Evidential Chains-based Reasoning For Robust Classification And Its Decision Support For Cardiac Diagnosis

Data-driven decision-making widespreads in engineering practice and management, and new challenges are identified in the knowledge based reasoning theories and methods of data fusion. Conventional decision-making methods are no longer applicable to these decision data, sourcing from multi-sensor data, relational databases, various levels of expertise, etc. Therefore, based on the framework of evidence fusion, this thesis investigates the robust classification decision mechanism with evidential chains-based reasoning(FUER). Theories and methodologies of multi-source data acquisition, message passing and information sharing, case/rule-based reasoning(CBR/RBR) classification and intelligent decision-making are synthetically utilized. The main work and innovations are as follows:First, FUER is investigated from the perspective of hierarchical association. The current researches on evidential reasoning and group decision making are systematically reviewed. Knowledge structure of evidential chains(ECs), belief preference and exponential similarity are defined, and their properties are analysised for classification decisions from the three layers of attributes, features and labels.Second, multi-resource evidential chains-based reasoning(mr FUER) model is constructed, revealing the reasoning mechanism for heterogeneous entities in the context of data-driven decision making. First based on the association of evidential chains from a single datasheet, the combined belief is derived for the query. Through improving the similarity frequency-weighted nearest neighbor algorithm(sf-NN), the impacts of the mis-labeled entities are analysised for the decision structures. Then, the orthogonal combination rule is formulated for ECs fusion. The mr FUER model is extended for evidential chains association among multiple datasheets, providing robust decision with high interpretability of the x-datasets NN algorithm(x D-NN).Third, to reveal decision-making mechanisms in temporal systems at different time scales, evidential chains is expanded into multi-scale reasoning from single-scale reasoning, through the proposed multi-scale evidential chains-based reasoning model(ms FUER). Using the similarity matrix and identification criteria, the quadratic optimal model is constructed for establishing the identification framework of classification, with the robust feature weight vector acquired. Thereafter, using multi-scale mutual information, we present the bilevel mixed-integer optimization model for generating the strategies of multi-scale feature selection, aiming to solve the problem of feature combination explosive growth and maximizing the information value of inference. The temporal similarity based nearest neighbor algorithm(ts-NN) is proposed with its reasoning mechanism superior to that of the conventional single-scale decision.Forth, to reveal the dynamic mechanism of the decision-making system under the process of sequential percepts, evidential chains evolves to sequential reasoning from non sequential reasoning, providing the rules of state transition and belief updating with ambiguous information. The assumption that the percepts of the query has full information has been relaxed, sequential data-percepts FUER model(sd FUER) is constructed for solving the queries with partial information. The likelihood probabilities are estimated with the sensitivity and specificity of a single feature. Using the proposed Dirichlet function based belief updating algorithm(sf-BN), the robustness of the decision is effectively achieved with the percept sequence.Finally, the proposed FUER model and its variants are verified with simulation experiments on medical decision-making support. The experimental data sources from the multiple benchmark data sets, including 4240 cases of cardiovascular disease from Framingham Heart Study, 920 heterogeneous cases from three medical institutions and temporal data from Multiparameter Intelligent Monitoring in Intensive Care(MIMIC II) database. Using these data, the knowledge based is constructed for the prototype of the proposed clinical decision making support system. From the perspectives of balanced accuracy and the length of the evidential chains, the results of the system demonstrate that the decision quality and efficiency is improved for the h and l types of experts. As a result, the robustness of decision is enhanced for classification.