Feature selection and neural network optimization using evolutionary algorithms for clinical decision making

Abstract

Decision making is much critical in medical domain where an error in judgment can lead to many problems such as imposing unnecessary expensive treatment to patients, denying treatment to patients who actually eed it and exposing a patient unnecessarily to invasive procedures. The useful knowledge extracted from Electronic Health Records can be used to develop Clinical Decision Support Systems (CDSS) that can aid physicians in disease diagnosis. Developing scalable and effective frameworks for CDSS is a significant topic of research in clinical knowledge mining. The objective of this research work is to develop classification frameworks for the development of CDSS. Precisely, the work concentrates on two applications of evolutionary algorithms to enhance clinical data classification. First, to improve classifier performance by selecting the optimal number of features, second, to optimize the topology and learnable parameters of neural networks. These two research objectives have been developed into four contributions. In the first contribution, k-NN imputing approach is used for handling the missing values present in the dataset. An embedded technique based on an Extremely Randomized Tree classifier is used for feature selection. The learnable parameters viz. weights and bias of a feed forward neural network named as DISON is optimized using a combination of Back propagation algorithm and Strawberry Plant Optimization algorithm and is used for clinical data classification. Four datasets namely Vertebral Column, Pima Indian Diabetes (PID), Cleveland Heart Disease (CHD) and Statlog Heart Disease (SHD) from the University of California Irvine (UCI) machine newline