Fei Yang, Hu Jian, Gao Kun, Tu Jianfeng, Wang Wei, Li Wei-Qin
Surgical Intensive Care Unit, Department of General Surgery, Jinling Hospital, Medical School of Nanjing University, Nanjing, China.
School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing, China.
Ann Vasc Surg. 2018 Feb;47:78-84. doi: 10.1016/j.avsg.2017.09.004. Epub 2017 Sep 22.
Acute pancreatitis (AP) can induce portosplenomesenteric vein thrombosis (PVT), which may generate higher morbidity and mortality. However current diagnostic modalities for PVT are still controversial. In recent decades, artificial neural networks have been increasingly applied in medical research. The aim of this study is to predict the risk of AP-induced PVT by radial basis function (RBF) artificial neural networks (ANNs) model.
A retrospective or consecutive study of 426 individuals with AP at our unit between January 1, 2011 and July 31, 2016 was conducted. All individuals were subjected to RBF ANNs. Variables included age, gender, red blood cell specific volume (Hct), prothrombin time (PT), fasting blood glucose, D-Dimer, concentration of serum calcium ([Ca]), triglyceride, serum amylase (AMY), acute physiology and chronic health evaluation II score, and Ranson score. All outcomes were derived after subjecting the variables to a statistical analysis.
In the RBF ANNs model, D-dimer, AMY, Hct, and PT were the important factors among all 11 independent variables for PVT. The normalized importance of them was 100%, 96.3%, 71.9%, and 68.2%, respectively. The predict sensitivity, specificity, and accuracy by RBF ANNs model for PVT were 76.2%, 92.0%, and 88.1%, respectively. There were significant differences between the RBF ANNs and logistic regression models in these parameters (95% CI: 110.9% [-0.4 to 15.8%]; 8.4% [-3.3 to 19.2%]; and 12.8% [1.6-20.7%], respectively). In addition, the area under receiver operating characteristic curves value for identifying thrombosis when using the RBF ANNs model was 0.892 ± 0.091 (95% CI: 0.805-0.951), demonstrating better overall performance than the logistic regression model (0.762 ± 0.073; 95% CI: 0.662-0.839).
The RBF ANNs model was a valuable tool in predicting the risk of PVT following AP. AMY, D-dimer, PT, and Hct were important prediction factors of approval for AP-induced PVT.
急性胰腺炎（AP）可诱发门静脉脾肠系膜静脉血栓形成（PVT），这可能导致更高的发病率和死亡率。然而，目前PVT的诊断方法仍存在争议。近几十年来，人工神经网络在医学研究中的应用越来越广泛。本研究旨在通过径向基函数（RBF）人工神经网络（ANNs）模型预测AP诱发PVT的风险。
对2011年1月1日至2016年7月31日期间在本单位的426例AP患者进行回顾性或连续性研究。所有患者均接受RBF ANNs分析。变量包括年龄、性别、红细胞比容（Hct）、凝血酶原时间（PT）、空腹血糖、D-二聚体、血清钙浓度（[Ca]）、甘油三酯、血清淀粉酶（AMY）、急性生理与慢性健康状况评分II以及兰森评分。所有结果均在对变量进行统计分析后得出。
在RBF ANNs模型中，D-二聚体、AMY、Hct和PT是所有11个PVT独立变量中的重要因素。它们的标准化重要性分别为100%、96.3%、71.9%和68.2%。RBF ANNs模型对PVT的预测敏感性、特异性和准确性分别为76.2%、92.0%和88.1%。在这些参数方面，RBF ANNs模型与逻辑回归模型之间存在显著差异（95%CI：分别为110.9%[-0.4至15.8%]；8.4%[-3.3至19.2%]；以及12.8%[1.6 - 20.7%]）。此外，使用RBF ANNs模型识别血栓时，受试者工作特征曲线下面积值为0.892±0.091（95%CI：0.805 - 0.951），显示出比逻辑回归模型（0.762±0.073；95%CI：0.662 - 0.839）更好的整体性能。
RBF ANNs模型是预测AP后PVT风险的有价值工具。AMY、D-二聚体、PT和Hct是AP诱发PVT的重要预测因素。
