AI-Powered Real-Time Multimodal Model for Predicting Recurrence and Survival in Head and Neck Cancer: A Multicenter, Multinational Study

CLINICAL QUESTION

Can an artificial intelligence (AI)-powered longitudinal surveillance model accurately predict recurrence-free survival (RFS) and overall survival (OS) after curative-intent surgery for head and neck squamous cell carcinoma (HNSCC)?

BOTTOM LINE

An AI-powered multimodal model integrating clinicopathologic and longitudinal laboratory data accurately predicted recurrence and survival outcomes across multiple post-operative time points in patients with HNSCC. The model demonstrated robust performance across institutions and HPV subgroups, supporting its potential role in individualized surveillance and risk-adaptive follow-up strategies.

BACKGROUND: Recurrence after curative-intent treatment remains a major challenge in HNSCC, occurring in up to 50% of patients depending on disease risk. Current surveillance approaches rely largely on routine imaging and clinical assessment, with limited ability to dynamically individualize recurrence risk or tailor follow-up intensity over time.

STUDY DESIGN: Retrospective multicenter prognostic study using an eXtreme Gradient Boosting (XGBoost)-based AI model. Baseline clinicopathologic variables were integrated with serial laboratory measurements obtained during post-operative surveillance to generate longitudinal predictions for OS and RFS at one, two, three, four, and five-year intervals.

SETTING: Samsung Medical Center, Republic of Korea, and Massachusetts Eye and Ear Infirmary/Massachusetts General Hospital, U.S.

SYNOPSIS: Investigators analyzed 975 patients with HNSCC involving the oral cavity, oropharynx, hypopharynx, and larynx who underwent curative-intent surgery between 2008 and 2024. The AI-based “Recurrence And Death AI-based Risk” (RADAR) model incorporated 68 variables, including baseline demographic and pathologic features as well as longitudinal laboratory markers collected during surveillance visits.

The predictive model demonstrated strong performance for both RFS and OS across one to five years of follow-up. Recurrence-free survival prediction showed areas under the curve (AUCs) ranging from 0.769 to 0.831, while OS prediction ranged from 0.788 to 0.820, with sensitivities and specificities generally above 70%. Subgroup analysis showed particularly strong OS prediction in HPV-positive oropharyngeal cancer, with AUCs reaching as high as 0.943 at one year. In non-HPV-positive HNSCC, the model also maintained robust predictive accuracy, with OS AUCs ranging from 0.780–0.813 and RFS AUCs from 0.774–0.830 over five years.

Model interpretability identified ECOG performance status, tumor size, T and N classification, albumin, hemoglobin, neutrophil count, lymphocyte count, and C-reactive protein among the most influential predictive variables. The authors emphasize that the model uses routinely collected clinical and laboratory data and could potentially be integrated directly into electronic medical records to facilitate personalized surveillance strategies without additional testing burden. Limitations include the retrospective design, lack of radiomic or genomic integration, and absence of peri-operative immunotherapy-era data.

CITATION: Jung HA, et al. Artificial intelligence-powered real-time multimodal model for predicting recurrence and survival in head and neck cancer: a multicenter, multinational study. ESMO Open. 2026;11:106046. doi:10.1016/j.esmoop.2025.106046.