ViT-BiLSTM Multimodal Learning for Paediatric ADHD Recognition: Integrating Wearable Sensor Data with Clinical...

Area of Science:

• Computational neuroscience
• Biomedical engineering
• Machine learning

Background:

• Traditional ADHD classification relies on limited tabular accelerometer data, missing dynamic activity patterns.
• Existing methods may underperform due to incomplete feature representation.

Purpose of the Study:

• To develop a multimodal deep learning framework for enhanced ADHD classification.
• To integrate dynamic activity patterns from accelerometer data with static clinical information.

Main Methods:

• Collected accelerometer data (Apple Watches) and clinical measures from children (7-13 years).
• Developed a deep learning framework transforming signals into images for analysis.
• Evaluated various fusion strategies and deep learning models (e.g., ViT-BiLSTM) for optimal integration.

Main Results:

• The multimodal approach significantly improved ADHD classification accuracy.
• The ViT-BiLSTM model with cross-attention fusion demonstrated superior performance.
• Combining dynamic activity images with clinical variables proved effective.

Conclusions:

• Multimodal learning offers a robust strategy for ADHD classification.
• Integrating spatial-temporal activity dynamics with clinical data enhances diagnostic capabilities.
• The developed framework and code will be publicly available to promote research reproducibility.