Abstract
OBJECTIVES: This study aimed to build a radiomics signature using machine learning methods to estimate overall survival in patients with locally advanced esophageal squamous cell carcinoma (ESCC) who underwent definitive chemoradiotherapy (dCRT), and to verify its prognostic value across independent patient cohorts. METHODS: We retrospectively included 200 ESCC patients with histological confirmation from three medical centers. Radiomics models were constructed employing machine learning algorithms. A predictive nomogram combining radiomics-derived risk metrics with clinical features was established. Model performance was assessed by the concordance index (C-index), time-dependent ROC curves, and decision curve analysis (DCA). Similar modeling approaches were also applied to an independent immunotherapy-treated cohort. RESULTS: The developed radiomics signature exhibited modest predictive ability for overall survival in advanced ESCC patients treated with dCRT. High-risk individuals experienced reduced survival in the training cohort (p = 0.028) and validation cohort (p = 0.021) datasets, with similar findings observed in two external validation cohorts. The integrated nomogram combining clinical and radiomic features outperformed other predictive models and demonstrated potential clinical value for survival prediction. Within the immunotherapy-treated subgroup, the radiomics signature remained a statistically significant predictor of survival (p = 0.002), and the combined nomogram consistently exhibited acceptable prognostic performance. CONCLUSIONS: A reliable radiomics signature was established to effectively estimate survival outcomes in patients with advanced ESCC undergoing chemoradiotherapy or immunotherapy. Combining this model with clinical data enhanced its predictive capacity, underscoring its value for personalized prognostic evaluation. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12967-025-07387-1.