Interpretable survival modeling integrating nutritional-inflammatory biomarkers in elderly patients with locally advanced esophageal squamous cell carcinoma treated with definitive radiotherapy.

Background: Elderly patients with locally advanced esophageal squamous cell carcinoma (ESCC) receiving definitive radiotherapy show marked survival heterogeneity. As anatomical staging alone cannot fully reflect host vulnerability, immune-inflammatory condition, and treatment tolerance, interpretable prognostic tools based on routinely available variables may help refine risk assessment in this population. Methods: This retrospective study enrolled 308 elderly patients with locally advanced ESCC treated with definitive radiotherapy. After excluding 12 patients with missing data, the remaining cases were randomly divided into training and testing cohorts at a 7:3 ratio for internal validation. Nine conventional and machine learning-based survival models were constructed using clinicopathological and nutritional-inflammatory variables, with performance evaluated via concordance index (C-index), time-dependent area under the curve (AUC), and integrated Brier score (IBS). SHAP analysis was performed on top-performing models to clarify predictor contributions over follow-up. Results: Median overall survival (OS) and progression-free survival (PFS) were 19 and 13 months, respectively. For OS, regularized linear survival modeling showed the most stable performance, with Ridge regression providing the most consistent discrimination in the testing cohort. For PFS, model ranking was more time- and endpoint-dependent, and no complex machine-learning algorithm demonstrated a stable overall advantage. Exploratory age-stratified analyses suggested heterogeneity across elderly subgroups, with different algorithms ranking highest in patients aged 65-74 years and those aged ≥75 years. SHAP analysis showed that nutritional indicators, especially GNRI and PNI, contributed more persistently to OS prediction than to short-term PFS, whereas baseline inflammatory markers such as NLR and SIRI showed relatively weak attribution in this multivariable prediction framework. Ridge-based predicted-risk thresholds identified distinct high- and low-risk groups for both OS and PFS. Conclusion: In elderly patients with locally advanced ESCC treated with definitive radiotherapy, regularized linear models showed robust prognostic utility, whereas more complex tree-based or boosting models did not demonstrate a stable advantage. Nutritional markers, particularly GNRI and PNI, provided more persistent prognostic information for OS than short-term PFS. Exploratory age-stratified analyses and Ridge-based threshold analyses further suggest that this interpretable framework may help refine follow-up and prioritize earlier nutritional assessment or supportive care in higher-risk elderly subgroups, while external validation remains necessary before clinical implementation.