Predictive models of lymph node metastasis in patients with gastrointestinal stromal tumors based on machine learning algorithms: a SEER-based retrospective study.

Abstract

Background: Gastrointestinal stromal tumor (GIST) is one of the most prevalent tumors in the digestive system. Due to the rarity of lymph node metastasis in patients with GISTs, there is a scarcity of related studies, which leads to ongoing debates regarding its impact on patient prognosis. This study aimed to analyze the impact of lymph node metastasis on the overall survival of GISTs patients and further building predictive models with machine learning algorithms.

Methods: This is a retrospective study based on the Surveillance, Epidemiology and End Results (SEER) database. The demographic and clinicopathological characteristics of GISTs patients who underwent surgical therapy were collected from database. Five different types of machine learning algorithms were used to build the models. The ensemble models were based on the three algorithms with the highest sensitivity in the validating cohort. At last, receiver operator characteristic (ROC) curve, precision-recall curve (PRC), calibration curve, decision curve analysis (DCA) and Kaplan-Meier survival curve (KMC) were used to evaluate our models.

Results: A total of 1,404 patients with GISTs were included in our study after data cleaning. Artificial Neural Network model [area under the ROC curve (AUC): 0.951, 95% confidence interval (CI): 0.901-0.992, sensitivity: 0.752] achieved the best performance in the validating cohort and was chosen to be the final model. Calibration plots showed good consistency between prediction and actual observations. Although the AUC between the final model and the baseline model showed no significant difference, the area under the PRC (PRAUC) of the final model (PRAUC =0.765) was significantly higher than that of the baseline model (PRAUC =0.455). DCA showed that the final model had high net benefit. Survival analysis indicated that the final model could distinguish the prognosis of patients significantly (all P<0.001).

Conclusions: We used machine learning algorithms to build models that can accurately predict lymph node metastasis in GISTs patients.