Enhanced Infant Movement Analysis Using Transformer-Based Fusion of Diverse Video Features for Neurodevelopmental...

Area of Science:

• Developmental neuroscience
• Computational neuroscience
• Pediatric neurology

Background:

• Early detection of neurodevelopmental abnormalities is crucial for intervention and optimizing infant outcomes.
• Accurate, cost-effective infant diagnostic methods are needed due to data heterogeneity and condition variability.
• Current methods for analyzing infant movement patterns face challenges in accuracy and cost.

Purpose of the Study:

• To develop and validate a novel, open-source deep learning method for classifying infant movement patterns.
• To enhance the early detection and prediction of neurodevelopmental delays using advanced AI.
• To analyze the contribution of different video features within a deep learning model.

Main Methods:

• Recruited twelve parent-infant pairs (infants 3-12 months).
• Captured 25 minutes of 2D video during natural play interactions.
• Developed a transformer-based fusion model integrating multiple video features for movement pattern analysis.

Main Results:

• The deep learning model achieved over 90% accuracy, outperforming traditional methods.
• Sensitivity analysis indicated transformer and CNN components were more critical than pose estimation.
• The method provides robust, accurate, and low-cost analysis of infant movement.

Conclusions:

• The novel deep learning approach significantly improves the analysis of infant movement patterns.
• This method holds promise for enhancing early detection and prediction of neurodevelopmental delays.
• The study offers insights into the efficacy of transformer-based fusion models for diverse video feature integration.