Abstract
<title>Abstract</title> <p>Background Patients with locally advanced esophageal squamous cell carcinoma (ESCC) may experience substantially different recurrence trajectories after neoadjuvant immunochemotherapy (nICT) and curative-intent surgery. Whether pretreatment PET/CT radiomics can add prognostic information to postoperative clinicopathologic assessment remains uncertain. This study developed and externally evaluated a PET/CT radiomics-clinicopathologic model for recurrence-free survival (RFS) risk stratification. Methods This multicenter retrospective study included 256 patients with ESCC who underwent pretreatment PET/CT before nICT and radical esophagectomy. A development cohort of 201 patients was used for feature selection and model development, and 55 patients from two external centers were pooled for validation. RFS was the primary endpoint. Radiomic features were extracted from manually segmented tumors and screened by segmentation reproducibility, variance filtering, univariable Cox analysis, redundancy control, and LASSO-Cox regression. Selected features were summarized as an RFS-Radscore. The primary combined Cox model integrated RFS-Radscore with objective clinicopathologic variables and excluded postoperative adjuvant therapy to reduce treatment-selection ambiguity. Model performance was assessed using Harrell's C-index, bootstrap internal validation, calibration, Kaplan-Meier risk stratification, and time-dependent ROC analysis. Results RFS events occurred in 93 of 201 patients in the development cohort and 28 of 55 patients in pooled external validation. The radiomics workflow reduced 1,702 PET/CT features to 11 LASSO-Cox-selected predictors. In the primary combined model, higher RFS-Radscore (HR, 5.62; 95% CI, 3.25–9.70; P < 0.001), PD-L1 expression (HR, 0.61; 95% CI, 0.40–0.95; P = 0.030), and neural invasion (HR, 1.90; 95% CI, 1.17–3.09; P = 0.010) were associated with RFS. The model achieved C-index values of 0.735 (95% CI, 0.684–0.785) in the development cohort and 0.740 (95% CI, 0.656–0.824) in pooled external validation; the bootstrap optimism-corrected C-index and calibration slope were 0.724 and 0.939, respectively. Model-defined high-risk patients had poorer RFS in both the development cohort (HR, 3.81; 95% CI, 2.40–6.04; P < 0.001) and pooled external validation cohort (HR, 4.03; 95% CI, 1.21–13.39; P = 0.023). However, external time-dependent AUCs for the combined model did not exceed those of the clinicopathologic model alone. Conclusions The PET/CT radiomics-informed clinicopathologic model showed exploratory potential for postoperative RFS risk stratification after nICT in ESCC. Radiomics provided complementary pretreatment information, but its incremental external discrimination over clinicopathologic variables was not consistent. The model should therefore be considered hypothesis-generating and requires prospective validation with harmonized imaging protocols before clinical implementation.</p>