Deep Learning for Analysis of Electronic Health Records (EHR)

Abstract

In current scenario, every medical equipment, clinical instrument, lab setup in healthcare centres and hospitals, is linked with digital devices which has brought about digital data explosion. Due to this, the amount of digital information generated and stored in Electronic Health Records (EHRs) has increased exponentially. Therefore, EHRs have become an area of booming research, as EHRs can provide a host of untouched possibilities which, the data contained in them, can bring about. EHRs have several classification schema and controlled vocabularies are present to record relevant medical information and events. Thus, harmonizing and analysing data among institutions and across terminologies is an ongoing field of research. Several clinical code representation forms have been proposed by various deep learning EHR systems that share themselves easily to cross institutional analysis and applications. EHR records have primary use in storing patient information such as patient medical history, progress, demography, diagnosis and medications. But researchers across the globe have invented secondary use of EHRs for several clinical and health informatics applications. Secondary usage of electronic health records (EHRs) promises to boost clinical research and result into better informed clinical decision making. Challenge in summarizing and representing patient data prevents widespread practice to predict the future of patients using EHRs. Simultaneously, over the span of time, the machine learning field has witnessed widespread advancements in the area of deep learning. The current research in healthcare informatics focusses on applying deep learning based on EHRs to clinical tasks. In this context, the deep learning techniques described here can be applied to various types of clinical applications such as extraction of information, representation learning, outcome prediction, phenotyping and de-identification. Several limitations of current research have been identified like model interpretability and heterogeneity of data.