Multi-omic prediction of severe acute treatment-induced oral mucositis and dysphagia in nasopharyngeal carcinoma patients

Abstract

Introduction: With the improvement in survival rates for nasopharyngeal carcinoma (NPC), it is increasingly important to address the impact of treatment-induced toxicity on patients' quality of life. Acute oral mucositis (OM) and dysphagia are two of the most common toxicities resulting from NPC treatment. Severe cases cause significant suffering and pose a threat to treatment outcome through unexpected hospitalization, weight loss, and treatment interruption. This thesis harnesses high-dimensional multi-omic data to identify patients at risk of severe acute OM and dysphagia to better target preventative interventions and personalized support. Methods: Four hundred and sixty-four NPC patients treated with radiotherapy (RT) at two Hong Kong hospitals were retrospectively recruited for analysis. Radiomic, dosiomic and contouromic features were extracted from planning CT images, RT dose distributions and tumour and organ-at-risk contours respectively. Machine learning models for predicting severe acute OM and dysphagia were developed. Model performance was comprehensively assessed and compared to that of conventional prediction models using clinical and dosimetric features alone. Results: Multi-omic prediction models for severe acute OM and dysphagia outperformed conventional clinical and dosimetric models developed on the same data. Radiomics, by describing pre-treatment tissue characteristics, dosiomics, by describing the spatial distribution of the planned RT dose, and contouromics, by describing the challenges posed by patient geometry, were demonstrated to have unique predictive value, and facilitated greater model discrimination by supplementing clinical features. Importantly, this study conducted external validation to assess the generalizability of the models, providing a greater level of evidence compared to other prediction models in the literature. Conclusion: Multi-omic features including radiomic, dosiomic and contouromic features enhanced the discrimination performance of models incorporating clinical and dosimetric features and demonstrated independent predictive value. The findings in this project provide an invaluable reference for future work and include important recommendations for future development of multi-omics for toxicity prediction.