2519 - HER2-Protein Based Radiomics Nomogram for Predicting Pathological Complete Response in HER2-Positive Breast Cancer Receiving Anti-HER2 Therapy

Purpose/Objective(s):

Early prediction of pathologic complete response (pCR) is vital for optimizing treatment strategies in Human Epidermal Growth Factor Receptor 2 (HER2) -positive breast cancer. While studies indicate tumors with higher HER2 expression achieve higher pCR, standard immunohistochemistry (IHC)-based classification fails to explain the variability in individual treatment responses. To address this, we developed a HER2 protein-based radiomics nomogram integrating Reverse Phase Protein Array (RPPA)-derived expression, tumor heterogeneity assessment in non-invasive imaging, and clinical insights to optimize pCR prediction for anti-HER2 therapy.

Materials/Methods:

Publicly accessible RPPA-derived HER2 protein expression, T1w precontrast and postcontrast MRI images of 222 HER2+ patients in I-SPY 2 Trial were used. The inclusion criteria were HER2+ patients, identified by an IHC 3+ or IHC 2+/FISH+, and hence receiving anti-HER2 drugs. Patients were randomly assigned to training and testing sets (7:3) using pCR- stratified splitting. In model construction, radiomics features were extracted within functional tumor VOI by PyRadiomics. Features with robust intraclass correlation (ICC >0.9), significant correlation with RPPA-HER2 expression (spearman p <0.05) , minimum inter-feature correlation and maximum relevance were used to construct the radiomics model using SVM classifier. PCR-correlated clinical data (X² p <0.05) were integrated to create radiomics nomogram. As a proven predictor of pCR rates, we used the HER2 expression derived from RPPA as a benchmark for comparison. All models were built in the training set with 3-fold cross validation and verified by the independent testing set.

Results:

Twenty HER2-correlated radiomics features (of 1,316) and two clinical variables (of 8) were used. Radiomics-based models outperformed expression-based models. The HER2 Radiomics and its nomogram achieved training AUCs of 0.777 and 0.844, and testing AUCs of 0.666 and 0.655, respectively. In contrast, the HER2 Expression and its nomogram yielded training AUCs of 0.694 and 0.766, with testing AUCs of 0.658 and 0.642. Although clinical integration improved training performance, nomograms yielded lower testing AUCs than standalone models. The clinical model alone showed the lowest testing AUC (0.543). Reduced testing performance in nomograms suggests clinical integration decreases model generalizability.

Conclusion:

We demonstrated the RPPA-HER2 radiomics model is a viable predictor for pCR. The clinical-radiomic integration was limited by a narrow range of clinical parameters. Expanding the diversity of clinical parameters in larger cohorts is essential to confirm its utility.

Model		AUC
		Training set	Testing set
1	HER2 Expression (benchmark)	0.694	0.658
2	HER2 Radiomics	0.777	0.666
3	Clinical data	0.702	0.543
1+3	HER2 Expression Nomogram	0.766	0.642
2+3	HER2 Radiomics Nomogram	0.844	0.655