Prediction model of gastrointestinal tumor malignancy based on coagulation indicators such as TEG and neural networks.

OBJECTIVES

Accurate determination of gastrointestinal tumor malignancy is a crucial focus of clinical research. Constructing coagulation index models using big data is feasible to achieve this goal. This study builds various prediction models through machine learning methods based on the different coagulation statuses under varying malignancy levels of gastrointestinal tumors. The aim is to use coagulation indicators to predict the malignancy of gastrointestinal tumors, expand the methods and ideas for coagulation index tumor prediction, and identify independent risk factors for gastrointestinal tumor malignancy.

METHODS

Clinical data of 300 patients with gastrointestinal diseases were collected from the Second Affiliated Hospital of Anhui Medical University from January 2024 to August 2024 and grouped according to TNM and G staging, representing tumor malignancy levels. First, independent influencing factors of gastrointestinal tumor malignancy were identified using stepwise multivariate logistic regression. ROC curves were used to assess the ability of TEG five items and other coagulation indicators to distinguish between malignancy levels of gastrointestinal tumors. Finally, we constructed a network model suitable for our task data based on residual networks, named the Residual Fully Connected Binary Classifier (RFCBC). This model was compared with other commonly used binary classification methods to select the optimal model.

RESULTS

The TEG five items (AUC values: R: 0.682; K: 0.731; α-angle: 0.736; MA: 0.699; CI: 0.747) showed better discrimination ability in the G group than other coagulation indicators. Although the TNM group showed moderate discrimination ability, it did not exhibit a significant advantage over other indicators. The R and MA values were identified as independent influencing factors in both TNM and G groups. Ultimately, the RFCBC prediction model showed the best predictive performance compared to other binary classification machine learning models (TEG five items: 87.56%; Thromboelastogram et al.: 88.6%).

CONCLUSION

This study found that the R and MA values are independent predictive factors for the malignancy of gastrointestinal tumors. Compared to other coagulation indicators, the TEG five items have better discrimination ability regarding tumor malignancy. The RFCBC model created in this study outperforms other commonly used binary classification methods in predicting the malignancy of gastrointestinal tumors, providing a new model construction method and feasible approach for future coagulation index prediction of gastrointestinal tumor malignancy.