Abstract

Accurate pain intensity recognition is vital for improving clinical care, especially in scenarios where patients cannot self-report. However, although existing datasets are often imbalanced, the main challenge is that current optimization and loss functions lack sensitivity to minority classes, adapt poorly to intra-class variability, and are less robust under imbalance, leading to biased recognition performance. To address these challenges, this study proposes a hybrid deep learning model based on ResNet-50 and BiLSTM to capture both spatial and temporal features from facial expression videos while incorporating strategies to mitigate the imbalance issue. To address the challenge of category imbalance inherent in real-world datasets, the study constructs an imbalanced version of the BioVid Part A database, reflecting realistic pain distribution with limited high-intensity samples. The study systematically compares three loss functions—Cross Entropy Loss, Focal Loss, and Cross Entropy Focal Loss (CEFL)—and evaluates their effectiveness in enhancing minority class recognition. A comprehensive benchmarking of four state-of-the-art optimization algorithms, AdamW, LAMB, AdaBelief, and NovoGrad, is conducted across a range of learning rates to systematically evaluate their convergence dynamics and generalization performance within the proposed framework. Experimental results show that the Focal Loss combined with LAMB or NovoGrad achieves superior performance, with the best accuracy reaching up to 84.89%, significantly outperforming traditional configurations. This research highlights the importance of tailored training strategies for imbalanced facial pain recognition and provides a robust baseline for future work. Future directions include expanding to real-world, in-the-wild pain assessment scenarios and integrating multimodal signals to enhance robustness and accuracy.