Evaluation of self-supervised pre-training for automatic infant movement classification using wearable movement...

Area of Science:

• Biomedical Engineering
• Developmental Pediatrics
• Machine Learning in Healthcare

Background:

• Infant motor performance assessment is crucial for early detection of developmental issues.
• Current methods often lack objectivity and scalability in out-of-hospital settings.
• The MAIJU wearable offers automated, objective infant motor evaluation.

Purpose of the Study:

• To investigate the impact of self-supervised pre-training on MAIJU data classification accuracy.
• To determine if context-selective quality-screening of pre-training data further improves classifier performance.
• To enhance the reliability and robustness of automated infant motor analysis.

Main Methods:

• Utilized self-supervised learning techniques for pre-training classification models on unlabeled MAIJU data.
• Implemented context-selective quality-screening to filter pre-training datasets, excluding periods of minimal infant movement or sensor data loss.
• Evaluated classifier performance based on accuracy improvements after pre-training and data screening.

Main Results:

• Self-supervised pre-training robustly increased the accuracy of infant posture and movement classification models.
• Context-selective quality-screening of pre-training data led to substantial additional performance gains.
• The findings demonstrate the effectiveness of advanced machine learning for wearable-based infant analysis.

Conclusions:

• Self-supervised learning is a viable method to enhance the accuracy of infant motor ability evaluation using smart wearables.
• Optimizing pre-training data through quality-screening significantly boosts classifier performance.
• This approach supports more reliable out-of-hospital developmental monitoring and clinical decision-making.