An Adaptive Classification and Recommendation Model for E-Health

Abstract

FACULTY OF TELECOMMUNICATION AND INFORMATION ENGINEERING Department of Software Engineering

Doctor of Philosophy

An Adaptive Classification and Recommendation Model for e-Health Systems By Anam Mustaqeem 13F-UET/PhD-SE-06 E-health based system is an advanced topic of research, showing an enormous amount of effort for providing an efficient response to cardiac diseases. Health monitoring of patients in particular senior citizens at their home based location is categorized among the wide ranged applications of today’s health care systems. Health professional track the clinical condition of elderly patients at remote locations using monitoring devices, which otherwise would have to admit to a medical caring unit. The purpose of e-health is to highlight the issues regarding health care and therefore, reduce the chances of hospitalization, help in improving quality of life style, and saving money. With the advent of e-health, the information about most critical issues of patients can be accessed from far away locations. A promising and active research area, which has benefited from the e-health based system is cardiac monitoring and care. In this research work, we intend to develop an adaptable and intelligent recommendation based model for e-health systems. As cardiac diseases are one of most critical and life threatening disease among chronic diseases, therefore cardiac disease classification and iv

recommendation is addressed in this study. The research work has been categorized into three different areas. The results have been evaluated using standard evaluation metrics and an improved accuracy is obtained in all research tasks. A model is proposed using a standard dataset for arrhythmia classification to provide improved classification accuracy. The approach for classification combines feature selection, pre-processing and classification techniques, and provides promising diagnosis results. Further, normalization is done for scaling and standardizing the data parameters. An improved feature selection method using a wrapper method around the random forest (RF) is employed to select the most significant features achieving higher classification accuracies for the UCI arrhythmia dataset. The selected features help in achieving better accuracy and efficient classification performance. In the second part of the thesis, the details are presented for collection, analysis and processing of a customized dataset to implement recommender system for cardiac patients under the supervision of medical experts. Although recommender systems have emerged in various domains, development of a clinically apporoved medical recommender systems still require a long way to go, as medical recommendations directly affect the life of patients. We have proposed a hybrid machine learning based prediction and risk analysis based recommendation model for detection of heart disease which provides suitable medical advice to a patient depending on the type of disease identified. The results show that the proposed medical recommender system will be a significant contribution in the field of cardiac disease classification and recommendation. A medical recommender system is implemented using a modular clustered based collaborative filtering model, which is an improvisation in the traditional collaborative filtering technique to target the issues of sparsity and scalability. Sub-clustering at two levels is introduced to ensure fast and robust similarity computations. The involvement of cardiac experts in the whole process is made possible for clinical approval and disapproval of the outcomes.