Katrina A. Wolters¹, Veith Veilnhammer², Jefferson Ortega¹, Zhihang Ren¹, Haley Frey¹, David Whitney¹; ¹University of California, Berkeley, ²Max Planck UCL Centre for Computational Psychiatry and Ageing Research

Clinician accuracy in dermatological malignancy judgments is essential for optimal patient outcomes. Previous work has identified serial dependence in skin lesion judgments (Ren et al., 2023), and recent findings suggest that human perception alternates between two distinct modes: an internal mode, characterized by decreased attentional engagement and attractive perceptual biases (e.g. serial dependence), and an external mode, where task engagement is heightened and perception is aligned more closely with ground truth (Weilnhammer et al., 2023). Here, we investigated if sequential biases were present in a 2AFC lesion malignancy classification task, and whether machine learning tools like Hidden Markov Models (HMMs) could identify distinct perceptual states. We analyzed over 750,000 lesion judgments made by medical students and residents on an online app, and found significant serial dependence in malignancy judgments (p < 0.01), replicating Ren et al. (2023). We extended these results by fitting an HMM to the behavioral data, finding that a two-state model was strongly favored over a one-state model (log Bayes Factor = 6681.36). The two inferred states were consistent with internal and external perceptual modes: the internal mode was associated with significant serial dependence, whereas the external mode showed a repulsive sequential bias, and exhibited 14% higher classification accuracy than the internal mode. Further, given that serial dependence is feature-similarity tuned (Manassi et al., 2023), we explored the impact of sequential lesion similarity (computed using Google’s Derm Foundation machine learning model) on malignancy judgments, and found that sequential lesion similarity significantly predicts perceptual bias, with net serial dependence increasing as image similarity increases. Taken together, these results suggest that there may not be one singular perceptual bias or decision-making state in effect when observers make sequential malignancy judgments; instead, sequential lesion similarity and fluctuations in perceptual mode both play a substantial role in bias and diagnostic accuracy.

Acknowledgements: Supported in part by NIH R01CA236793.