Predicting radiation-induced xerostomia in head and neck cancer: a critical review of radiomics, machine learning and methodological quality

OBJECTIVE: To evaluate the predictive performance, methodological quality and clinical readiness of radiomics and machine learning approaches for predicting radiation-induced xerostomia. METHODS: A structured narrative review was conducted. PubMed and Embase were searched for studies evaluating radiomics, machine learning and deep learning approaches for predicting xerostomia following head and neck radiotherapy. Methodological quality was assessed using the Radiomics Quality Score. RESULTS: Twenty-two eligible studies, with cohorts ranging from 19 to 1208 patients, were identified. Delta-radiomics and deep learning approaches generally demonstrated the strongest predictive performance, although gains over conventional models were typically modest. Mean Radiomics Quality Score was 14 out of 36, with common limitations including limited external validation, absence of prospective studies, lack of open science and no cost-effectiveness analyses. CONCLUSION: Radiomics and machine learning show potential as adjuncts to existing xerostomia prediction models; however, methodological limitations and poor generalisability remain barriers to routine clinical implementation.