6P - Diagnostic accuracy of artificial intelligence in classifying HER2 status in breast cancer immunohistochemistry slides: A systematic review and meta-analysis

Background

HER2-positive breast cancer accounts for approximately 15-20% of cases and often results in poor clinical outcomes. The DESTINY-Breast04 trial demonstrated that trastuzumab-deruxtecan (T-DXd) significantly improved survival in patients with immunohistochemistry (IHC) scores of 1+, and 2+ with negative in situ hybridisation (HER2-low). Consequently, accurate differentiation between scores is crucial. Manual HER2 IHC classification, however, is labour-intensive and prone to significant interobserver variability. This study evaluates the performance of artificial intelligence (AI) in distinguishing HER2 IHC scores of 0, 1+, 2+, and 3+.

Methods

We conducted searches in MEDLINE, EMBASE, Scopus, and Web of Science up to 3rd May 2024. We included original studies evaluating the performance of AI compared to pathologists' manual scoring as the reference standard in classifying HER2 IHC. Meta-analysis was performed employing the bivariate random-effects and hierarchical summary receiver operating characteristics models to estimate pooled sensitivity and specificity, and the area under the curve (AUC). Heterogeneity was assessed using the I2 Higgins’ score. Statistical analysis was carried out using Stata v17.0. The risk of bias was evaluated using QUADAS-2 tool.

Results

We evaluated the performance of 9 AI algorithms across 8 publications, covering 872 breast cancer cases (patients and images). AI demonstrated excellent overall accuracy in HER2 scoring, as evidenced by an AUC = 0.98 [95% CI 0.96-0.99]. Meta-analysis revealed a pooled sensitivity of 0.88 [95% CI 0.81-0.93] and a pooled specificity of 0.96 [95% CI 0.94-0.98]. Heterogeneity was significantly high for both metrics, with an I2 of 83.05 [95% CI 78.14-87.96] for sensitivity and 83.13 [95% CI 78.25-88.01] for specificity.

Conclusions

AI shows significant potential in aiding pathologists with accurate classification of HER2 status in breast cancer IHC samples. Variability in results may be attributed to the diverse technological tools employed in each AI algorithm. This high level of accuracy underlines AI's capability to standardise HER2 diagnosis and reduce interobserver variability in clinical practice.

Editorial acknowledgement

Clinical trial identification

Legal entity responsible for the study

D. Arruda Navarro Albuquerque.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.