Predictive Analytics Based Integrated Framework for Intelligent Healthcare Applications

Abstract

Healthcare analytics is a field that deals with the examination of underlying patterns in healthcare data in order to determine ways in which clinical care can be improved - in terms of patient care, hospital management and cost optimization. Towards this end, health information technology systems such as Clinical Decision Support Systems (CDSSs) have received extensive research attention over the years. A CDSS is designed to provide physicians and other health professionals assistance with clinical decision-making tasks, based on automated analysis of patient data and other knowledge sources. Recent advancements in Big Data and Healthcare Analytics have seen an emerging trend in the application of Artificial Intelligence techniques to healthcare data for supporting essential applications like disease prediction, mortality prediction, symptom analysis, epidemic prediction etc. Despite such major advantages o↵ered by CDSSs, there are several issues that need to be overcome to achieve their full potential. There is scope for significant improvements in terms of patient data modeling strategies and prediction models, especially with respect to clinical data of unstructured nature. In this research thesis, various approaches for building decision support systems towards patient-centric and population-centric predictive analytics on large healthcare data of both structured and unstructured nature are presented. For structured data, an empirical study was performed to observe the e↵ect of feature modeling on mortality prediction performance, which revealed the need for extensive study on the relative relevance of features contributing to mortality risk prediction. Towards this, a Genetic Algorithm based wrapper feature selection method was proposed, for determining the most relevant lab events that help in e↵ective patient-specific mortality prediction. Clinical data in the form of unstructured text, being rich in patient-specific information sources has remained largely unexplored, and could be potentially used to leverage e↵ective CDSS development. Towards this, an Extreme Learning Machine based patient-specific mortality prediction model built on ECG text reports of cardiac patients was proposed. The approach, which involved word iiiiv embedding based feature modeling and an unsupervised data cleansing technique to filter out anomalous data, underscored the importance of e↵ective word embeddings. Therefore, our next objective was to study the word embedding models and their role in feature modeling for building e↵ective CDSSs. A benchmarking study on performance of word representation models for patient specific mortality prediction using unstructured clinical notes was performed. Our next objective involved analyzing and utilizing the unstructured clinical notes for building e↵ective disease prediction models. An ontology-driven feature modeling approach was proposed, for designing a disease group prediction model built on unstructured radiology reports. In order to solve the problems of sparsity and high dimensionality of this approach, another feature modeling approach based on Particle Swarm Optimization (PSO) and neural networks was proposed to further enhance the performance of disease group prediction models using unstructured radiology reports. With the objective of analyzing physician notes, a hybrid feature modeling approach was proposed to leverage the latent information embedded in unstructured patient records for disease group prediction. Towards addressing the incremental and redundant nature of unstructured clinical notes, aggregation of nursing notes using TAGS and FarSight approaches were also explored for e↵ective disease group prediction, which demonstrated significant potential towards enabling early disease diagnosis. For population health analysis (flu vaccine hesitancy, flu vaccine behaviour and depression detection), a generic model called Multi-task Deep Social Health Analyzer (MDSHA) was proposed which uses a PSO based topic modeling approach for e↵ective feature representation and predictive modeling. All proposed approaches were compared to existing state-of-the-art approaches for respective prediction tasks using standard datasets. The promising results achieved underscore the superior performance of the approaches designed in this research, and reveal much scope for adaptation in the healthcare field for improving quality of healthcare.