Uncovering the associative mediation pathway from EUS features to malignancy through pathological mediators in gastrointestinal stromal tumor.

Background: Endoscopic ultrasound (EUS) plays a pivotal role in preoperative risk stratification of gastrointestinal stromal tumors (GISTs); however, the biological mechanisms underlying the association between EUS imaging features and malignant risk remain incompletely understood. Most existing studies focus on predictive performance, with limited exploration of interpretable pathways linking imaging phenotypes, pathological biomarkers, and risk classification. Methods: We proposed an integrated, network-driven analytical framework to investigate the mechanistic pathways underlying GIST risk stratification. Graphical Lasso modeling was first applied to construct a sparse feature interaction network incorporating EUS and pathological variables, enabling network-based feature selection through composite influence scores derived from topological and outcome-related metrics. Subsequently, counterfactual mediation analysis was performed to systematically evaluate whether pathological biomarkers mediated the associations between key EUS features and GIST risk category. Indirect effects were assessed using stratified bootstrap with 2000 resamples, and mediation proportions were quantified to characterize the relative contribution of mediating pathways. False discovery rate (FDR) correction was applied for multiple comparisons across all exposure-mediator pathways. Results: Graphical Lasso modeling identified maximum tumor size, shape regularity, and ulceration as highly influential EUS features within the interaction network. Mediation analysis revealed that Ki-67 was the only pathological biomarker exhibiting consistent and statistically significant mediation effects across multiple EUS features (FDR < 0.05). Specifically, Ki-67 significantly mediated the effects of maximum tumor size (13.1% of total effect), shape regularity (19.73%), and ulceration (21%) on GIST risk category. In contrast, other pathological markers-including CD34, CD117, and Dog-1-did not demonstrate significant indirect effects. These findings indicate that tumor proliferative activity, rather than immunophenotypic expression alone, serves as the primary biological conduit linking EUS phenotypes to malignant risk. Conclusion: This study provides a novel, interpretable framework that integrates network-based feature discovery with counterfactual mediation analysis to uncover biologically meaningful pathways in GIST risk stratification. The results highlight Ki-67 as a central mediator bridging EUS imaging features and malignant potential, offering mechanistic insights beyond conventional predictive modeling. This approach may facilitate more transparent imaging-based risk assessment and support precision decision-making in the management of GISTs.