Dynamic prediction of Radiotherapy toxicities in Head and neck cancer using clinical and imaging data

Head and neck cancer (HNC) radiotherapy (RT) is effective but causes significant toxicity. We aimed to develop a dynamic deep learning model to predict three major HNC RT toxicities-nasogastric (NG) tube placement, hospitalization, and radionecrosis-by integrating clinical data and daily cone-beam computed tomography (CBCTs), assessing whether serial imaging or dosimetry features improve early prediction.

We retrospectively analyzed 1,012 HNC patients treated with RT between 2017 and 2022. A multibranch 3D ResNet50 and multilayer perceptron model was trained using 5-fold cross-validation. Inputs included anatomical deformations from daily CBCTs (converted to Jacobian determinant matrices, J_f), radiomics, and clinical variables (demographics, tumor and treatment details, early weight loss). Each toxicity was modeled using weighted binary cross-entropy loss to address class imbalance. Prediction at the 10th RT fraction was compared with and without J_f integration.

The cohort was 78% male, median age was 63 years (range 35-84). Primary sites were mainly oropharynx (47%), larynx (19%), and oral cavity (16%). Concurrent chemoradiation was given to 57%, induction chemotherapy to 7%, and postoperative RT to 18% of patients. Incidences of NG tube, hospitalization, and radionecrosis were 16.6%, 4.2%, and 4.6%, respectively. Clinical features alone yielded highest predictive accuracy: 70% for NG tube, 67.3% for hospitalization, and 74.2% for radionecrosis. Early weight loss was the strongest predictor. Early J_f or radiomics did not improve performance. NG tube prediction accuracy improved with later RT fractions (up to 75% at fraction 25).

Clinical data combined with weight loss remains the most reliable early predictor of toxicity without added benefit from imaging data.