Study on risk stratification and treatment strategy of blood indicators in patients with moderate risk of GIST.

Background: Intermediate-risk gastrointestinal stromal tumor (GIST) patients exhibit marked prognostic heterogeneity. The traditional NIH risk classification often results in undertreatment of latent high-risk patients and overtreatment of truly low-risk ones. This study aimed to develop an interpretable machine learning model integrating hematologic inflammatory markers to achieve precise risk re-stratification and optimize adjuvant therapy strategies for intermediate-risk patients. Methods: Primary GIST patients were retrospectively enrolled. LASSO regression was applied to select key features from eight inflammatory markers (including NLR, PLR, and SII). A random survival forest model was then constructed, followed by 5-fold cross-validation. SHAP values were used to interpret feature contributions, and Kaplan-Meier survival analysis was conducted to evaluate stratification performance. Results: LASSO regression identified seven inflammatory markers, among which PLR, SII, and PIV were the top three key variables. The optimal random survival forest model (five-feature model) achieved an AUC of 0.777, with an internally validated mean AUC of 0.782 (95% CI: 0.679-0.878) and an out-of-bag (OOB) error of 0.124. SHAP analysis revealed that PLR, NLR, and PAR were the major contributors to model prediction. The model effectively stratified intermediate-risk patients into "intermediate-high-risk" and "intermediate-low-risk" subgroups with significantly different survival outcomes (p<0.0001). Conclusion: This study represents the first construction of an interpretable predictive model integrating blood-based inflammatory markers with machine learning algorithms. The model accurately identifies occult high-risk individuals among patients with intermediate-risk GIST, thereby providing exploratory evidence and a foundation for hypothesis generation for future individualized management strategies.