| ObjectiveClinical laboratory examination is one of the basic means of disease diagnosis,formulation of therapy protocols, and evaluation of therapeutic effects for doctors.However, studies have shown that some clinical laboratory examination are unnecessaryor inappropriate. There are many reasons for this phenomenon, one of which is the lack ofclinical decision supporting systems on clinical laboratory examination. The main purposeof this study is to explore a knowledge discovery method based on Self-OrganizingMaps(SOM) in historical database and apply it into clinical decision supporting systemson clinical laboratory examination and to discover the application regularity of the clinicallaboratory examination, based on expert experience and majority suggestions, in order tosupply scientific evidence for proper selection and standardized application of clinicallaboratory examination items, and to provide new ideas for SOM application in the medical field，which has great practical significance to improve the level of clinicallaboratory examination and efficiency of health care services and to control the excessivegrowth of health care costs.Methods1. The2009-2011outpatients’ data from hospital information system and laboratoryinformation system of two largest comprehensive hospitals in Xi’an were collected usingSQL query language and were connected, transformed, cleaned,and screened. Accordingto papers, information such as age, sex, the numbers of clinical laboratory examination inthree years, the academic title of doctors, payment mode, clinic tests, and preliminarydiagnostic characteristics of patients were selected as model variables.2. Model variables of5756patients with abdominal pain, fever and arthralgia wereused to establish SOM and k-means clustering. Different parameters and step lengths wereapplied to train SOM model.The number of clustering of k-means was formulated basedon the the number of clustering of SOM. X~2test was applied for difference comparison ofsex, the academic title of doctors, payment mode of patients. The further comparison incouples was tested by X~2segmentation test. One-way ANOVA was applied for differencecomparison of age, the numbers of clinical laboratory examination in three years ofpatients. The further comparison in couples was tested by SNK-q. DBI of two models wascalculated to evaluate dipartite degree and clustering effect of two model.3. Combined with the practical application of each cluster of patients’ clinicallaboratory examination, reference for each cluster of patients was made based on the resultof clustering.15clinical outpatient experts were invited to test clustering effect based ontheir clinical expertise. Evaluation results were divided into proper and improper results,and proper rate of clustering results of two models were calculated. The quantitative indexof consistent rate of two models and kappa value were calculated to evaluate clusteringeffect, and to show which was more suitable in the clinical laboratory examination field.4. Model variables of15,999patients with connective tissue disease, ankylosingspondylitis, respiratory infections, chronic gastritis, epilepsy, liver damage, viral hepatitisB, prostatitis, rheumatoid arthritis and abdominal pain were used to establish SOM.Different parameters and step lengths were applied to train SOM.―Clever variable‖were confirmed by feature map of data sets attributes distribution.2test was applied fordifference comparison of sex, the academic title of doctors, pay mode of patients.One-way ANOVA was applied for difference comparison of age, the numbers of clinicallaboratory examination in three years of patients. Results of SOM were described bycolumnar section to summarize up characteristics of each cluster of patients.5. The quantitative index of DBI was calculated to evaluate clustering effect. Wedrafted reference scheme of clinical laboratory examination of each clustering patientsbased on the practical application of clinical laboratory examination.50drafted schemesof each clustering were evaluated by20clinical outpatient experts. And the evaluationresults were divided into proper and improper types. Appropriate rate of each cluster wascalculated to evaluate two models.The SPSS16.0software was used for k-means, One-way ANOVA, test. And SOMwere established by MatlabR2009b based on SOM tool box.Results1.5756patients were clustered by Self-Organizing Maps and the number ofclustering was3. There was statistical significance between the groups of each cluster(P<0.05). Proportion of patients with male sex and non-professional admission in cluster Iwere significantly higher than that of cluster Ⅱ (P<0.015). Proportion of patients withtheir own expenses in cluster I was significantly higher than that in other clusters(P<0.015). Age of patients in cluster Ⅱ was significantly higher than that in otherclusters, which was58.48±9.35. The number of clinical laboratory examination in threeyears of cluster Ⅲ was significantly higher than that of other cluster, which was2.87±1.85.2. In order to compare with the results of SOM, we defined the k as3in k-meansmodel. There was statistical significance between the groups of each cluster (P<0.05).Proportion of patients with male sex and non-professional admission in cluster I wassignificantly higher than that in other clusters (P<0.015). Age of patients in cluster Ⅲwas significantly higher than that in other clusters, which was61.81±8.47. The number ofclinical laboratory examination in three years of clustert Ⅱ was significantly higher thanthat in clusters Ⅲ, but there was no statistical significance among other clusters (P>0.05). 3. Clustering results of comparison between Self-Organizing Maps and K-meansclustering model showed that difference of proportional of preliminary diagnosticcharacteristics of patients was lagest in SOM and DBI of SOM was smallest(DBI=0.82)and dipartite degree of SOM was best. Results of experts showed that appropriate rate ofSOM was61.29%, and appropriate rate of K-means was50.69%, and the consistent rate oftwo models was77.87%and the kappa value was0.556.4. Outpatients data of top ten preliminary diagnostic characteristics of patients wereclustered to8clusterings based on SOM. Analysis results showed that the difference ofsex, the academic title of doctors and payment mode of patients had statistical significance(P<0.05). ANOVA-analysis results showed that the difference of age and the number ofclinical laboratory examination in three years had statistical significance (P<0.05).5. Sex, age, the number of clinical laboratory examination in three years andpreliminary diagnostic characteristics of patients had made great contribution to theclustering models, which were called clever variable‖.6. The quantitative index of DBI was calculated to evaluate clustering effect andDBI of SOM was0.86. Experts results showed that qualified rate was76.4%.Conclutions1. Although consistency of the two models is better, the result of SOM has betterdipartite degree, much more significantly difference in each cluster and higher appropriaterate than that in k-means. It indicates that SOM has more advantages than k-means in datamining of clinical laboratory examination domain, and higher value of application andpopularization.2. Sex, age, and preliminary diagnostic characteristics of patients have made greatcontributions to the clustering models. According to these characteristics, doctors couldprescribe the proper clinical laboratory examination items for the patients.3. SOM has better effects on patients clustering who have used clinical laboratoryexamination items. And results of Self-Organizing Maps are in accord with medical theory,which could guide clinical laboratory examination practice. |
PDF Full Text Request