Development and validation of a prediction model for distinguishing upper gastrointestinal stromal tumor and leiomyoma based on white-light endoscopy and ultrasound endoscopy
10.3760/cma.j.cn311367-20231016-00117
- VernacularTitle:基于白光内镜和超声内镜的上消化道间质瘤与平滑肌瘤鉴别预测模型的建立和验证
- Author:
Jianying LU
1
;
Yijie GU
;
Wenjuan SHEN
;
Tingting XIA
Author Information
1. 苏州大学附属第一医院内镜中心，苏州 215006
- Keywords:
Ultrasound endoscopy;
Submucosal tumor;
Gastrointestinal stromal tumor;
Leiomyoma;
Nomogram
- From:
Chinese Journal of Digestion
2024;44(5):314-320
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To analyze the image characteristics of gastrointestinal stromal tumor(GIST) and leiomyoma under white-light endoscopy and ultrasound endoscopy, so as to establish a nomogram model and to validate its performance.Methods:From August 1, 2019, to December 1, 2022, the clinical data of 224 patients with GIST or leiomyoma who underwent endoscopic ultrasound examination at the First Affiliated Hospital of Soochow University were retrospectively analyzed. The 224 patients were divided into the modeling group of 145 cases (78 cases of GIST and 67 cases of leiomyoma), and the validation group of 79 cases (41 cases of GIST and 38 cases of leiomyoma). The basic data of patients, parameters of white-light endoscopy and ultrasound endoscopy were screened to establish a binary logistic regression model and draw a nomogram. The receiver operating characteristic curve (ROC) was drawn, and the area under the curve (AUC) was used to evaluate the diagnostic efficiency of the model, and calibration curve was used to evaluate the consistency of predicted and observed probabilities. The model′s performance was compared with the diagnostic results of junior physicians (attending physicians) and senior physicians (associated chief physician). Decision curve analysis (DCA) was performed to evaluate the net benefit of the model. Independent sample t-test and chi-square test were used for statistical analysis. Results:Under white-light endoscopy, there were statistically significant differences in the lesion locations (esophagus: 0 vs. 56.7% (38/67); cardia: 11.5% (9/78) vs. 13.4% (9/67); gastric: 88.5% (69/78) vs. 29.9% (20/67)) and tumor morphyology (spherical or spheroid: 80.8% (63/78) vs. 28.4% (19/67); shuttle: 19.2% (15/78) vs. 71.6% (48/67)) between GIST and leiomyoma in the modeling group ( χ2=64.51 and 46.37, both P<0.001). Under ultrasound endoscopy, the proportion of patients with GIST whose lesions originated from the muscularis propria layer, with indistinct borders and with internal hyperechoic area were all higher than those of patients with leiomyoma (96.2% (75/78) vs. 62.7% (42/67); 53.8% (42/78) vs. 13.4% (9/67); 35.9% (28/78) vs. 10.4% (7/67)), and the differences were statistically significant ( χ2=25.91, 25.82 and 12.75, all P<0.001). Based on the logistic regression model, a nomogram model was established with age, tumor morphology, lesion origin, boundary clarity, and hyperechoic foci as predictive indicators. In the modeling group, the accuracy of nomogram model in the diagnosis of GIST and leiomyoma was 89.7% and 83.6%, respectively. In the validation group, the sensitivity, specificity, and accuracy in GIST and leiomyoma diagnosis of the nomogram model and senior physicians were all higher than those of junior physicians in differentiating GIST from leiomyoma (90.2%, 87.8% vs. 82.9%; 81.6%, 84.2% vs. 78.9%; 86.1%, 86.1% vs. 81.0%, respectively), and the sensitivity, specificity, and accuracy of the nomogram model were equal to those of senior physicians in differentiating GIST from leiomyoma. The AUCs of the nomogram model in the modeling group and validation group were 0.932 (95% confidence interval 0.891 to 0.974) and 0.916 (95% confidence interval 0.854 to 0.978), respectively. The calibration curves of the model indicated that the consistency between the predicted probabilities and observed probabilities was good, and DCA suggested good clinical net benefits of the model. Conclusion:The model exhibits good test efficiency, discrimination, prediction consistency and clinical net benefit when age, tumor morphology, lesion origin, boundary clarity, and hyperechoic foci are selected as indicators.
