Preliminary establishment of a sample clot warning model for coagulation screening tests based on machine learning algorithm
10.3760/cma.j.cn114452-20250213-00081
- VernacularTitle:基于机器学习算法的血栓与止血筛查试验样本凝固预警模型初步建立
- Author:
Weiling SHOU
1
;
Qian CHEN
;
Zhejun FANG
;
Chengxiang CUI
;
Lin ZHENG
;
Siyu MA
;
Wei WU
Author Information
1. 中国医学科学院北京协和医院检验科，北京 100730
- Publication Type:Journal Article
- Keywords:
Thrombus;
Coagulation screening tests;
Clot;
Prediction model;
Machine learning;
XGBoost
- From:
Chinese Journal of Laboratory Medicine
2025;48(5):603-608
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To preliminarily establish a sample clot warning model for coagulation screening tests using 5 machine learning methods.Methods:This cross-sectional study collected 7 401 routine screening test samples from Peking Union Medical College Hospital, Chinese Academy of Medical Sciences, from January 1st, 2015, to August 18th, 2024, including 4 786 clotted (positive) and 2 615 qualified (negative) samples for model development. The dataset was divided into Dataset 1 and Dataset 2 based on a reagent change for APTT in December 2018, with separate models developed for each. An additional 2 493 samples (October 31st to November 8th, 2024) were used to evaluate consistency between the model and manual assessment, while 23 200 samples (October 17th to December 31st, 2024) were used for assessing real-world predictive performance. Five machine learning algorithms were employed to develop the clot prediction model: logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), naive bayes (NB), and artificial neural network (ANN), with the ANN model constructed using two different hidden layer and neuron parameter settings. Model selection was based on AUC, accuracy, sensitivity, specificity, F1-score, PPV, and NPV, with the optimal model integrated into the LIS for validation.Results:Among the six models using 5 machine learning algorithms, XGBoost demonstrated the highest performance (AUC=0.961, sensitivity=0.945, F1-score=0.934) and robustness to reagent changes ( Z=-1.333, P=0.113). When deployed, the differences between the model's predictions and manual pre-judgment were statistically significant ( Z=-5.289 to 8.933, all P<0.01). The predictive efficacy indices AUC (95% CI), sensitivity, specificity, and accuracy of the XGBoost model deployed in real-world operation of the LIS were 0.939 (0.918—0.960), 0.958, 0.921, and 0.921 respectively. Conclusion:In this study, a clot warning model for coagulation screening samples was established based on the XGBoost algorithm, and its prediction efficacy is good, providing a foundation for intelligent pre-analytical quality control for coagulation screening tests.
