基于神经网络模型和血常规指标的孕妇地中海贫血预测研究

Prediction of thalassemia in pregnant women based on neural network model and blood routine indexes

目的运用神经网络模型,研究重庆市江津区地中海贫血(简称地贫,即珠蛋白生成障碍性贫血)孕妇的血常规指标对地贫的预测效果。方法收集2016年1月至2022年12月该院9652例产前检查孕妇的临床资料、血常规及地贫基因检测结果。在9652例孕妇中,检出地贫基因阳性847例,阴性8805例。在地贫基因阳性和阴性孕妇中分别随机抽取424例和4402例作为试验组,剩余423例地贫孕妇与4403例阴性孕妇作为对照组。构建随机森林模型、Logistic回归模型分析地贫预测效果较好的血常规指标,建立神经网络模型区分地贫类型为α或β。结果847例地贫阳性孕妇中,α地贫阳性569例(5.90%),β地贫阳性267例(2.77%),α合并β地贫阳性11例(0.11%)。通过随机森林模型筛选出对诊断地贫重要性较高的5项血常规指标平均红细胞体积(MCV)、平均血红蛋白含量(MCH)、红细胞体积分布宽度变异系数(RDW-CV)、平均血红蛋白浓度(MCHC)、红细胞计数(RBC),采用Logistic回归模型进行验证,受试者工作特征(ROC)曲线分析显示,5项指标的曲线下面积(AUC)分别为0.906、0.904、0.785、0.783、0.780,模型AUC为0.906。采用随机森林筛选的血常规指标,建立神经网络模型预测α地贫、β地贫并优化,得出α、β地贫预测值均为100%。结论该研究建立的基于血常规指标的神经网络模型对α、β地贫预测率较高,可为地贫的早期筛查提供了新的思路。

Objective To investigate the prediction effect of blood routine indexes on thalassemia in pregnant women in Jiangjin district of Chongqing by using neural network model.Methods From January 2016 to December 2022,the clinical data,blood routine and thalassemia gene test results of 9652 pregnant women were collected.Among 9652 pregnant women,847 were positive and 8805 were negative for thalassemia gene.Among the positive and negative cases,424 and 4402 were randomly selected as the test group.In contrast,the remaining 423 thalassemia pregnant women and 4403 negative pregnant women were assigned to the control group.The random forest model and Logistic regression model were constructed to analyze the blood routine indexes of thalassemia with good prediction effect,and the neural network model was established to differentiate betweenα-thalassemia andβ-thalassemia.Results Among 847 pregnant women with thalassemia genes,569(5.90%)were positive forα-thalassemia,267 cases(2.77%)were positive forβ-thalassemia,and 11(0.11%)were positive forαcombined withβthalassemia.Furthermore,five blood routine indexes MCV,MCH,RDW-CV,MCHC and RBC,which were of high importance in the diagnosis of thalassemia,were selected by random forest model.These five indexes were verified by Logistic regression model.Receiver operating characteristic(ROC)curve analysis showed that the area under the curve(AUC)of the five indexes were 0.906,0.904,0.785,0.783 and 0.780,respectively,and the overall AUC for the model was 0.906.Using the blood routine indexes of random forest screening,a neural network model was established to predict and optimizeαandβthalassemia,and the predicted values of bothαandβthalassemia were 100%.Conclusion The prediction rate ofαandβthalassemia based on the neural network model established in this study is high,which provides a new idea for the early screening of thalassemia.