Improving resectable gastric cancer prognosis prediction: A machine learning analysis combining clinical features and body composition radiomics

Abstract

We evaluate the significance of body composition radiomics in predicting outcomes for resectable gastric cancer (GC) patients, as these parameters can enhance optimal surveillance strategies and risk-stratification models. Automated segmentation using deep learning algorithms was employed on CT images to analyze body composition in 276 GC patients, retrospectively recruited from the Clinical Hospital of the University of Campinas. Radiomics features of skeletal muscle (SM), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) were calculated. Body composition radiomics were integrated with clinicopathological factors using machine learning (ML) algorithms trained for patient outcome prediction. We compared results using Random Forest, Logistic Regression and Boosted Decision Tree algorithms. To identify the relevant features for the prognosis, recursive feature inclusion (RFI) was performed using SHAP Importance ranking. Our study uncovered novel body composition radiomic features that enhance patient prognosis, particularly the 90th percentile radiodensity value (HU) for SM and VAT. The ML model output also refined pathological staging: Stage II patients with a higher predicted mortality risk by the model had overall survival (OS) similar to Stage III patients, while Stage III patients with lower predicted risk showed OS comparable to Stage II. This approach demonstrates that the integration of clinical and radiomic features enhances the accuracy of pathological staging and offers more detailed information to refine treatment strategies for gastric cancer patients. Skeletal muscle and visceral adipose tissue radiodensity percentiles emerged as crucial determinants of patient outcomes.