A Pathomics-Based Prognostic Model for Disease-Free Survival in Resected Gastric Cancer

TL;DR

This study creates a model combining pathomics and clinical data to better predict survival after gastric cancer surgery.

Contribution

A novel clinic–pathomics model is developed for improved disease-free survival prediction in gastric cancer patients.

Findings

The clinic–pathomics model showed high predictive accuracy with AUCs up to 0.851 in the training cohort.

High-risk patients had significantly worse survival outcomes in both training and validation cohorts.

The nomogram outperformed clinic-only and pathomics-only models in decision curve analysis.

Abstract

Gastric cancer has a high postoperative recurrence rate, and traditional staging systems cannot accurately predict individual recurrence risk. Pathomics can extract quantitative features from pathological slides to reflect tumor biological characteristics, but there is a lack of reliable prognostic models combining pathomics and clinical data. This study aimed to develop and validate a disease-free survival prediction model for postoperative gastric cancer patients by integrating pathomics features and clinical factors. The final model showed better predictive performance, and a practical nomogram was built. This study provides a valuable tool for individualized risk stratification and postoperative management, and promotes the clinical translation of pathomics in gastric cancer.

Objectives: This study aims to develop and validate a prognostic risk model by integrating pathomics features with clinical variables to predict disease-free survival (DFS) in patients with gastric cancer (GC). Methods: Patients with GC who were pathologically diagnosed and subsequently treated with curative gastrectomy and D2 lymphadenectomy at the Fifth Affiliated Hospital of Wenzhou Medical University between January 2017 and April 2023 were retrospectively enrolled and assigned to a training cohort (n = 275) and an independent validation cohort (n = 118). Pathomics features were extracted from pathological images, and LASSO-Cox regression was used to identify pathomics features significantly associated with DFS. The selected pathomics features were integrated with clinical factors to create a prognostic model. Predictive accuracy was evaluated using time-dependent ROC analysis, and the model’s performance was compared with the clinic-only and pathomics-only models. A nomogram was constructed to provide individualized DFS predictions. Results: A total of 16 pathomics features were selected, and the cut-off for the pathomics scores was set at 0.27. High-risk patients exhibited significantly worse DFS compared to low-risk patients in both the training cohort (HR = 4.57, 95% CI: 3.118–6.697, p < 0.0001) and the validation cohort (HR = 2.264, 95% CI: 1.255–4.083, p < 0.0001). The clinic–pathomics model demonstrated strong predictive performance in both cohorts, with AUCs for 1-, 3-, and 5-year survival of 0.832, 0.821, and 0.851 in the training cohort, and 0.671, 0.702, and 0.682 in the validation cohort. The nomogram, incorporating the pathomics score, T stage, differentiation degree, and ECOG performance status, showed high calibration accuracy, as confirmed by calibration plots, and outperformed both the clinic-only and pathomics-only models in decision curve analysis. Conclusions: A clinic–pathomics model integrating pathomics features with clinical data provides a reliable tool for DFS prediction in patients with GC, which facilitates individualized DFS predictions and personalized treatment strategies.

Figures