The Differential Diagnosis Model Study Of Hypocellular Myelodysplastic Syndrome And Aplastic Anemia

Objectives To build hypo-MDS and AA identification & classification models by applying logistic regression,decision tree,BP neural network and support vector machine,so as to provide new ideas and methods for their identification and diagnosis.Methods The hypo-MDS patients and AA patents,diagnosed by Institute of Hematology of Chinese Academy of Medical Sciences and Affiliated Hospital of North China University of Science and Technology during January 1,2008 and December 31,2016,are taken as the research objects.All hypo-MDS patients and AA patents meet MDS standard revised by WHO in 2008 and the Diagnosis & Curative Effect Standard for Hematology(the Third Edition),and all information of the research objects are fully recorded.We collect the basic information of the research objects,peripheral blood count,blood smear,bone marrow smear.Exeel software,Medeler14.1softwareand Medcale software are applied to build logistic regression model,decision tree model,BP neural network model and SVM model for hypo-MDS and AA.We obtain the optimal classification model by comparing the sensitivity,specificity,Youden index,+LR,-LR,AUC,precision,Kappa value,PV+ and PV-of the four classification models,and analyze the characteristics of misdiagnosis data of the optimal model.Results 1.286 patients are researched: there are 130 hypo-MDS patients,including 69 male cases(accounting for 53.08%)and 61 female cases(accounting for 46.92%);and there are 156 AA patients,including 83 male cases(accounting for 53.2%)and 73 female cases(accounting for 46.8%).The differences between the two groups of patents in age and occupation are statistically significant(P<0.05),and other basic information is not statistically significant(P>0.05).2.The blood cell count displays that,compared with AA patients,hypo-MDS patients have lower content of erythrocyte and content of hemoglobin,which is statistically significant(P<0.05).The proportion of mature lymphocytes of hypoMDS patients is lower than that of AA patients,which is statistically significant(P<0.05).The bone marrow cell morphology displays that,compared with AA patients,hypo-MDS patients have lower proportion of basophilic-nucleus neutrophil,orthochromatic-nucleus neutrophil,segmented neutrophil,mature lymphocyte,while have higher proportion of basophilic erythroblast,polychromatic erythroblasts and orthochromatic erythroblast,which is statistically significant(P<0.05).3.Hypo-MDS and AA models based on logistic regression,decision tree,BP neural network and SVM are built.For the samples of the training set,the sensitivity of logistic regression,decision tree,BP neural network and SVM is 67.03%,86.60%,69.23% and 67.03% respectively,the specificity is 75.47%,97.17%,78.30% and 78.30% respectively,the Youden index is 0.43,0.84,0.48 and 0.45 respectively,the positive likelihood ratio is 2.73,30.60,3.19 and 3.09 respectively,the negative likelihood ratio is 0.44,0.14,0.39 and 0.42 respectively,the area under ROC curve is 0.71,0.95,0.74 and 0.73 respectively,the classification precision is 71.57%,94.92%,74.11% and 73.10% respectively,the Kappa value is 0.43,0.90,0.48 and 0.46 respectively,the positive predictive value is 70.11%,96.55%,73.26% and 72.62% respectively,and the negative predictive value is 72.73%,93.64%,74.77% and 73.45% respectively.The differences in sensitivity between logistic regression model and decision tree model and between decision tree model and SVM model are statistically significant(P<0.05);the differences in specificity between logistic regression model and decision tree model,between decision tree model and BP neural network model and between decision tree model and SVM model are statistically significant(P<0.05);the differences in precision between logistic regression model and decision tree model,between decision tree model and BP neural network model and between decision tree model and SVM model are statistically significant(P<0.05).the differences in AUC between logistic regression model and decision tree model,between decision tree model and BP neural network model and between decision tree model and SVM model are statistically significant(P<0.05).For the samples of the test set,the sensitivity of logistic regression,decision tree,BP neural network and SVM is 82.05%,84.62%,82.05% and 79.49% respectively,the specificity is 72.00%,76.00%,68.00% and 68.00% respectively,the Youden index is 0.55,0.61,0.50 and 0.47 respectively,the positive likelihood ratio is 2.93,3.53,2.56 and 2.48 respectively,the negative likelihood ratio is 0.25,0.20,0.26 and 0.30 respectively,the area under ROC curve is 0.77,0.80,0.73 and 0.76 respectively,the classification precision is 76.40%,79.78%,74.16% and 73.03% respectively,the Kappa value is 0.53,0.60,0.49 and 0.46 respectively,the positive predictive value is 69.57%,73.33%,66.67% and 65.95% respectively,and the negative predictive value is 83.72%,86.36%,82.93% and 80.95% respectively.The differences among the four models in sensitivity,specificity,precision and area under ROC curve are not statistically significant(P > 0.05).4.130 hypoMDS patients are identified and classified,and 13 hypo-MDS patients are misclassified as AA patients.Through comparison between the misdiagnosed cases and correctly diagnosed cases,it is found that,compared with the correctly diagnosed cases,misdiagnosed cases have higher content of erythrocyte and content of hemoglobin in peripheral blood cell count,have lower proportion of basophilic erythroblast,proportion of polychromatic erythroblasts and proportion of orthochromatic erythroblast in bone marrow smear,while have higher proportion of mature lymphocyte.The differences are statistically significant(P<0.05).156 AA patients are identified and classified,and 15 AA patients are misclassified as hypo-MDS patients.Through comparison between the misdiagnosed cases and correctly diagnosed cases,it is found that,compared with the correctly diagnosed cases,misdiagnosed cases have lower content of erythrocyte and content of hemoglobin in peripheral blood cell count,have higher proportion of basophilic erythroblast,proportion of polychromatic erythroblasts and proportion of orthochromatic erythroblast in bone marrow smear,while have lower proportion of mature lymphocyte.The differences are statistically significant(P<0.05).Conclusions Among the four classification algorithms including logistic regression,decision tree,BP neural network and SVM,decision tree algorithm has the optimal effect in identification of hypo-MDS and AA and can assist clinicians in identifying and diagnosing the two diseases.