Deep Learning-Enabled Multimodal Fusion for Early Prediction and Risk Stratification of Head and Neck Squamous Cell Carcinoma: A Cross-Modal Attention Framework with Synthetic-Cohort Validation

Histopathological, radiological, molecular and clinical information is interpreted separately, partly because head and neck squamous cell carcinoma (HNSCC) is often diagnosed late. We study the benefit of learned cross-modal attention fusion of these sources, given the individual value of each modality. In the absence of controlled clinical repositories, we built a disclosed synthetic cohort from an explicit generative model (development n=2,847; distribution-shifted external cohort n=612) with structural missingness in radiomic and molecular data. Modality encoders are trained jointly for malignancy detection, three-tier risk stratification and risk-score regression and feed gated pooling to multi-head cross-modal attention. The comparators were leave-one-modality-out ablation, concatenation fusion, four unimodal models, logistic regression and gradient boosting. Fusion yielded internal AUC of 0.852 and external AUC of 0.863, against true generative risk ceilings of 0.880 and 0.905. It was not different from concatenation (external ΔAUC +0.004, p = 0.22) or logistic regression, and was better than histopathology alone by 0.016–0.023. Only removing histopathology degraded detection (external AUC = 0.698). Multimodal input improved risk stratification more (external tier accuracy 68.6% vs 61.8%). Attention fusion was robust but did not improve much over simpler fusion, validation on real patients is required.