Enhancing prediction of primary site recurrence in head and neck cancer using radiomics and uncertainty estimation.

Introduction: Head and neck squamous cell carcinomas (HNSCC) present a significant clinical challenge due to high recurrence rates despite advances in radiation and chemotherapy. Early detection of recurrence is critical for optimizing treatment outcomes and improving patient survival. Methods: We developed two artificial intelligence (AI) pipelines-(1) machine learning models trained on radiomic and clinical data and (2) a Vision Transformer-based model directly applied to imaging data-to predict HNSCC recurrence using pre- and post-treatment PET/CT scans from a cohort of 249 patients. We incorporated Test-Time Augmentation (TTA) and Conformal Prediction to quantify prediction uncertainty and enhance model reliability. Results: The machine learning models achieved an average AUC of 0.820. The vision transformer model showed moderate performance (AUC = 0.658). Uncertainty quantification enabled the exclusion of ambiguous predictions, improving accuracy among more confident cases. Discussion: Our machine learning models achieved strong performance in predicting HNSCC recurrence from radiomic and clinical features. Incorporating uncertainty quantification further improved predictive performance and reliability.