Prediction of HER2 changes post-neoadjuvant therapy based on fusion of ultrasound radiomics and clinicopathological features empowered by explainable AI: A multicenter study.

[BACKGROUND] Dynamic HER2 expression changes during neoadjuvant therapy (NAT) challenge precision oncology. Current biopsy-dependent evaluation inadequately meets clinical needs for dynamic monitoring. We developed a non-invasive predictive model integrating pretreatment ultrasound radiomics and clinicopathological parameters to forecast post-NAT HER2 status evolution in breast cancer.

[METHODS] In this multicenter retrospective study (January 2017-May 2023), 655 patients with paired pre- and post-NAT HER2 assessments were enrolled from three institutions. Pretreatment ultrasound images underwent manual tumor segmentation and radiomic feature extraction. Clinicopathological parameters, including baseline HER2 status and NAT regimen, were retrospectively collected. An Ultrasound Image Clinical Feature Fusion (UICFF) framework incorporating attention-guided multimodal feature selection was developed for HER2 transition prediction. Model performance was compared with six classical and two unimodal baselines. Survival outcomes were evaluated using Kaplan-Meier analysis.

[RESULTS] Dynamic HER2 alterations occurred in 29.7 %, 34.6 %, and 25.5 % of the training, internal, and external cohorts, respectively. The multimodal UICFF achieved superior discrimination (AUC_internal = 0.811; AUC_external = 0.823), outperforming radiomics-only and clinicopathological-only models (external ΔAUCs: +0.193 and +0.125, respectively). SHapley Additive exPlanations analysis identified age, menopausal status, Ki-67 index, and wavelet-based texture features as dominant predictors. Younger age and larger tumor size were positively associated with HER2 dynamics. Dynamic HER2 changes correlated with improved pathological response to HER2 blockade (58.6 % vs. 38.1 %) but showed no independent effect on 5-year event-free survival.

[CONCLUSIONS] The interpretable UICFF framework enables individualized, pretreatment prediction of HER2 evolution in patients undergoing NAT, providing a clinically actionable and noninvasive alternative to repeated biopsies.