Interpretable Pretreatment Prediction of Adaptive Replanning Needs In Head and Neck Cancer Radiotherapy
Abstract
Purpose
Head and neck cancer (HNC) patients undergoing radiotherapy often necessitate replanning due to anatomical changes. However, replanning is resource-intensive and typically decided on short notice, causing workflow disruptions and treatment delays. Therefore, we aimed to identify pretreatment clinical predictors of replanning and develop an interpretable prediction model.
Methods
We analyzed 547 HNC patients treated with radiotherapy between 2017 and 2024. We collected 22 pretreatment clinical variables suspected to influence replanning, including: sex, age, tumour site, TNM staging, pathology, grade, chemotherapy, immunotherapy, surgery, smoking history, p16 status, laterality, PEG insertion, BMI, PTV volumes, neck volume, and days between diagnosis and treatment. Data were split into stratified training (n=437) and testing (n=110) sets. Univariate analysis (chi-squared and Mann-Whitney U tests) was performed to identify potential predictors, followed by correlation and multivariate analysis to determine independently predictive variables. Based on these predictors, we developed an interpretable points-based scoring system where points are assigned according to replanning rates for each variable category. Patients were classified as requiring replanning if their score exceeded a threshold. For comparison, classical machine learning models were trained on the same data.
Results
The overall replanning rate was 47.6%. Univariate analysis identified 13 significant predictors (p<0.05). In multivariate analysis, the strongest independent predictors, subsequently used in our models, were N stage, M stage, chemotherapy, and PTV volumes. When evaluated on the hold-out test set, the points-based system achieved an AUC of 0.786, while the best machine learning model (logistic regression) achieved an AUC of 0.797, demonstrating comparable predictive performance despite the point model’s simplicity and interpretability.
Conclusion
We developed an interpretable risk-scoring system for HNC radiotherapy replanning that performs comparably to machine learning algorithms. This simple tool can support early identification of patients likely to require replanning, enabling proactive scheduling of rescans and reducing workflow disruptions.