Fine-Grained Fidgety Movement Classification Using Active Learning.

Area of Science:

• Developmental neuroscience
• Computer vision
• Machine learning

Background:

• Infant fidgety movements (9-20 weeks corrected age) are key indicators of typical development.
• General Movement Assessment (GMA) by trained professionals is the standard but faces accessibility issues.
• Existing computer vision solutions often lack interpretability by not modeling movement dynamics.

Purpose of the Study:

• To develop a computer vision approach for directly modeling and classifying infant fidgety movements.
• To improve model interpretability by focusing on explanatory movement factors.
• To address the challenge of limited labeled data for short infant movements.

Main Methods:

• Proposed a novel method to directly model and classify short infant movements as fidgety or non-fidgety.
• Utilized active learning to minimize the need for labeled data, focusing on informative examples.
• Validated the framework by analyzing hip movements in typically developing and at-risk infant cohorts.

Main Results:

• Active learning proved suitable for modeling infant movements.
• The approach achieved adequate performance even with labels from novice annotators.
• Demonstrated the feasibility of training interpretable models for infant movement analysis.

Conclusions:

• Active learning is an effective strategy for training computer vision models for infant fidgety movement assessment.
• This approach enhances the potential for accessible, automated infant movement analysis.
• The method offers improved interpretability compared to direct video-to-status mappings.