Abstract / Summary
Background: Response to radiotherapy combined with immune checkpoint inhibition in esophageal cancer is markedly heterogeneous, and available indicators identify likely responders poorly. This study developed and internally validated a model integrating pretreatment CT radiomics with clinicopathological features to predict early RECIST 1.1-defined radiographic response. Methods: In total, 156 patients with esophageal cancer (151, 96.8%, squamous cell carcinoma) treated with radiotherapy plus an immune checkpoint inhibitor between January 2021 and December 2024 were retrospectively enrolled and split 7:3 into training (n = 109) and internal validation (n = 47) cohorts. From whole-tumor regions of interest on pretreatment venous-phase contrast-enhanced CT, 1218 radiomics features were extracted and a radiomics score (Rad-score) was derived in the training cohort. Clinical, radiomics and integrated models predicting RECIST 1.1 response 8-12 weeks after radiotherapy were assessed for discrimination, calibration and net benefit. Results: Objective response and disease control rates were 47.4% and 86.5%. Eight features were retained; the Rad-score was higher in responders (0.35 ± 0.81 vs −0.38 ± 0.82, P < 0.001). The Rad-score (OR = 3.111), neutrophil-to-lymphocyte ratio <3.0 (OR = 2.877), tumor length <5 cm (OR = 2.312) and completion of at least four immunotherapy cycles (OR = 2.180, an on-treatment variable) were independent predictors. The integrated model achieved AUCs of 0.800 (95% CI 0.716-0.884) and 0.776 (95% CI 0.640-0.912) in the training and validation cohorts, numerically higher than the clinical and radiomics models (DeLong test, all P > 0.05). Brier scores were 0.182 and 0.190, and net benefit was positive between threshold probabilities of 0.10 and 0.80. Objective response rates in higher-versus lower-predicted-probability groups were 76.9% vs 21.1% (training) and 66.7% vs 30.8% (validation). Conclusion: Developed predominantly in esophageal squamous cell carcinoma, the integrated model predicts early radiographic response rather than long-term outcome. As one predictor becomes known only after treatment initiation, it applies to the early-treatment setting and remains hypothesis-generating until externally validated.