Artificial intelligence vs human-based frameworks for breast cancer molecular biomarker status assessment: A benchmarking analysis from 2500 patients.

Abstract

e12568 Background: Accurate assessment of breast cancer biomarkers (HR-ER/PR, HER2, Ki-67 index) is essential for prognostication and treatment selection but is limited by inter-observer variability and subjective interpretation. The emergence of HER2-low disease and quantitative thresholds has highlighted limitations in conventional pathology. Artificial Intelligence (AI) has potential to improve consistency and infer molecular phenotypes, however performance relative to human pathologists remains unclear. We conducted a meta-analysis to compare AI and human performance in biomarker assessment. Methods: 670 MEDLINE records evaluating AI-based diagnostic frameworks in breast cancer pathology were identified (PROSPERO ID: CRD420251161868). 9/52 studies directly comparing AI with human pathologists in biomarker assessment were included. Data extracted included study characteristics, biomarkers assessed, AI methodology, comparator performance, and diagnostic metrics, including accuracy, sensitivity, specificity, and area under the curve (AUC). Results: Analysis of nine studies encompassing 2,575 patients reveals that AI-driven biomarker assessment significantly enhances diagnostic precision in breast oncology. Evaluating Hormone Receptors, HER2, and Ki-67, AI frameworks achieved a mean diagnostic accuracy of 92.0% (range: 85.0%–97.8%), consistently outperforming human benchmarks, which yielded a mean accuracy of 82.2% (range: 60.0%–85.3%). Head-to-head comparisons underscored AI's capacity to mitigate inter-observer variability, particularly in HER2 scoring. In this domain, AI achieved 92.1% accuracy against consensus standards, exceeding manual reads (85.3%) and demonstrating superior inter-rater reliability (Cohen's κ = 0.84) compared to human evaluators (0.675). In Ki-67 quantification, AI demonstrated high expert concordance (Spearman's ρ = 0.745-0.861), providing crucial standardisation at clinical thresholds. The relationship between sample size (n = 12 to 1,341) and performance remained remarkably robust; mean accuracy varied by only ±3.5% across heterogeneous cohorts. While AUC reporting was limited to 0.71-0.72, the data confirm that AI frameworks reliably resolve "borderline" cases, such as the HER2 0 vs. 1+ distinction, effectively matching or exceeding traditional pathology standards. Conclusions: AI-based pathology frameworks demonstrate consistently high diagnostic performance for breast cancer molecular biomarker assessment compared with human pathologists. However, heterogeneity in reported metrics and limited use of molecular reference standards highlight the need for standardised validation and reporting frameworks to support clinical translation.