Artificial intelligence-powered real-time model for predicting recurrence and survival in head and neck squamous cell carcinoma after curative intent surgery.

Abstract

6083 Background: Head and neck squamous cell carcinoma (HNSCC) accounts for approximately 5.3% of cancer-related mortality worldwide, with an estimated 660,000 new diagnoses and 325,000 deaths annually. Curative-intent surgery or definitive chemoradiotherapy remain the only two curative treatment modalities for patients with HNSCC but recurrence rates vary from 10-50% and survival is still limited for some patients emphasizing the need for accurate predictive and prognostic models. This study developed and validated an AI model that integrates clinical, pathological, laboratory, and radiologic data to predict recurrence and survival in HNSCC, aiming to optimize patient outcomes and personalize treatment strategies. Methods: The model was developed using XGBoost and Cox regression, internally validated, and tested using data from Samsung Medical Center (SMC). Data in the model included baseline data collected at the time of surgery and longitudinal laboratory data gathered during surveillance. An 80/20 ratio was applied to randomly allocate patients to the developing set and internal validation sets from the SMC dataset. The dataset included patients with HNSCC who underwent curative intent surgery between January 2008 and August 2024. Two models were developed: one to predict progression-free survival (PFS) and overall survival (OS) within 12 months after the surgery, and another to predict PFS and OS within 12 months of the surveillance monitoring point, thus creating a real-time prediction model. External validation was conducted using data from Massachusetts General Hospital (MGH). Results: A total of 1,062 patients with HNSCC (oral cavity cancer, oropharyngeal cancer, and laryngeal cancer) were included in the study. The AUC for predicting 12-month PFS after surgery was 0.804 (sensitivity: 82.4%, specificity: 77.3%), with a C-index of 0.802 for RFS. For predicting OS at 12 months after surgery, the AUC was 0.875 (sensitivity: 100%, specificity: 73.1%), with a C-index of 0.862 for RFS. For external validation using MGH data, the AUC for predicting 12-month PFS was 0.875, with a C-index of 0.793 for RFS. The C-index for OS in the MGH dataset was 0.75. In the longitudinal surveillance model, the AUC for predicting 12-month PFS at each monitoring point was 0.883, while the AUC for 12-month OS was 0.902. Conclusions: This study successfully developed and validated an AI-powered model for predicting RFS and OS in HNSCC patients, achieving strong performance in both internal and external validations. These findings highlight the potential of AI-based approaches to support personalized treatment strategies and improve prognostic accuracy in HNSCC.

Ähnliche Arbeiten