Interobserver Reproducibility of Treatment-Defining HER2-Low and HER2-Ultralow Immunohistochemical Scoring in Breast Carcinoma: A Systematic Review

Eligibility for trastuzumab deruxtecan now depends on separating HER2 immunohistochemistry (IHC) score 0 from 1+ (HER2-low) and, within score 0, null from ultralow staining—boundaries the assay was not designed to resolve. We systematically identified and appraised studies reporting interobserver or interlaboratory agreement among pathologists for HER2 IHC scoring of breast carcinoma at these contrasts. PubMed/MEDLINE, Scopus, and Web of Science were searched (17 to 18 June 2026) without date or language limits, following a prospectively registered protocol and the PRISMA 2020 guidelines. Two reviewers independently screened records, extracted data verbatim, and appraised quality using QAREL; no values were estimated or imputed. Eighty-nine studies were included: 34 at first submission, and 55 added after an audit of the screening record and retrieval of previously unavailable reports, both completed during revision and disclosed in full. Agreement at the low-end contrasts was consistently lower than for the full HER2 category set: coefficients operationalized directly at 0 versus 1+ ranged from 0.41 to 0.69 across kappa and intraclass-correlation families, per-category coefficients for score 1+ ranged from 0.19 to 0.65, with most below 0.55, and overall multicategory agreement was generally higher (most coefficients 0.43 to 0.85). Agreement for null versus ultralow was examined in 10 of the 89 studies across 4 constructs and was low and spectrum-dependent; in the largest direct examination, per-category κ for ultralow was 0.09 to 0.21. The evidence was fragmented across at least 9 agreement metrics and across binary, per-category, and within-multicategory operationalizations of the same contrast; many studies reported no confidence interval; and no metric family and operationalization was shared by 2 or more studies with a derivable standard error on comparable cohorts. Sixty-two of 89 studies raised some concerns, and 27 were at high risk of bias. Analytical reproducibility at these treatment-defining boundaries is therefore not yet established; the fragmentation that prevents a pooled estimate is itself a finding, underscoring the need for harmonized reporting and analytical validation as priorities.