Predicting Replanning In Head-and-Neck Patients Using Dosiomics and Clinical Features
Abstract
Purpose
Adaptive replanning in head and neck (H&N) radiotherapy can be resource-intensive and can also disrupt clinical workflow. We developed and validated a machine learning model that combines dosiomics and clinical features to predict which patients are likely to require replanning and the associated risk factors.
Methods
A retrospective cohort of 149 H&N patients was analyzed. A total of 1,009 candidate predictors were extracted from dose distributions (DVH and spatial dose features) and clinical data. A multi-stage feature selection pipeline was applied (variance and correlation filtering, ANOVA and mutual information screening, L1-regularized logistic regression, permutation importance, and bootstrap stability selection). Iterative filtering to remove low-variance, redundant, and unstable features, and retaining a small set of consistently predictive variables, was necessary to avoid overfitting and improve generalization in high-dimensional, limited-sample datasets. Finally, a logistic regression classifier was trained and evaluated using stratified 5-fold cross-validation.
Results
53 of 149 patients (35.6%) required replanning. The pipeline reduced dimensionality by 99.5% and identified five robust predictors: extent of volume receiving dose ≥50 Gy, fraction of the parotid volume that receives at least 100% of the prescription dose, optic nerve D0.1cc, post-operative status, and chemotherapy. The model achieved an ROC AUC of 0.76, sensitivity of 81.1%, specificity of 64.6%, and accuracy of 70.5%. The model output categorized patients into groups: 27 (18.1%) were very high risk (probability >70%), and 47 (31.5%) were high risk (50-70%). In each category, the final five predictors that determine the decision and the potential for early replan prediction are identified.
Conclusion
A model using five clinically interpretable dosiomics and clinical predictors demonstrated moderate discrimination for replanning need in H&N patients and may enable early identification of high-risk patients and the factors involved, for proactive replanning. A larger cohort and refinement of the dosiomics predictors can further improve the model.