# Deep EHR: A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record (EHR) Analysis
Shickel et al. survey the recent research on EHRs using Deep Learning(DL) techniques . They conclude that information extraction, representation learning, outcome prediction, phenotyping, and de-identification are the common clinical tasks performed.
The authors classify the results from their literature search into 5 sub-categories based on the applied DL techniques, followed by a section that explains these 5 sub-categories in detail.
Shickel et al. classify the papers by their DL techniques, as shown in Figure 1 below.
Papers are firstly divided into Supervised and Unsupervised categories, and further divided into sub-categories based on the applied DL techniques.
The authors identify 5 challenges in existing implementations.
- Data Heterogeneity, EHRs may contain written clinical notes, laboratory test results, continuous time-series of vital signs etc.
- Unified Representation, each system discussed by the authors has its own representation of patient data, an universal representation will benefit all researchers in the field.
- Patient De-identification, strict privacy laws and regulations require sensitive patient data to be protected, this prevents cross-institutional knowledge sharing.
- Benchmarks, the authors believe that the results reported from the research need to be verified using a standard set of benchmarking metrics, the community needs to introduce such a standard.
- Interpretability, the nature of DL is that the models trained cannot be naturally interpreted, there is no transparency for layers between the input and the output. However in the clinical domain, the transparency is of utmost importance.
Deep Learning on EHRs
Shickel, B., Tighe, P. J., Bihorac, A., & Rashidi, P. (2018). Deep EHR: A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record (EHR) Analysis. IEEE Journal of Biomedical and Health Informatics, 22(5), 1589–1604. https://doi.org/10.1109/JBHI.2017.2767063 ↩︎