Developing and validating a multi-omics prediction model for severe acute oral mucositis in nasopharyngeal carcinoma patients undergoing radiation therapy

Abstract

Background

Oral mucositis is a common and painful toxicity which can severely affect patients’ quality of life. This study focused on developing and externally validating a prediction model for severe acute oral mucositis (OM) in nasopharyngeal carcinoma (NPC) patients undergoing radiation therapy (RT). We attempted to harness pre-treatment clinical, dose-volume-histogram (DVH), radiomic and dosiomic features to better predict the occurrence of severe OM.

Methods

A retrospective analysis of 464 histologically confirmed NPC patients treated at two public institutions in Hong Kong was performed. Model development and internal validation was conducted on institution A (N=363) and external validation was evaluated on institution B (N=101). Severe OM was defined as the occurrence of CTCAE/RTOG grade 3 or higher during treatment. Two predictive models were constructed: 1) conventional clinical and DVH features, and 2) a multi-omics approach to include clinical, DVH, radiomic, and dosiomic features. Both models underwent rigorous optimization, involving mRMR feature selection, data scaling and model fitting conducted within 20-fold cross-validation on institution A.

Results

The multi-omics model outperformed the conventional model in internal and external validation. Specifically, the multi-omics model achieved area under the receiver-operating characteristic curve (AUC) scores of 0.67 (0.61, 0.73), and 0.65 (0.53, 0.77) respectively, compared to the conventional model's scores of 0.63 (0.56, 0.69) and 0.56 (0.44, 0.67). The 95% confidence intervals show that only the multi-omics model significantly outperformed random guessing (AUC=0.5) in external validation.

Conclusion

These findings suggest that radiomics and dosiomics features, by quantifying pre-treatment tissue radiodensity and spatial dose distribution, can enable better identification of patients at risk of severe OM. Further exploration of radiomics and dosiomics-based prediction models is warranted to facilitate improved clinical decision-making, thereby enabling more personalized care for the prevention and management of OM.