Sangwon Shin

Human epidermal growth factor receptor 2 (HER2)-targeted therapies have revolutionized breast cancer treatment, necessitating standardized HER2 testing. However, current immunochemistry-based HER2 assessment faces challenges due to subjectivity and variability among observers, prompting the guidance of artificial intelligence (AI). We evaluated AI’s efficacy in HER2 status assessment and treatment response prediction, especially focusing on complete and intense circumferential HER2-positive (3+) tumor cells. An AI-powered HER2 analyzer (Lunit SCOPE HER2, Lunit Inc., Seoul, South Korea) and three pathologists independently assessed HER2 3 + tumor cell proportions from whole-slide images of 191 breast cancer cases from Kyung Hee University Hospital. Logistic regression and receiver operating characteristic (ROC) curves determined predictive accuracy. AI-detected 3 + tumor cell proportions strongly correlated with fluorescent in situ hybridization (FISH) positivity (area under ROC curve [AUC]: 0.783) and pathological complete response (pCR) rates (odds ratio [OR]: 1.003–1.052), outperforming pathologists’ assessments. Combining AI improved prediction accuracy of pathologists, with an AUC increased from 0.712–0.813 to 0.790–0.821 for predicting FISH positivity and median OR increased from 0.996–1.061 to 1.003–1.055 for predicting pCR. Overall, this preliminary study suggests that AI could enhance HER2 status determination and treatment response prediction, complementing traditional pathological evaluation of HER2 immunohistochemistry.

Baseline characteristics of the test sets (test set A, B, and C).

Figure S6. Prediction scores for EGFR mutation prediction between two magnifications in test set C (n = 2,261).

Performance of the ensemble model to predict EGFR mutation status across test set A (n = 1,461) and test set B (n = 599). ROC curves for test set A (A–D) and test set B (E–I) are shown. For each test set, the following subgroup analyses were performed: entire dataset (A and E), specimen type (surgical resection and biopsy; B and F), EGFR mutation subtypes (E19del, L858R point mutations, E20ins, T790M, and other mutations; C and G) and scanner type (D and H). For test set B, data source subgroups are additionally presented (I). The sample size (“n”) for each subgroup reflects the number of cases and controls. For EGFR subtypes, each ROC curve was generated using cases positive for the specific EGFR subtype (with or without co-mutations) as EGFR mutant and all EGFR wild-type cases as EGFR wild type. The data represent the AUC together with corresponding 95% CIs.

Baseline characteristics of the training set and tuning set.

Figure S3. Prediction scores of the ensemble model for EGFR mutation prediction across EGFR mutation subtypes in test set A.

Figure S5. Prediction scores for EGFR mutation prediction across scanner types in test set C (n = 2,261).

Generalized performance across different whole-slide scanners and magnifications in test set C (n = 2,261). A and B, Binary classification performance for EGFR mutation prediction across six scanner types (A) and two magnifications (B). The x-axis represents the slides sorted by average value of prediction scores from six scanners. The predicted EGFR classification determined by the model is denoted by color (blue, EGFR mutant; black, EGFR wild type). C and D, Concordance rates of binary classification performance across scanner types (C) and magnifications (D). E, Pearson correlation coefficients (r) for the continuous prediction scores across scanner types and magnifications.

Figure S1. Distribution of prediction scores from the ensemble model in Test Set A (n = 1,461) and Test Set B (n = 599).

Supplementary Materials and Methods