Dual-stream transformer approach for pain assessment using visual-physiological data modeling.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Pain Research

Background:

• Traditional pain assessment relies on subjective self-reporting, limiting objectivity.
• Existing machine learning models often use single data modalities, failing to capture pain's complexity.
• Multimodal approaches integrating diverse data sources offer a more comprehensive pain assessment.

Purpose of the Study:

• To propose a robust dual-stream framework for automatic pain assessment.
• To leverage multimodal data (facial video and physiological signals) for enhanced pain recognition.
• To improve adaptability across diverse clinical scenarios.

Main Methods:

• Utilized masked autoencoder pre-training for facial video and bio-psychological signals.
• Implemented a dual hybrid positional encoding embedding and cross-attention fusion classifier.
• Evaluated the framework on AI4Pain and BioVid datasets for electrode-based and heat-induced pain.

Main Results:

• The proposed dual-stream framework demonstrated superior performance in pain assessment.
• The model effectively captured complex patterns from high-dimensional multimodal data.
• Achieved enhanced robustness and adaptability in classifying pain correlates.

Conclusions:

• The multimodal dual-stream framework offers a significant advancement in automatic pain assessment.
• Integrating facial video and physiological signals improves the accuracy and reliability of pain detection.
• This approach holds promise for more objective and effective clinical pain management.