Answers are grounded in this paper and cite their sources.
BACKGROUND: HER2 immunohistochemistry (IHC) plays a central role in therapeutic decision-making for breast cancer. However, interpretation of equivocal (2+) IHC results remains challenging and is subject to interobserver variability, necessitating reflex in situ hybridization testing. This study evaluated the diagnostic performance and reproducibility of ChatGPT-4o, a general-purpose large language model, in scoring HER2 IHC in breast cancer cases initially classified as IHC 2+. METHODS: We retrospectively analyzed 81 formalin-fixed, paraffin-embedded invasive carcinoma of no special type (NST) cases with prior HER2 IHC 2+ scores and corresponding dual in situ hybridization (DISH) results. Five high-power field images per case were independently analyzed by ChatGPT-4o across three sessions, using a standardized prompt aligned with the ASCO/CAP 2023 guidelines. Cases remaining equivocal after AI-assisted interpretation were excluded from diagnostic performance calculations. HER2 DISH served as the reference standard. RESULTS: Fourteen cases (17.3%) remained equivocal following AI interpretation. Among the 67 reclassified cases, ChatGPT-4o demonstrated an overall diagnostic accuracy of 79% (95% CI: 67-88%), a sensitivity of 30%, specificity of 100%, positive predictive value of 100%, and negative predictive value of 77%. Intra-model reproducibility was good (intraclass correlation coefficient = 0.78), whereas agreement with HER2 DISH was fair (Cohen's κ = 0.375). Misclassification predominantly involved false-negative interpretations among HER2-positive cases. CONCLUSION: ChatGPT-4o demonstrated high specificity and reproducibility for identifying HER2 IHC 3+ cases but showed limited sensitivity and only fair concordance with HER2 DISH. These findings indicate that, in its current general-purpose form, ChatGPT-4o is not suitable for independent HER2 assessment and may serve, at best, as an exploratory adjunct to pathologist interpretation.
No takes yet. Share an insight, caveat, or question.
Charoenlap et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: