Article Preview
TopIntroduction
The availability of electronic medical records and laboratory results has triggered the applicability of machine learning in health care domain. Medical records, patient history, scan reports, laboratory test results have become potential sources for data analysis. The information gleaned from these analyses can be used to help health care practitioners diagnose and treat diseases, resulting in better patient care. For example, prediction of diseases, tumour classification from MRI images, and patient monitoring using data from remote sensors together with clinicians’ intervention enhances decision making process.
Zikos et al. (2018) have developed a conceptual model for Clinical Decision Support System (CDSS) by integrating patient history data, diagnostic test results and physical examination results to arrive at disease diagnosis, selecting appropriate treatments and to keep track of discharge and patient transfer information. Elucidating hidden patterns from medical data using data mining have widely been reported in the literature. Prediction of diseases and abnormalities using different classification algorithms can be found in literature. A CDSS for classifying acute renal failure using various classification algorithms has been proposed by Al-Hyari et.al (2013). A rule based classifier for the diagnosis and treatment of Allergic Rhinitis has been proposed by Christopher et.al (2015). A neural network based computer aided diagnostic system which identifies the extremity of Parkinson’s disease is described by Jane et.al (2016). A CDSS framework with a feature selection technique based on correlation is presented for diagnosing hepatitis and breast cancer in Elgin Christo e.al (2019). A method for determining rough set indiscernibility using a back propagation neural network is mentioned in Nahato et.al (2016) for the prognosis of breast cancer, cardiac disease and hepatitis. A CDSS framework addressing class imbalance and feature selection of clinical datasets is discussed by Sreejith et.al (2020). A Time delay neural network for classifying clinical data having time stamped data is proposed by Jane et.al (2021). A clinical classification framework for diagnosing vertebral disease, heart disease and diabetes using a customized neural network is proposed by Sreejith et.al (2020). A neural network based classification which uses genetic algorithm has been used for diagnosis of cardiac infarction in Amin et.al (2013). An adaptive genetic algorithm with fuzzy logic model has been proposed by Reddy et.al (2020) for diagnosing heart disease at an earlier stage.
Data mining is a collection of techniques that uses widely available data in a variety of disciplines to create models that aid decision-making (Han et.al 2011). The available clinical data can be used to uncover previously unknown knowledge, and the details gained from such analysis are used to assist the prognosis. The three main data mining tasks are association rule mining, classification, and clustering.
Association rule mining is an unsupervised task of mining frequent patterns from data thereby finding associations between data points that appear together. Classification is a supervised data mining strategy which is used to predict class labels of unknown instances by building a trained model which learns from labelled data. The two step process in classification involves a training phase built on known labelled data and testing phase which makes future predictions on test data. Clustering is an unsupervised data mining technique in which comparable observations are categorized based on their similarity, and then each group is given a label. ANN, a paradigm modelled after the neuron of the human brain, is widely deployed for solving classification problems. An ANN is made up of nodes that are stacked in layers to help map input to output. Connections existing between these layers transform the data by adjusting the learnable parameters viz. weights and bias with them. By passing the training data through these layers, the algorithm attempts to optimize the learnable parameters.