Applications of interpretable machine learning models in the prediction of 5-year risk of death in patients with gastric gastrointestinal stromal tumors

Abstract

Aim: To evaluate the performances of seven different machine learning models in predicting 5-year risk of death in patients with gastrointestinal stromal tumors(GIST) of the stomach. Meanwhile, the Shapley Additive explanation (SHAP) value is applied to construct a visual machine learning model. Methods: The data of GIST of the stomach patients derived from Surveillance, Epidemiology, and End Results (SEER) database from 1975 to 2020. Patients were divided into training (n = 748) and validation sets (n = 321). The data were used to construct seven machine learning models to predict 5-year risk of death in patients with GIST of the stomach. A total of 7 clinical variables were input into these models. Model performance was measured with the area under the receiver operating characteristic curve (AUC) and average precision (AP). The models were visualized and interpreted using the SHAP method. Results: The 5-year survival rate was 39.9% (426/1068). Of the seven machine learning models, Catboost had the best AUC (0.64) and AP (0.73). We explored the significance of features in the model through SHAP analysis. Surgery, patient age and risk of GIST were the heavily weighted factors used by the Catboost. Conclusion: This is the largest study of GIST of the stomach patients from the SEER registry to show that surgery, patient age and risk of GIST are significant independent prognostic factors for 5-year risk of death. Based on simple baseline patient information, Catboost model can accurately predict the 5-year risk of death. Also, SHAP values can be good for interpreting machine learning models, as well as for predicting, guiding follow-up and monitoring individuals.