Automated Facial Pain Assessment Using Dual-Attention CNN with Clinically Calibrated High-Reliability and...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Pain Medicine

Background:

• Accurate pain assessment is challenging, particularly for non-verbal patients, as current methods rely on subjective self-reports or clinician observations.
• Existing pain assessment tools lack objectivity and consistency, hindering effective clinical management and research.

Purpose of the Study:

• To develop and validate a novel automated facial pain assessment framework using a dual-attention convolutional neural network (CNN).
• To achieve clinically calibrated, high-reliability, and interpretable pain level detection from facial expressions.
• To provide an objective and consistent tool for pain monitoring in clinical settings.

Main Methods:

• Developed a dual-attention CNN architecture incorporating multi-head spatial attention and an enhanced channel attention block with triple-pooling.
• Employed label smoothing (α = 0.1) and AdamW optimization for stable model convergence.
• Evaluated the framework on a clinically annotated dataset using subject-wise stratified sampling across five pain classes.

Main Results:

• Achieved a test accuracy of 90.19% ± 0.94% and a 5-fold cross-validation accuracy of 83.60% ± 1.55%.
• Obtained an F1-score of 0.90, Cohen's κ = 0.876, and AUCs of 0.991 (macro) and 0.992 (micro).
• Grad-CAM visualizations confirmed the model's focus on physiologically relevant facial regions.

Conclusions:

• The proposed framework offers a robust, explainable, and reproducible method for automated facial pain assessment.
• The AI-driven system demonstrates significant potential for integration into real-world automated pain-monitoring systems.
• The biomimetic approach aligns with natural pain perception mechanisms, enhancing objective pain evaluation.