PainFedMVL: A Federated Multi-View Learning Approach for Multi-Level Pain Recognition.

Area of Science:

• Artificial Intelligence in Medicine
• Computer Vision for Healthcare
• Machine Learning for Clinical Diagnostics

Background:

• Pain assessment is crucial in rehabilitation and neurology but faces challenges due to subtle facial cues, patient variability, and data heterogeneity.
• Existing methods struggle with reliable multi-level pain recognition in distributed medical environments.

Purpose of the Study:

• To develop a robust and privacy-preserving federated multi-view learning framework for accurate, distributed pain assessment.
• To address the challenges of inter-subject variability and non-IID data in clinical pain recognition.

Main Methods:

• Proposed PainFed-MVL framework integrating Local Binary Patterns from Three Orthogonal Planes (LBP-TOP) for texture and Bi-Weighted Oriented Optical Flow (Bi-WOOF) for micro-dynamics.
• Utilized a multi-scale CNN-biLSTM for hierarchical spatiotemporal feature extraction.
• Implemented an adaptive aggregation strategy using Jensen-Shannon (JS) divergence to handle non-IID decentralized data.

Main Results:

• PainFed-MVL demonstrated superior performance over baseline methods in both binary and multi-level pain classification on the BioVid dataset.
• The adaptive aggregation strategy effectively mitigated the impact of data heterogeneity, stabilizing optimization and improving generalization.
• The framework achieved robust and privacy-preserving pain assessment.

Conclusions:

• PainFed-MVL offers a significant advancement in automated clinical pain assessment, particularly in distributed and privacy-sensitive settings.
• The integration of complementary features and adaptive federated learning addresses key limitations in current pain recognition technologies.
• This approach provides a foundation for more reliable pain monitoring in rehabilitation and neurological disorder management.