KEY POINTS
- This retrospective single-institution study evaluated 60 patients with head and neck cancer treated with helical tomotherapy to 50–70 Gy in 25–35 fractions between 2024 and 2025. The primary endpoint was patient-reported severe xerostomia at three months, defined as a University of Michigan Xerostomia Questionnaire score ≥60/100. Severe symptoms occurred in 16 patients (26.7%), leaving only a small number of positive events for predictive modeling.
- Six feature frameworks were compared using identical machine-learning pipelines. These included conventional parotid DVH parameters, first-order dosiomics, comprehensive dosiomics, combined DVH and dosiomics, and two additional frameworks incorporating clinical variables. The investigators extracted 125 dosiomic features per parotid gland, including dose-distribution texture and shape descriptors, and evaluated Random Forest and support vector machine classifiers using stratified five-fold cross-validation and three class-balancing strategies.
- The best-performing model achieved only modest discrimination. A Random Forest combining DVH and dosiomic features achieved a mean cross-validated AUC of 0.603 ± 0.186 (95% CI 0.466–0.763). By comparison, logistic regression using mean parotid dose alone achieved AUC 0.454 ± 0.211, while a clinical-only model achieved AUC 0.491 ± 0.160. The best model's confidence interval included chance-level discrimination.
- Adding spatial dose features did not produce a statistically significant performance improvement. Comparisons of Random Forest models showed no significant difference between DVH-only and combined DVH–dosiomics features (p=0.548) or between full dosiomics and combined features (p=0.421). Across all 36 modeling configurations, mean AUC values ranged from approximately 0.32 to 0.60. Adding clinical covariates to the combined model did not consistently improve prediction.
- Class-balancing methods failed to resolve poor predictive performance. Synthetic minority oversampling (SMOTE) increased the mean AUC of one full-dosiomics Random Forest model from 0.511 to 0.552, but did not improve the overall best-performing combined model. Feature selection was also highly unstable: the most frequently selected feature appeared in only 6.67% of 7,500 selection events, and no feature reached the prespecified 50% stability threshold. These results argue against assigning biological importance to individual selected descriptors.
- Calibration and decision-curve results suggested possible utility despite weak discrimination. For the best combined model, mean Brier score was 0.199 ± 0.025 and expected calibration error was 0.136 ± 0.071. Decision-curve analysis estimated a net benefit of 4.63 per 100 patients at a 30% risk threshold, with favorable results over approximately 20–40% thresholds. However, these estimates were internally evaluated in the same small cohort and do not establish clinical usefulness.
- The model has not been externally validated and does not address long-term xerostomia. Only parotid dose distributions were incorporated, without comprehensive consideration of submandibular glands, minor salivary glands, or additional biological predictors. The outcome was assessed at three months, whereas late salivary dysfunction may evolve over 12–24 months. With only 16 severe events, substantial fold-to-fold variation, and no independent validation cohort, the reported performance may overestimate generalizability.
CLINICAL TAKEAWAY
Three-dimensional dosiomic features did not significantly improve prediction of early severe xerostomia compared with conventional DVH parameters in this small head and neck radiotherapy cohort. Although spatial dose information may contain additional predictive signals, the observed discrimination was too weak and unstable to support clinical NTCP-based decision-making. Larger multicenter studies with external validation and long-term toxicity endpoints are required.