Prediction models of severe radiation-induced oral mucositis: a systematic review and meta-analysis

Radiation-induced oral mucositis (RIOM) is a common complication in cancer survivors following radiotherapy, and numerous studies have established predictive models to evaluate the risk of severe RIOM. However, there are significant differences in the methodological quality, predictive performance, and clinical application value of these models. This systematic review evaluated risk prediction models for severe RIOM in cancer survivors, analyzed the limitations of the current research, and proposed recommendations for optimization, thereby providing a reference for clinical nursing staff in selecting appropriate assessment tools. The search period spanned from the database's inception to February 2025 and included CNKI, VIP, Wanfang, Chinese Biomedical Literature Database, CINAHL, PubMed, Cochrane Library, and Embase. A meta-analysis was conducted to evaluate the incidence of severe RIOM and its predictive factors. The systematic evaluation was registered in the Platform for International Prospective Systematic Evaluation Registry (PROSPERO) database under registration CRD420250655070. A total of ten studies were included, involving 2,881 cancer survivors, encompassing 14 distinct models. Seven studies conducted internal validation, while only two studies performed external validation. Fourteen models reported the area under the Area Under Curve (AUC), which ranged from 0.657 to 0.942. The quality of the literature was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST), revealing that nine out of the ten studies evaluated were at a high risk of bias. The meta-analysis revealed a 36% incidence of severe RIOM (95%CI = 25%-48%). Age ≥ 60 years, diabetes, smoking, and a history of periodontal disease were identified as independent risk factors for severe RIOM (P < 0.05). The relevant predictive models exhibited excellent performance. However, most existing models lack external validation, limiting their extrapolation and clinical applicability. Moving forward, medical researchers should focus on developing models with exceptional predictive accuracy while minimizing the risk of bias, following standardized development guidelines.