Case-Based Medical Decision Making Support Technology For Diagnosis&treatment

The knowledge requirement of hospital diagnosis and treatment decision making on theknowledge has typical features which are consistant to the advantage of the case-based reasoningtechnology. However, due to the diversity, variability, complexity, uncertainty and dynamicevolution characteristics of of human illness, as well as the development and change from medicaltechnology and diagnostics, the existing case-based reasoning techniques are unable to meet therequirements of specific problem-solving in medical diagnosis and treatment decision making.Concerning key scientific issues in medical diagnostic decision-making around, the traditionalcase-based reasoning technology was improved and optimized for a more powerful CBR systemwith higher performance to adapt to a variety diagnosis and treatment decision-makingrequirements via the integration of statistical, fuzzy, gray system theory, genetic algorithms andother relevant theories. Then, CBR-IMDSS, an intelligent decision support system frameworkbased on advanced CBR was built and a improved CBR-based were used for the construction ofmedical decision support system which integrates the technical advantages from DSS and CBRrespectively for medical knowledge discovery. CBR-IMDSS can be used to assist the decisionmaking for the physicians during the process of early detection, diagnosis, prediction of a varietyof diseases. In addition to the integration of an open technical framework for intelligent treatmentdecisions based on the history case base, other related specific research content and the maincontributions are as follows.(1)The intelligent integrating and open medical decision support technology based oncase-based reasoning was explored and CBR-based Intelligent Medical Decision SupportSystems(CBR-IMDSS) was developed. The improved case-based reasoning was presented toconstruct medical decision support system. This technology combined the advantages of DSS andCBR for the data mining of medical case knowledge and can be used for early detection, diagnosis,prognosis and other medical decision process.(2) The studies with respect to the feasures' reduction methods were reviewed and then thefeasure reduction of cases for clinical decision making cases in hospitals was examined.â‘ theprinciple component analysis(PCA) and its mathematical models and methods were presented andits necessary conditions for use and the correlation test methods among variables were discussed.Wisconsin cancer data set was used for empirical tests. After the Barlett test, the calculation offeasure vectors and other computing processes, three main components were abtained. It showedthat PCA can reduce the data dimension.â‘¡Logistic regression, a common approach ofstatistical analysis and its applications in weight determination of the case in medical diagnosticdecision-making, and the corresponding theories and models were presented. Using UCI breastcancer data, we conducted an experiment and obtained the weights of feasures;â‘¢Three different approaches based on information entropy method for weight acquisition algorithm werestudied. The subsequent experiment showed that the entropy method can be used to weight theacquisition, and had better performance than that of Delphi expert weighting method.(3) In view of a kind of special clinical diagnosis issues in the hospitals, a fuzzymulti-attribute decision-making method based on case-based reasoning was studied. We exploreda comprehensive case retrieval method for fuzzy multi-attribute knowledge acquision requirementwas presented and multi-attribute fuzzy case-based reasoning framework for multi-attribute fuzzyprognosis was built. Different type of knowledge representation was examined and the complexcase retrieval problems with with symbolic attributes, the logic values, and two-dimensionalspatial values were addressed. According to the characteristics of fuzzy decision making cases, aswell as actual situation of hospitals, we performed some modifications on traditional nearestneighbor algorithm based on Euclidean distance to adapt the retrieval needs of spatial orientationvariables, and formed IKNN-CSFV, an improved algorithm for computing fuzzy spacial variables.In terms of fuzzy intervals among characteristic attribute values of hospital diagnosis andtreatment decision making cases, by integrating the concept of fuzzy sets into case-basesreasoning, we transformed the problem of case-based reasoning into fuzzy multi-attributedecision-making and developed FCA-PSIS algorithm for the calculation of interval fuzzy attributevariables in cases. Integrating the nearest neighbor algorithm based on Euclidean distance,IKNN-CSFV, FCA-PSIS, we can obtain FHRA-M method, a comprehensive fuzzy case retrievalalgorithm which is suitable for the characteristics of hospital diagnosis and treatmentdecision-making. Based on Columbia Saint Mary's Cancer Dataset, we empirically completed theeffectiveness validation of this method. Further, we empirically compared KNN, RBFNetwork,J48, and other relevant methods. The result showed FHRA-M has higher performance.(4) Knowledge acquisition of another kind of special diagnosis and treatment decisionmaking cases in which non-continuous feasures are dominant was studied. We integratedconditional probability and GAs into case-based reasoning technology and developed CRMGACPalgorithm which includes a GAs-based weight determation method and an improved similarityalgorithm integrating the conditional probability. CRMGACP method can be seen as an extensionmethod of traditional KNN. Compared to traditional KNN, its continuitious independent variablesremain the same but the logic variables are extended to all discrete ones. The computing ofcontinutious variables is based on the similarity algorithm based on Euclidean distance but thecalculation of discrete variables is based on conditional probability theory. Using VC++weimplemented a program called CancerCBRSys. Based on Anhui-based cancer data sets, weempirically compared the performance of four different case-based reasoning methods, KNNunder fixed weight (const. w), KNN under the weight from expert evaluation (expert. W),CRMGACP and CRM-IECP, a fixed case retrieval algorithm based on information Entropy andthe conditional probability, respectively. The selected statistics for performance evaluation of algorithms are accuracy, sensitivity, specificity and F-values. The results showed the CRMGACPhas the best performance in accuracy, sensitivity, precision and F-values. Naive bayes, Logistics,RBF network and the Simple Cart were selected to compare the performance based on the sametesting set and reference set and the results show that CBR (CRMGACP) has better performancethan other methods. In general, CRM-GACP shows significant advantage in comparison trials andis hopeful to be a powerful decision-making tool in clinical diagnosis.(5) The knowledge mining problem in gray diagnosis and treatment decision making case isexplored. In this study, we integrate information entropy theory and gray theory. The gray theoryis optimized and then integrated into the case reasoning. The knowledge mining of cases withincomplete Information, discrete attributes and point-to-point distance calculation is a kind ofcommom problem in hospital diagnosis and treatment decision-making. Considering differentimportance of attributes, weight is brought into the computing of comparative situation and anadvanced local gray relational algorithm is acquired. The experiment based on UCImammography data set suggests the information entropy can be obtained higher accuracy thanDelphi method. Specially, amongst four different fusion methods, the fusion of gray system theoryand information entropy obtained the best results. The second is the fusion algorithm of theinformation entropy and the nearest neighbor method based on Euclidean distance algorithm. Inaddition, the fusion of information entropy and gray theory gets the best sensitivity. The otherthree fusions have poor sensitivity, as well as weak stability. Another interesting discovery is thatalmost all of these fusion algorithms perform the worst in specificity when K is equal to one, butNOT always in sensitivity. Generally, in this experimental study of breast cancer decision-making,information entropy weight method performs better that Delphi method. Considering a variety ofcomprehensive performance, the integration of information entropy theory and case-basedreasoning method has significant advantage in grey case knowledge mining for decision making.In addition, the selection of K values in the KNN is explored. In theory, it seems tha the bestmatching target case should be the most similar case, not the second (or the third, or even later)case. In other words, accuracy should be the highest level when k=1. In the study of the gray casedecision-making in Chapter VI, we examined the influence from changing k value on the retrievalaccuracy. Subsequent experimental result shows that the accuracy is not at the highest point whenis equal to one.