Development and validation of NTCP models for predicting acute esophagitis in carbon ion radiotherapy for thoracic malignancies

Acute esophagitis (AE) is a common toxicity following carbon ion radiotherapy (CIRT) for thoracic malignancies (TM). Despite CIRT’s physical advantages, specific normal tissue complication probability (NTCP) models for predicting AE remain underdeveloped. This study aims to develop NTCP models using clinical data to identify key risk factors for AE during CIRT, enabling personalized treatment optimization to reduce esophageal toxicity. 168 patients with TM (non-small cell lung cancer (NSCLC): 95, tracheobronchial adenoid cystic carcinoma (TACC): 43, thoracic bone and soft tissue sarcomas (TBSTS): 30) treated with CIRT (59.5–80 Gy (RBE) in 16–23 fractions) were analyzed. AE was graded per Common Terminology Criteria for Adverse Events v4.03. Dosimetric parameters and clinical variables were evaluated using logistic regression. The Lyman-Kutcher-Burman (LKB) NTCP model was optimized via maximum likelihood estimation to establish volume effect (n), slope (m), and tolerance dose for 50% complication probability (TD50) parameters. Model performance was assessed by the area under curve (AUC) of the receiver operating characteristic (ROC) curve and validated internally using 1000-fold bootstrapping. No grade ≥ 3 AE was observed. TACC showed highest rates (G1:88.4%; G2:20.9%), followed by TBSTS (G1:50.0%; G2:16.7%) and NSCLC (G1:35.8%; G2:1.1%). Females had higher G1 (63.0% vs. 47.5%) but doubled G2 risk (13.0% vs. 7.4%) versus males. Toxicity peaked in patients < 60 years (G1:64.4%; G2:16.4%), declining with age. Paradoxically, non-drinkers had 9-fold higher G2 risk (14.0% vs. 1.5%), while smoking showed protection (non-smokers vs. smokers, G2:1.2% vs. 17.1%). Notably, multivariate analysis identified V44Gy (RBE) as the independent predictor for G1 AE (OR: 1.06, 95% CI: 1.013–1.115; p < 0.001) with AUC 0.83. While G2 toxicity was driven by D9cc (1.07, 1.02–1.13; p = 0.004), age years (0.94, 0.895–0.986; p = 0.012), and smoking status (0.12, 0.013–1.04; p = 0.05) with AUC 0.936.The LKB models demonstrated for G1 AE: n = 0.23 (95% CI:0.05–0.45), m = 0.81 (0.2-1.0), TD50 = 62.8 Gy (RBE) (35.5–75.5) with AUC = 0.792, and for Grade 2 AE: n = 0.06 (0.01–0.20), m = 0.16 (0.01–0.25), TD50 = 80 Gy (RBE) (50.0–95.0) with AUC = 0.764. The first validated NTCP models for CIRT-induced AE highlight V44Gy (RBE), D9cc, age, and smoking as critical predictors. Derived LKB parameters enable personalized planning to reduce esophageal toxicity, advancing particle therapy optimization.