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

Pain is a critical clinical indicator in rehabilitation and neurological disorders, yet reliable multi-level recognition remains challenging due to subtle facial variations, inter-subject variability, and heterogeneous clinical data. To address these issues, we propose PainFedMVL, a federated multi-view learning framework for distributed medical settings. The framework integrates Local Binary Patterns from Three Orthogonal Planes (LBP-TOP) to capture illumination-robust spatiotemporal textures and Bi-Weighted Oriented Optical Flow (Bi-WOOF) to encode localized facial micro-dynamics, thereby constructing complementary feature representations. A multi-scale CNN-biLSTM is employed to extract hierarchical spatiotemporal dependencies, enabling robust modeling of both global texture and fine-grained motion cues. To mitigate the non-IID nature of decentralized medical data, we introduce an adaptive aggregation strategy based on Jensen-Shannon (JS) divergence, which explicitly measures the distributional distance between each client and the global model to achieve adaptive weighting. The mechanism reduces the adverse effects of inter-client heterogeneity, stabilizes optimization, and enhances generalization across diverse populations. Experiments on the BioVid dataset demonstrate that PainFedMVL consistently outperforms representative baselines in both binary and multi-level pain classification, providing a robust and privacy-preserving solution for clinical pain assessment.