AI-assisted immunohistochemistry algorithm for enriching HER2-amplified cases among equivocal (2+) breast cancers: a two-cohort proof-of-concept study

Background HER2 assessment is a fundamental step in breast cancer diagnosis and treatment decision-making. Immunohistochemistry (IHC) remains the primary screening method; however, cases scored as equivocal (2+) require confirmatory in situ hybridization (ISH). This reflex testing prolongs turnaround time, increases costs, and consumes additional laboratory resources, while only a minority of 2 + cases ultimately show HER2 amplification. Methods Pathologist-assigned HER2 IHC 2 + invasive breast cancers were re-scored by the uPath HER2 (4B5) algorithm, blinded to ISH, in two cohorts: a retrospective enriched discovery cohort of 100 cases (50 ISH-amplified and 50 non-amplified) and a consecutive, routine-practice validation cohort of 47 cases reflecting real-world prevalence. The primary analysis used uPath HER2 3 + as the algorithm-positive threshold (uPath HER2 1+/2 + = non-positive); a secondary, exploratory analysis was restricted to non-equivocal outputs (HER2 2 + excluded). Diagnostic performance was reported with exact (Clopper–Pearson) 95% confidence intervals and complete confusion matrices, following the STARD and CLAIM reporting guidelines. Results uPath scores were significantly higher in ISH-amplified than non-amplified cases ( p < 0.0001). In the discovery cohort, uPath re-scored 45/100 cases as 3+ (35 amplified, 10 non-amplified); the primary analysis yielded a sensitivity of 70.0% and specificity of 80.0%. In the consecutive validation cohort, uPath re-scored only 4/47 cases (8.5%) as 3+ (2 amplified, 2 non-amplified), for a sensitivity of 25.0% and specificity of 94.9%. A secondary, exploratory analysis (2 + excluded) reached 100% sensitivity but only 9% specificity. Conclusions An off-the-shelf IHC algorithm enriched for HER2 amplification among equivocal cases but it was still not reliable enough to eliminate the need for confirmation by ISH. These proof-of-concept findings support combining pathologist assessment with purpose-trained AI and warrant prospective, multicenter validation before any change to clinical workflow.