Interpretable deep learning-based hierarchical multi-modal fusion model for predicting HER2 expression in gastric cancer.

Objectives: Aimed to develop and evaluate a hierarchical multimodal framework to improve HER2 status prediction in Gastric Cancer (GC) patients. Methods: This retrospective study included 402 endoscopically identified GC patients (2014-2023), among whom 92 had pathologically confirmed HER2 status. Endoscopic images were analyzed using ResNet-50 to extract deep learning (DL) features and to develop a separate model for predicting invasion depth to aid in HER2 prediction. Radiomic features were extracted from contrast-enhanced CT images using Pyradiomics, with optimal features selected based on reproducibility (ICC > 0.75) and feature importance. Six machine learning algorithms were used to build multimodal models, and performance metrics such as area under the curve (AUC) and sensitivity were evaluated. Additionally, SHapley Additive exPlanations (SHAP) values were used to assess the contributions of individual features within the best-performing model. Results: Using ResNet-50 analysis of endoscopic images, an invasion depth prediction model was developed (validation AUC: 0.79; 95% CI: 0.77-0.81). The probability output from this model was integrated as a variable and combined with three optimal DL endoscopic features (1, 2, 8) for HER2 prediction. Additionally, seven discriminative features were extracted via radiomic analysis. When integrated with endoscopic, radiomics and clinical data, the multimodal model consistently outperformed unimodal approaches, with Logistic Regression yielding the highest validation AUC (0.83, 95% CI: 0.74-0.91). SHAP analysis highlighted Wavelength-HHL_GLSZM_SAHGE and DLfeature_8 as key HER2 predictors. Conclusion: By integrating clinical, endoscopic, and radiomic data, the hierarchical multimodal approach enhances HER2 expression prediction in GC, thereby optimizing clinical decision-making, and enabling tailored HER2-targeted therapies.