Semantically Enhanced Tensor Factorization For Medical Information Retrieval And Recommendation Systems

With the dramatic growth of digital publications and the explosion of information on the Internet,medical information retrieval and recommendation systems are playing an increasingly important role to assist physicians and patients to better access knowledge and information.The recent advances in data mining have the potential to enable medical knowledge discovery to support decision making and improve the healthcare outcomes.The Semantic Web has been widely used for intelligent information processing by providing the machine-processable knowledge to enable semantic reasoning.To address problems such as incompleteness and sparsity,this study incorporates tensor-based methods to take advantages of their capability to deal with multiaspect data.This dissertation aims to leverage the semantic web and tensor factorization methods for data analytics in a number of healthcare problems.(1)To enable better understanding of the unstructured text and address the semantic gap,semantic expansion networks are presented to capture more semantic evidence for analysis using the complex network theory.A rescaled centrality is proposed to enhance the performance for semantic relevance analysis.Through leveraging the widely-used word embeddings and semantic enrichment,an automated clinical evidence grading system is presented to identify and recommend high-quality medical research articles.Experiments show that the proposed methods could improve the performance for medical text classification.(2)This study explores factorization-based methods for healthcare data mining.The widely-used recommender system techniques are adopted for data-driven chronic disease prediction.Tensor-based methods are proposed to incorporate the clinical attributes or temporal information for personalized chronic disease prediction based on nonnegative tensor factorization.Experiments with large-scale electronic health records demonstrate the superior performance of the proposed methods.For location-based fall incidents prediction,a fourth-order tensor with two moving average methods is presented.The tensor-based methods could improve the predictive performance compared with other widely-used machine learning methods.To address the problem of missing data for medical data mining,this study explores the factorization machine for medical data classification tasks.(3)A semantically enhanced medical information retrieval system is proposed with two-stage query expansion strategies extracting information from structured and unstructured resources respectively.To address problems of incompleteness and sparsity,a tensor factorization method is applied to estimate the importance of semantic association triples with a minor contribution of the incremental pseudo relevance feedback method.Experiments with a publicly available dataset demonstrate that the performance of the proposed system is significantly better than the baseline systems and comparable with the state-of-the-art systems.(4)A recommendation system is presented for medical community question answering.The proposed method models the ternary relationship of question-tag-expert using a tensor-based approach and adapts a tensor completion algorithm to identify high quality answers.Experiments demonstrate that the proposed system is able to improve the performance benefitting from the combination of information retrieval and recommendation methods.This dissertation demonstrates the advantages of incorporating the semantic web and tensor-based methods for semantic analysis and healthcare data mining.The semantically enhanced tensor-based framework presents superior performance for tasks such as medical information retrieval and recommendation systems.