Table_1_Prediction Models for Sepsis-Associated Thrombocytopenia Risk in Intensive Care Units Based on a Machine Learning Algorithm.DOCX

Sepsis-associated thrombocytopenia (SAT) is a common complication in the intensive care unit (ICU), which significantly increases the mortality rate and leads to poor prognosis of diseases. Machine learning (ML) is widely used in disease prediction in critically ill patients. Here, we aimed to establish prediction models for platelet decrease and severe platelet decrease in ICU patients with sepsis based on four common ML algorithms and identify the best prediction model. The research subjects were 1,455 ICU sepsis patients admitted to Dongyang People's Hospital affiliated with Wenzhou Medical University from January 1, 2015, to October 31, 2019. Basic clinical demographic information, biochemical indicators, and clinical outcomes were recorded. The prediction models were based on four ML algorithms: random forest, neural network, gradient boosting machine, and Bayesian algorithms. Thrombocytopenia was found to occur in 732 patients (49.7%). The mechanical ventilation time and length of ICU stay were longer, and the mortality rate was higher for the thrombocytopenia group than for the non-thrombocytopenia group. The models were validated on an online international database (Medical Information Mart for Intensive Care III). The areas under the receiver operating characteristic curves (AUCs) of the four models for the prediction of thrombocytopenia were between 0.54 and 0.72. The AUCs of the models for the prediction of severe thrombocytopenia were between 0.70 and 0.77. The neural network and gradient boosting machine models effectively predicted the occurrence of SAT, and the Bayesian models had the best performance in predicting severe thrombocytopenia. Therefore, these models can be used to identify such high-risk patients at an early stage and guide individualized clinical treatment, to improve the prognosis of diseases.

败血症相关性血小板减少症（SAT）是重症监护病房（ICU）中的一种常见并发症，它显著提高了死亡率，并导致疾病预后不良。机器学习（ML）在危重患者的疾病预测中得到了广泛应用。本研究旨在基于四种常见的机器学习算法建立ICU败血症患者血小板减少及严重血小板减少的预测模型，并确定最佳预测模型。研究对象为2015年1月1日至2019年10月31日期间入住温州医科大学附属东阳人民医院的1,455例ICU败血症患者。记录了基本临床人口统计学信息、生化指标和临床结局。预测模型基于四种机器学习算法：随机森林、神经网络、梯度提升机和贝叶斯算法。结果显示，732名患者（占49.7%）出现血小板减少。与未出现血小板减少的患者组相比，血小板减少组的机械通气时间和ICU住院时间更长，死亡率更高。模型在国际在线数据库（重症监护医疗信息市场III）上进行了验证。四种模型在预测血小板减少的受试者工作特征曲线下面积（AUC）介于0.54至0.72之间。模型预测严重血小板减少的AUC介于0.70至0.77之间。神经网络和梯度提升机模型有效地预测了败血症相关性血小板减少症的发生，贝叶斯模型在预测严重血小板减少症方面表现最佳。因此，这些模型可用于早期识别高风险患者，并指导个体化临床治疗，以改善疾病的预后。