Interaction-aware radiomics-dosiomics signatures improve the prediction of xerostomia after radiotherapy in nasopharyngeal carcinoma

Background Xerostomia remains a common toxicity after radiotherapy for nasopharyngeal carcinoma (NPC), emphasizing the need for accurate early prediction. This study aimed to develop an interpretable radiomics–dosiomics model for predicting clinically relevant xerostomia after radiotherapy in patients with NPC. Methods A total of 109 patients with radiotherapy planning data and xerostomia assessments approximately six months after completion of radiotherapy were included, of whom 105 with unambiguous grades were used for binary modeling of grade 2–3 versus grade 0–1 xerostomia. Pretreatment CT radiomics, dosiomics, segmental dose-distribution features and clinical variables were extracted from salivary-related structures. Stable radiomics and dosiomics features were selected using univariable analysis, Mann–Whitney testing, Spearman correlation filtering, mRMR analysis and LASSO stability selection. Interaction features were then constructed to model joint effects between image phenotype and regional dose exposure. Machine learning models were then evaluated using repeated stratified five-fold cross-validation. Results The interaction-aware L2 logistic regression model achieved the best overall balance between discrimination and calibration, with a pooled AUC of 0.897(95% CI: 0.836-0.949), AUC from patient-averaged predictions of 0.907(95% CI: 0.841-0.958), accuracy of 0.881(95% CI: 0.820–0.942), sensitivity of 0.870(95% CI: 0.783–0.957), specificity of 0.891(95% CI: 0.804-0.965) and Brier score of 0.124(95% CI: 0.089-0.163). Compared with clinical, radiomics-only, dosiomics-only and additive combined models, the interaction-aware model showed improved predictive performance. Interpretation analysis indicated that radiomics features and radiomics–segmental dose interaction features contributed substantially to model prediction. Conclusion An interaction-aware radiomics–dosiomics framework can improve prediction of clinically relevant xerostomia after radiotherapy for nasopharyngeal carcinoma. This approach provides an interpretable strategy for individualized toxicity risk assessment, demonstrating high discrimination accuracy and potential clinical utility.