A machine learning model utilizing CT radiomics features and peripheral blood inflammatory markers predicts the prognosis of patients with unresectable esophageal squamous cell carcinoma undergoing PD-1 inhibitor combined with concurrent chemoradiotherapy.

Abstract

Objective: To investigate the value of a machine learning model that integrates radiomics features and peripheral blood inflammatory markers in predicting the prognosis of patients with unresectable esophageal squamous cell carcinoma (ESCC) receiving PD-1 inhibitor combined with concurrent chemoradiotherapy. Methods: A retrospective collection was conducted involving 105 patients with unresectable ESSC who received PD-1 inhibitors combined with concurrent chemoradiotherapy at the First Affiliated Hospital of the University of Science and Technology of China from January 2020 to August 2023. These patients were randomly divided into a training set (n=74) and a validation set (n=31). Radiomics features were extracted from arterial phase CT images obtained before initial treatment, with feature selection performed using Pearson Correlation and LASSO-COX methods. Baseline clinical characteristics were analyzed, and hematological parameters were collected before the start of immunotherapy and within 4-6 weeks post-treatment to calculate inflammatory markers. Subsequently, independent radiomics features influencing patient prognosis were identified using a multivariate Cox proportional hazards model, and these features were incorporated into a clinical feature-based multivariate Cox model to derive independent prognostic factors combining radiomics and clinical characteristics. Nomograms were constructed to predict the 2-year progression-free survival (PFS) of patients based on the results of COX analysis involving clinical characteristics, radiomic features, and combined indicators. The models were evaluated and assessed using ROC curves and calibration curves. Results: In the training cohort, the AUC was 0.705 for the clinical model, 0.573 for the radiomics model, and 0.834 for the combined model. In the validation cohort, the AUC was 0.784 for the clinical model, 0.775 for the radiomics model, and 0.872 for the combined model. Conclusion: The combined model integrating the radiomic feature NGTDM-busyness, the inflammatory marker ΔNLR, and the clinical characteristic M stage offers the optimal predictive value for the 2-year PFS in patients.