An Evolutionary Approach To Cascade Multiple Classifiers: A Case-Study To Analyze Textual Content Of Medical Records And Identify Potential Diagnosis
Hegler Correa Tissot · International journal of scientific and technology research · 2014
Abstract : This paper describes an experiment where classifiers are used to identify potential diagnoses on examining textual content of medical records. Three classifiers are applied separately (k-nearest neighborhood, multilayer perceptron and support vector machines) and also combined in two different approaches (parallel and cascading); results show that even accuracy point to a best alternative, ROC analysis show that choosing an approach depends on an acceptable error level. Index Terms : Pattern Recognition, Multilayer Perceptron, Support Vector Machine, K-Nearest Neighborhood, Differential Evolution Algorithm, Parallel and Cascading Classifiers, Medical Records, Potential Diagnosis ———————————————————— 1 I NTRODUCTION WITH the increasing volume of textual content that is being made available, science related to information management has evolved in recent years to develop new system modeling and building techniques to deal with unstructured data formats. As the interest in finding and sorting information from text documents is growing, text mining emerges as a technology which the purpose of extracting non-trivial and interesting knowledge from large collections of unstructured documents [43]. Many intelligent diagnostic systems have been employed to assist condition monitoring tasks, such as expert systems and Artificial Neural Networks (ANNs), support vector machines and fuzzy logic systems, with promising results of such techniques [16], [33], [35]. However, individual decision system can only acquire a limited classification capability that is only appropriate for special data and may not be enough for a particular application. The application of a decision fusion system (DFS) has received considerable interest in recent years, achieving considerable successes to solve complex pattern recognition tasks. DFS can be also called multiple classifiers fusion (MCF), combination of classifiers, multiple experts or hybrid method. Due to the integration of different decisions from multiple classifiers, the technique can boost the recognition accuracy of in may applications [29]. Several applications and experiments have been developed using textual information in order to classify documents according to certain criteria, and using classifiers is a common method to find patterns in such documents. A large number of case studies can be found in the literature comparing and combining different classifiers [3], [5], [7], [10], [17], [23], [24], [25], [26], [27], [31], [34], [41], [46]. This paper describes and compares the results obtained with separately and combined classifiers when trying to identify possible diagnosis from analyzing textual content of medical records. To support this experiment, three classifiers were used: 1) K-Nearest Neighborhood (KNN), 2) Multilayer Perceptron (MLP) and 3) Support Vector Machines (SVM). Section 2 gives the basic pattern recognition concepts and describes classifiers used in this experiment; sections 3 and 4 explain how classifiers can be combined and how an evolutionary algorithm can be used to find cascading parameters; section 5 describes a case study and objectives of the experiment; section 6 details the methodology used to run this experiment; section 7 presents the obtained results and section 8 closes with the conclusions and future work suggestions.