An electronic health record holds information including patient demographics, historical admissions and discharges, lab test results, prescriptions, medical images and clinical notes. Doctors use this information to make individual care decisions and keep track of clinical information. Data scientists are starting to make headway in exploring the secondary use of EHR data, however current mathematical models to interpret health information usually only look at single areas of an EHR individually, using one data type to predict one outcome.

Our research, presented this week by Dr Edmond Zhang at the 12th International Symposium for Medical Information and Communication Technology in Sydney, has built a more comprehensive concept of many inputs to many outputs. It aims to apply a more holistic approach to EHR data and explore deep learning of an entire EHR to find patterns and combinations of risk factors. 

Healthcare is a prime example of a field that is full of data that is not being used to its full potential. With a huge variety of data being collected and stored, analysing this data and making it meaningful for both patients and clinicians is at the heart of precision medicine. The way the data is collected means that it won’t necessarily make sense to the end user, therefore research projects like this must involve some sort of translation of data. It is essential to ensure that the findings from data science find their way into the hands of those who will benefit. 

The research focuses on analysing all information on an EHR to give clinicians a more holistic view of a patient. The deep learning strategy proposed uses multiple neural networks to learn from the different input data and then translate it into features that can be combined to present a holistic view of the patient and thus allow it to be compared to other EHRs.

Ideally, once a person’s health data is collected and stored on their electronic health record, deep learning could be applied to EHRs to help clinicians find “patients like me”. This means if a patient comes into a clinic with a health issue, the GP can examine the EHR and subsequently search for other EHRs that present a range of similar information. For example, a patient could come in with an unusual set of symptoms, and the GP could ask the model to find what health conditions were prevalent amongst “similar” patients. Here, the GP would no longer need to define what aspects of similarity were important, the deep learning model can do that for them.

Rather than relying on experience and previous knowledge of certain cases, clinicians are starting to draw on information from entire populations to treat a person’s health. The healthcare industry is experiencing this massive transformation to digital health, and it is crucial to ensure people’s health data is being used in meaningful ways. Deep learning offers a new way to discover patterns in data, which enables healthcare professionals to move closer to personalised medicine.

There are many possibilities for how machine learning can be used in healthcare, and all of them depend on having sufficient data from multiple sources. This research will be used to implement Amadeus Intelligence, which applies machine learning to the large volume of health data that is ingested and managed by Orion Health’s Rhapsody and Amadeus platforms, to present clinicians with predictive analysis and help to identify “patients like me”. Amadeus Intelligence has the potential to leverage data insights from over 100 million patient records managed by Orion Health, and help clinicians to identify health issues more quickly and accurately through deep learning of EHRs.

*Presenting at the 12th International Symposium for Medical Information and Communication Technology (ISMICT) today in Sydney, Edmond Zhang will be sharing his research on deep learning on electronic health records (EHR) and how this can help clinicians find what we call “patients like me”.