Body composition dynamics and their biological correlates in gastric cancer (GC) prognosis remain insufficiently understood. This study integrated longitudinal clinical and CT-derived body composition data with interpretable machine learning to explore the prognostic relevance of body composition after radical gastrectomy. Body composition features were quantified at the L3 level using automated CT segmentation in 353 institutional patients, with longitudinal changes derived from baseline and approximately 3-month follow-up scans. These imaging-derived features, together with clinicopathological variables, were used to develop and internally evaluate machine learning models for postoperative recurrence and 3-year survival prediction. A public TCGA/TCIA cohort of 38 patients was used only for exploratory imaging-transcriptomic analysis rather than external model validation. The selected models showed exploratory predictive performance, and body composition-related variables, particularly VAT and longitudinal body composition changes, showed potential prognostic relevance. Exploratory transcriptomic analyses generated preliminary hypotheses linking imaging-defined risk patterns with lipid metabolism- and coagulation-related pathways. These findings remain investigational and require prospective multicenter external validation, threshold refinement, and mechanistic confirmation before clinical implementation.
Hanzheng et al. (2026) studied this question.