PurposeTo determine whether the addition of metabolic parameters from fluorine-18-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) scans to clinical factors could improve risk prediction models for radiotherapy-related esophageal fistula (EF) in esophageal squamous cell carcinoma (ESCC).Methods and MaterialsAnonymized data from 185 ESCC patients (20 radiotherapy-related EF-positive cases) were collected, including pre-therapy PET/CT scans and EF status. In total, 29 clinical features and 15 metabolic parameters from PET/CT were included in the analysis, and a least absolute shrinkage and selection operator logistic regression model was used to construct a risk score (RS) system. The predictive capabilities of the models were compared using receiver operating characteristic (ROC) curves.ResultsIn univariate analysis, metabolic tumor volume (MTV)_40% was a risk factor for radiotherapy (RT)-related EF, with an odds ratio (OR) of 1.036 [95% confidence interval (CI): 1.009–1.063, p = 0.007]. However, it was excluded from the predictive model using multivariate logistic regression. Predictive models were built based on the clinical features in the training cohort. The model included diabetes, tumor length and thickness, adjuvant chemotherapy, eosinophil count, and monocyte-to-lymphocyte ratio. The RS was defined as follows: 0.2832 − (7.1369 × diabetes) + (1.4304 × tumor length) + (2.1409 × tumor thickness) – [8.3967 × adjuvant chemotherapy (ACT)] − (28.7671 × eosinophils) + (8.2213 × MLR). The cutoff of RS was set at −1.415, with an area under the curve (AUC) of 0.977 (95% CI: 0.9536–1), a specificity of 0.929, and a sensitivity of 1. Analysis in the testing cohort showed a lower AUC of 0.795 (95% CI: 0.577–1), a specificity of 0.925, and a sensitivity of 0.714. Delong’s test for two correlated ROC curves showed no significant difference between the training and testing sets (p = 0.109).ConclusionsMTV_40% was a risk factor for RT-related EF in univariate analysis and was screened out using multivariate logistic regression. A model with clinical features can predict RT-related EF.