Development and external validation of a recursive attention-based multimodal deep learning model (RAMs-NPC) for prognostic stratification in nasopharyngeal carcinoma: A retrospective study.

Abstract

e18005 Background: Accurate prognostication is essential for personalized therapy in nasopharyngeal carcinoma (NPC), yet conventional risk stratification based on TNM staging and EBV DNA fails to capture tumor heterogeneity. We developed and validated a recursive attention mechanism (RAMs)-enhanced multimodal deep learning model (RAMs-NPC) that integrates MRI, digital pathology and clinical data to improve survival prediction for NPC. Methods: This single-center retrospective study enrolled 618 patients with non-metastatic NPC treated from 2017 to 2020, all with pretreatment MRI, H&E slides and ≥5 years of follow-up. Patients were randomly divided into training (n=433) and validation (n=185) cohorts at a 7:3 ratio. The RAMs-NPC model adopted U-Net for MRI tumor segmentation, ResNet for pathological feature extraction and MLP for clinical variable processing. RAMs dynamically focused on prognostically salient regions and weighted multimodal feature contributions adaptively; fused features via an attention gate were input into a Cox proportional hazards model for survival prediction. The primary endpoint was the concordance index (C-index) for overall survival (OS). Results: In the validation cohort, the model achieved an OS C-index of 0.84 (95% CI: 0.80-0.88), a progression-free survival (PFS) C-index of 0.81, a 3-year OS time-dependent AUC of 0.86 and a Dice similarity coefficient of 0.82 for MRI tumor segmentation. Ablation experiments demonstrated that RAMs significantly improved the model performance, with a 6.3% increase in OS C-index and a 6.6% increase in PFS C-index (all p<0.01). Kaplan-Meier analysis revealed significant separation of survival curves between high- and low-risk groups (log-rank p<0.001). Decision Curve Analysis (DCA) confirmed the model’s clinical utility, and the model-derived risk score was an independent prognostic factor for OS (HR=2.95, 95% CI: 1.88-4.63, p<0.001). Conclusions: The RAMs-NPC model synthesizes multimodal data via a recursive attention mechanism, providing a highly accurate and interpretable tool for NPC prognostic stratification. It outperforms conventional staging systems, shows promising potential to guide personalized treatment decisions, and warrants further external validation.