Deep Representation Learning from Electronic Health Records
As Electronic Health Records (EHRs) become more ubiquitous, healthcare providers are beginning to appreciate the benefits of using this data for ‘secondary applications’, such as disease diagnosis and readmission prediction, by applying machine learning techniques (in particular, deep neural networks).
Principal Investigator:
Dr Edmond Zhang – Senior Data Scientist, Orion HealthResearch Team:
Reece Robinson – Principal Engineer, Orion Health Prof. Bernhard Pfahringer – Professor of Computer Science, University of WaikatoTimeline
Completed March 2018
As Electronic Health Records (EHRs) become more ubiquitous, healthcare providers are beginning to appreciate the benefits of using this data for ‘secondary applications’, such as disease diagnosis and readmission prediction, by applying machine learning techniques (in particular, deep neural networks). The difficulty with this approach is that data from EHRs is often inconsistent, episodic, exists in a multitude of formats, and is sometimes incomplete.
This research looks at previous efforts to apply deep neural networks to EHR data and provides evidence for adopting a new approach, which the researchers refer to as “bringing the model to the data”, as opposed to “transforming the data to the model”. These researchers found that by extracting key features from EHR data using multiple neural networks, and then combining these outputs, the accuracy of predicting diagnostic codes from the data is greatly improved.
This research project has been highlighted on the following webpages:
This research was presented by Dr Zhang at the 12th International Symposium on Medical Information and Communication Technology (ISMICT) 2018 conference in Sydney in March 2018.
Publication of the ISMICT conference proceedings is in progress:
E. Zhang, R. Robinson and B. Pfahringer. “Deep Holistic Representation Learning from EHR” in Proc. ISMICT, 2018
A follow-on project is now underway: A Deep Learning Platform for GP Referral Triage