Deep learning empowered sensor fusion boosts infant movement classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Prechtl general movement assessment (GMA) is crucial for diagnosing infant neurological impairments.
• Machine learning aims to standardize GMA but single-sensor deep learning models lag behind human expert performance.
• Current AI approaches for motor pattern classification are limited by proprietary datasets.

Purpose of the Study:

• To introduce and evaluate a sensor fusion approach for assessing fidgety movements (FMs) in infants.
• To determine if a multi-sensor system surpasses single-modality assessments for infant movement classification.
• To enhance the accuracy and scalability of automated infant neurodevelopmental assessment.

Main Methods:

• Compared pressure, inertial, and visual sensors for recording fidgety movements (FMs) in 51 typically developing infants.
• Investigated various sensor combinations and fusion techniques (early and late fusion).
• Utilized convolutional neural network (CNN) architectures for movement pattern classification.

Main Results:

• The three-sensor fusion approach achieved a classification accuracy of 94.5%.
• Multi-sensor fusion significantly outperformed any single sensor modality in classifying infant movements.
• Demonstrated the superior performance of integrated sensor data over individual data streams.

Conclusions:

• Sensor fusion presents a promising method for the automated classification of infant motor patterns.
• A robust sensor fusion system can enhance AI-based early recognition of neurofunctions.
• This technology facilitates the automated early detection of neurodevelopmental conditions.