Conference Coverage

Synthetic-Face AI Shows Similar Pain Attention Patterns to Human-Trained Model

Edited by:
Brian McCurdy

Key Highlights

  • Researchers compared pain classifiers trained on synthetic facial images or human facial images, evaluating both on the Delaware Pain Dataset.
  • Facial attention maps showed high agreement, with an aggregate Pearson correlation of 0.991.
  • The synthetic-trained model emphasized regions around the eyes, while the human-trained model used more broadly distributed facial cues.
  • The human-trained model achieved modestly better class separation at the learned decision layer.

A comparative deep learning study found that a pain-detection model trained exclusively on synthetic facial images focused on anatomically meaningful regions similar to those used by a model trained on human images. The models showed highly correlated facial attention patterns, although the human-trained model separated pain and non-pain images more effectively at its learned decision layer.

The findings are scheduled for presentation on October 18 at the ANESTHESIOLOGY Annual Meeting 2026 in San Diego. Investigators examined whether synthetic training images could produce pain-related features like those learned from human data, addressing the difficulty of recording naturally occurring pain episodes.

The institutional review board–approved study compared two models: one trained on human faces collected at the San Francisco VA Medical Center and another trained on synthetic faces generated using Imagen3 from Alphabet Inc. Both models were evaluated on a separate dataset, the Delaware Pain Dataset.

Researchers used dimension reduction to visualize how each model internally organized pain and non-pain faces. They also used Grad-CAM++ attention maps to identify the facial regions the model emphasized during pain detection. A divergence map compared regional attention differences between the models.

Study Findings

The authors reported identical feature representations for the 2 models, with a silhouette score of 0.023 and an intra/inter ratio of 3.011. At the learned decision layer, however, the human-trained model achieved modestly better class separation: its silhouette score was 0.132 and intra/inter ratio was 1.043, compared with 0.032 and 2.730, respectively, for the synthetic-trained model.

Grad-CAM++ analysis showed high spatial agreement, with a per-image Pearson correlation of 0.916 ± 0.046 and an aggregate correlation of 0.991. Divergence mapping indicated that the synthetic-trained model preferentially focused on periorbital regions. The human-trained model distributed attention more broadly across the brows, nasolabial folds, and lower face, regions corresponding to established Facial Action Coding System pain action units.

Clinical Implications

The investigators argue these findings support using synthetic data for pain classification because the synthetic-trained model identified anatomically meaningful pain regions consistent with established facial pain cues. They concluded that the learned patterns mirrored those of the human-trained model.

The authors suggested that the synthetic model’s greater periorbital emphasis likely reflected exaggerated upper-face pain expressions in its AI-generated training data. In contrast, the human-trained model captured more distributed cues.

Expert Commentary

“These results support the use of synthetic data for pain classification and reveal learned pain-related regions that mirror human-trained models,” the researchers concluded.

Reference

Heintz T, Wallace A, Cobert J. Pain prediction from synthetic only facial images mirrors human-like pain regions. Abstract scheduled for presentation at: Anesthesiology 2026; October 16–20, 2026; San Diego, CA. Abstract A1318.