Confident and Trustworthy Model for Fidgety Movement Classification.

Area of Science:

• Neuroscience
• Developmental Pediatrics
• Artificial Intelligence

Background:

• General movements (GMs) are crucial for assessing infant neurological development up to five months.
• The General Movement Assessment (GMA) qualitatively evaluates GMs, with Fidgety Movements (FM) absence indicating potential cerebral palsy (CP).
• Automated GMA solutions aim to increase accessibility, but current models lack confidence estimation, leading to errors.

Purpose of the Study:

• To develop a deep learning model for automated GMA that classifies movements and abstains from uncertain predictions.
• To improve the reliability and interpretability of automated infant movement analysis.
• To address the issue of overconfident mistakes in current automated GMA systems.

Main Methods:

• A deep learning approach was employed for classifying fidgety and non-fidgety movements.
• Two novel regularization losses were introduced to ensure balanced confidence across movement types.
• The model was designed to selectively abstain from classification when uncertain.

Main Results:

• The proposed model demonstrates the ability to gauge its own classification confidence.
• Regularization losses effectively maintained similar confidence levels across different movement types.
• Selective abstention had minimal impact on video-level coverage, and confident predictions improved overall performance.

Conclusions:

• The developed deep learning model enhances automated GMA by providing confidence-aware classifications.
• This approach improves diagnostic accuracy for conditions like cerebral palsy by reducing overconfident errors.
• The model offers a more reliable and accessible tool for infant neurological development assessment.