Bone disease prediction and phenotype discovery using feature representation over electronic health records
Hui Li, LI Xiao-yi, Xiaowei Jia, Murali Krishna Ramanathan, Aidong Zhang · 2015
With the expansion of the healthcare industry and the overwhelming amount of electronic health records (EHRs) shared by healthcare institutions and practitioners, we now wish to take advantage of EHR data to develop an effective disease risk management model that not only improve disease prediction but also shows clinically meaningful feature grouping which can be used to define the bone disease phenotypes. In this paper, we explore the feasibility of extracting critical risk factors for osteoporosis and bone fracture prediction based on heterogeneous electronic health records (EHRs). This will improve our understanding of bone disease risk arising from the complex interplay of the human bone mineral density (BMD) assessment with major risk factors such as gender, age, family history, and life styles. We focus on addressing the problem of identifying individual or integrated risk factors (RFs) to predict osteoporosis and bone fractures, which are common diseases associated with aging and may be clinically silent but cause significant mortality and morbidity. On the other hand, the unbiased, EHR-driven phenotype discovery could be achieved using a massive EHR dataset and a computationally intense analysis capable of identifying all of the phenotypes in the dataset. We infer the precise phenotypic patterns from the new feature representation from EHR data for grouping risk factors for osteoporosis. We present a 2-layer deep graphical model to use EHR with minimal human supervision which derives a new representation of medical objects by embedding them in a low-dimensional vector space. This new low-dimensional risk factor representation will ultimately present the risk factor embedding/clustering by offering intuitive visualization by projection onto a 2D plane. We demonstrate the capability of our framework on a 9704 Caucasian women of 20 years of prospective dataset under osteoporosis and bone fracture assessment. The integrated representation not only improves the risk prediction accuracy but also presents clinically meaningful risk factor grouping for bone diseases.