Beyond expertise: exploring behavioral personas in AI-assisted rare renal cancer diagnosis

Fumarate hydratase-deficient renal cell carcinoma (FHdRCC) is a rare, aggressive malignancy that poses morphological screening challenges. We developed a deep learning morphological screening aid and conducted a crossover reader study with 21 pathologists evaluating 30 challenging cases with and without AI. To explore human-AI interaction, we operationalized Relative AI Reliance (RAIR) and Clinical Overreliance Rate (CORR). AI assistance significantly improved overall diagnostic accuracy (60.0–73.3%, P = 0.012) and inter-rater reliability (Fleiss’ kappa 0.311–0.482). Exploratory clustering based on RAIR and CORR identified two distinct behavioral phenotypes: Receptive ( n = 14) and Resistant ( n = 7). The Receptive group derived greater diagnostic benefit from correct AI advice (median RAIR, 0.732 vs. 0.444) but exhibited higher susceptibility to incorrect outputs (median CORR, 0.800 vs. 0.600). Conversely, the Resistant group maintained uniform oversight but limited potential gains. Notably, these behavioral personas were independent of traditional clinical expertise ( P = 1.0) and years of experience ( P = 0.881). These preliminary findings suggest that AI assists in the morphological triage of rare cancers. Furthermore, because human-AI interaction proficiency is a distinct trait decoupled from traditional diagnostic skill, persona-tailored training may optimize collaboration. This hypothesis-generating framework requires validation in larger studies.