Multimodal Deep Learning Prognostic Model for Predicting Treatment Outcomes for Oropharyngeal Carcinoma.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiation Oncology, The Arthur G. James Comprehensive Cancer Center, The Ohio State University, Columbus, OH.
- Department of Radiation Oncology, Henry Ford Health, Detroit, MI.
Abstract
Accurate prediction of locoregional recurrence (LRR) in oropharyngeal squamous cell carcinoma (OPSCC) is needed to guide treatment de-escalation decisions, but validated prognostic models integrating imaging and clinical data are lacking. We developed and evaluated a multimodal deep learning model combining pretreatment computed tomography (CT) with clinical variables to predict LRR. In this retrospective multicenter prognostic study, we assembled 1,384 patients with histologically confirmed OPSCC treated with definitive radiotherapy with or without concurrent chemotherapy and without surgical resection of the primary tumor. Data were drawn from 1 private institutional archive and five publicly available archives. A cross-attention model fusing CT and clinical features was compared with feature concatenation, clinical-only, and imaging-only models. The primary outcome was LRR. Discrimination was assessed by Harrell Concordance Index (C-index) across 3-fold cross-validation, and risk stratification by Kaplan-Meier analysis with log-rank testing. Attribution analysis was used to identify influential predictors. LRR occurred in 161 of 1,384 patients (11.6%) over a median follow-up of 59.8 months (interquartile range, 25.8-89.0 months). The cross-attention model achieved the highest mean C-index, 0.718 (range, 0.699-0.734), followed by the clinical-only (0.705; range, 0.680-0.735), feature concatenation (0.697; range, 0.673-0.721), and imaging-only (0.674; range, 0.666-0.680) models. High-risk patients had significantly shorter LRR-free survival than low-risk patients in all three folds (log-rank <i>P</i> < .001 for each). Smoking history and human papillomavirus status were the most influential clinical predictors. A multimodal cross-attention deep learning model integrating CT with clinical variables showed numerically better discrimination for LRR prediction in OPSCC than unimodal approaches. Prospective external validation is required before such models inform individualized treatment decisions.