基于数据挖掘的低增生性骨髓增生异常综合征与再生障碍性贫血分类模型研究

目的 构建低增生性骨髓增生异常综合征（hypo-MDS）与再生障碍性贫血（AA）鉴别诊断的决策树、贝叶斯、卷积神经网络、改进的支持向量机四种模型并选择出最优模型。方法 收集2010—2019年华北理工大学附属医院的AA与hypo-MDS患者的病例资料，使用统计学方法筛选指标，将处理后的样本以4DK]∶1随机分为训练集和测试集，构建决策树、贝叶斯、卷积神经网络、改进的支持向量机四种模型，采用五折交叉验证法多次重复验证，通过灵敏度、AUC等指标评价鉴别诊断效果。结果 hypo-MDS患者红细胞、血红蛋白含量等指标低于AA患者，成熟单核细胞比例等指标高于AA患者，年龄和职业分布也存在差异(P＜0.05）；最终选出21个特异性指标。四种模型的分类效果比较:灵敏度分别为82.56%、65.12%、87.21%、79.07%；AUC分别为0.81、0.68、0.82、0.83；准确率分别为75.32%、69.48%、77.27%、74.03%。对卷积神经网络的误判病例分析得出年龄、血成熟淋巴细胞等7个指标均存在差异（P＜0.05）。结论 在决策树、贝叶斯、卷积神经网络、改进的支持向量机四种诊断模型中，卷积神经网络具有最佳分类效果。

Classification model of hypocellular myelodysplastic syndrome and aplastic anemia based on data mining

To construct four models of Decision Tree，Bayes，Convolutional Neural Network，Genetic Algorithm ofthe Support Vector Machines for the differential diagnosis of hypocellular myelodysplastic syndrome ( hypo － MDS) and aplasticanemia ( AA) and choose the best model. Methods The case data of AA and hypo － MDS patients of the Affiliated Hospitalof North China University of Science and Technology from 2010 to 2019 was collected，statistical methods to screen indicatorswere used，and the processed samples were divided into training set and test set randomly at 4∶ 1 to construct decision trees andshellfish. The four models of Decision Tree，Bayes，Convolutional Neural Network，Genetic Algorithm of the Support VectorMachines repeated the verification many times by using the five － fold cross － validation method，and we evaluated thedifferential diagnosis effect through indicators such as sensitivity and AUC. Ｒesults The red blood cell and hemoglobincontent of hypo － MDS patients were lower than those of AA patients，and the ratio of mature monocytes was higher than that ofAA patients. There were also differences in age and occupational distribution ( P ＜ 0. 05) . 21 specific indicators were finallyselected. Comparison of the classification effects of the four models were as follows: sensitivities were 82. 56% ，65. 12% ，87. 21% ，79. 07% ，AUC were 0. 81，0. 68，0. 82，0. 83 and accuracy rates were 75. 32% ，69. 48% ，77. 27% ，74. 03% .Analysis of the misjudgment cases of the convolutional neural network showed that there were differences in 7 indicatorsincluding age and blood mature lymphocytes ( P ＜ 0. 05) . Conclusion Among the four diagnostic models of Decision Tree，Bayes，Convolutional Neural Network，Genetic Algorithm of the Support Vector Machines，Convolutional Neural Network hasthe best classification effect