Processing of Electronic Health Records using Deep Learning: A review

Health care systems across the world are adopting Electronic Medical Records (EMR), providing a collection of longitudinal data pertaining to patients' health. Use of electronic records will significantly increase the availability of clinical data as well as impact on the potential of discovering of new disease patterns as well as providing personalized patient care by automatically processing of this vast quantity of data. Considering that richness of information contained in EMRs [1] and their potential to transform delivery of care [2], understanding the information contained in EMRs is becoming an important challenge. In this respect the rise of Deep Learning is accelerating automatic processing and understanding of EMR data. Especially the use of novel Deep Learning methods and architectures that can handle multi-dimensional, heterogenous and incomplete data are seen as particularly promising. These developments are foreseen to provide an increasing uptake in precision medicine through development of personalized health services enabled by analysis of individual and aggregated multimodal data residing into patients' Electronic Medical Records (EMR). In this paper we provide a review of the combination of these technologies and their impact on the delivery of care.

We conducted and reported the review according to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) statement. Methods of the review process and eligibility criteria were established in and remained unchanged during the review. We have conducted a literature search using electronic databases Google Scholar, PubMed, IEEE, and ACM. The literature search was supplemented by manual retrieval of references contained in the articles. The literature search was conducted without time restrictions, using the following keywords: (EMR OR EHR OR PHR OR “Electronic Medical Records” OR “Electronic Health Records” OR “Personal Health Records”) and (“Deep Learning”). Majority of articles covered a period from 2010 (beginning of the popularization of Deep Learning) up to 2017. Our search strategy retrieved 1790 articles.

All identified titles and abstracts were screened for eligibility by 2 researchers (VO and LL). Articles were excluded based on two criteria: 1) They did not primarily focus on processing of health records using deep learning or 2) the work did not make use of Deep Learning methods. After this screening, there were 169 full text articles that were further assessed for relevance. After this screening, there were 36 articles that were considered for this review.

Out of 1790 articles identified in the initial search, there were 36 articles that fulfilled eligibility criteria and were selected for review. We have divided the studies into two categories namely i) disease modelling using EMR data and ii) an overview of Deep Learning architectures and their use with EMR data.

Broadly, EMR are the collection of healthcare related data captured during a patient visit to a hospital or clinic, including disease diagnoses, medication prescriptions, procedures, surgeries, and lab test results among others. One of the most powerful applications of EMR data is for the refinement of personalized medicine strategies [3], as it provides an unparalleled amount of populationsized, patient-level data that can be mined to directly inform clinical practice. In fact, Precision Medicine Initiative (https://www.whitehouse.gov/precisionmedicine ) was launched with a $215 million investment to enable clinical practice to move from a “one-size-fitsall approach” to more individualized care. When EMR data is combined with other modalities of data such as molecular features, biosensors, social determinants of health, and environmental exposures among others, a rich opportunity is created to study an individual patient’s disease in multiple dimensions and at multiple scales. Initiatives such as the personalized cancer therapy program [4] have utilized EMR resources to tailor treatment regimens to participating patients. Moreover, additional data parameters from EMR are used to augment current traditional Modified Early Warning Score (MEWS) algorithms, which “track-and-“trigger” warnings of patient condition deterioration based on six cardinal vital signs [5], especially when combined with continuous monitoring of patients [6][7][8][9][10] Furthermore, discoveries from data-driven, EMR-based research can lead to actionable findings, such as identifying medication adverse reactions [11] [12] and predicting future disease risk [13]. In addition to enabling the provision of clinical care, EMR are also a powerful tool to assist fundamental research [14]. There are countless examples of research studies that discovered patterns of disease susceptibility [15,16], comorbidity [17] [18], and trajectories [19,20] using EMR. Linking EMR to other -omics data types within a network biology framework [21] has helped to elucidate contributions of various r

This content is AI-processed based on open access ArXiv data.