Zero-shot multimodal pain estimation via synthetic pain simulation and domain-invariant learning

Pain assessment in non-communicative populations-particularly neonates and cognitively impaired patients-remains a critical clinical challenge, as current automated methods require labeled pain datasets that are both ethically problematic and scarce for vulnerable populations.

We propose a framework trained on zero labeled real pain examples from the target population, combining synthetic pain simulation with unsupervised domain adaptation. Using latent diffusion models, we generate 50,000 realistic pain scenarios spanning facial expressions, physiological signals (heart rate variability and electrodermal activity), and temporal dynamics across diverse demographics. A transformer-based architecture integrates these multimodal cues, and a three-stage domain alignment procedure (contrastive learning, adversarial adaptation, and consistency regularization) bridges the synthetic-to-real gap using only unlabeled real data. We explicitly distinguish this synthetic-source unsupervised domain adaptation setting from classical zero-shot learning (Section 2.4).

Evaluation across three benchmarks shows competitive performance: MAE of 0.89 on the UNBC-McMaster dataset (31% gap vs. supervised methods), 78.3% accuracy on the BioVid Heat Pain Database (10.5% gap), and a preliminary point estimate of 81.5% accuracy for neonatal assessment (12.7 percentage points above a clinical AU-based baseline, with wide uncertainty given the limited N = 120 cohort; prospective validation on a larger cohort is planned). The proposed approach achieves 40.4% better cross-dataset generalization than supervised transfer learning and shows consistent performance across demographics (no significant disparities, p > 0.05; several subgroup comparisons are statistically underpowered). New analyses in this revision include Monte Carlo Dropout uncertainty quantification, robustness evaluation under acquisition perturbations, and a computational-complexity assessment.

This study provides preliminary evidence that ethically developed pain assessment AI-trained without exploiting vulnerable populations-can achieve clinically useful performance, and proposes a methodological pathway toward this goal that warrants confirmation through prospective, adequately powered clinical validation.