multimodal large language model.
• Employed a co-learning strategy during training to enhance representation learning.
• Achieved efficient and accurate prediction during inference using only skeleton data and clinical information.

Main Results:

• CoGMA demonstrated robust performance in evaluating both writhing and fidgety movement stages in infants.
• The framework excelled in zero-shot evaluation of fidget movements, overcoming limited training sample issues.
• CoGMA significantly enhances GMA methodology for early detection of neuromotor disorders.

Conclusions:

• CoGMA offers a significant advancement in infant neuromotor behavior assessment and early cerebral palsy detection.
• The framework's efficiency and accuracy pave the way for broader clinical application and research.
• InfantAnimator tool developed to facilitate anonymized data sharing and collaborative research.