GBMM: A graph-based multimodal Mamba model for response prediction in patients with esophageal squamous cell carcinoma receiving neoadjuvant immunochemotherapy.
Authors
Affiliations (5)
Affiliations (5)
- Department of Digital Medicine, College of Biomedical Engineering and Medical Imaging, Army Medical University (Third Military Medical University), Chongqing, China.
- Institute of Medical Robotics, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China.
- Department of Thoracic Surgery, The First Affiliated Hospital of Army Medical University (Third Military Medical University), Chongqing, China.
- Department of Digital Medicine, College of Biomedical Engineering and Medical Imaging, Army Medical University (Third Military Medical University), Chongqing, China. Electronic address: [email protected].
- Institute of Medical Robotics, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China. Electronic address: [email protected].
Abstract
Neoadjuvant immunochemotherapy (NICT) can improve outcomes in advanced esophageal squamous cell carcinoma, but its efficacy varies significantly among patients. Accurately predicting response to NICT before treatment can help identify those who may benefit and avoid ineffective treatment. However, this remains challenging in clinical practice. This study developed and validated a graph-based multimodal Mamba model (GBMM) for noninvasive prediction of treatment response using pretreatment CT imaging, radiomics features, and clinical variables. A total of 274 patients with esophageal squamous cell carcinoma who received NICT were enrolled retrospectively. The proposed GBMM integrated clinical features, radiomics features, deep image representations, and expert annotations through graph modeling, cross-modal interaction, Mamba-based representation learning, and teacher-student knowledge accrual. Model performance was evaluated in internal and external validation cohorts and compared with single-modality and conventional multimodal approaches. The model achieved area under the curve (AUC) values of 0.802 and 0.740 when evaluated on the internal and the external validation cohorts, respectively. These results were better than those of the state-of-the-art (SOTA) competing multimodal models. Decision curve analysis showed positive net benefit across a broad range of threshold probabilities. Shapley additive explanation (SHAP)-based interpretation highlighted tumor texture, morphology, systemic inflammatory status, physiological reserve, and regional disease burden as relevant predictive factors. These findings suggest that GBMM may provide a useful framework for pretreatment prediction of NICT response in esophageal squamous cell carcinoma. Further validation on larger multicenter prospective cohorts is required before its potential clinical utility can be established.