Using the Power of Deep Learning for Clinical Risk Prediction

02 March 2020

Portrait of Dr Yajie Zhu

by Dr Yajie Zhu
Machine Learning Scientist

Yajie Zhu is a Machine Learning Scientist on the Oxford Martin Programme on Deep Medicine and at The George Institute for Global Health, UK at the University of Oxford.

Adobe Stock 276702008
What is Deep Learning and why is it useful in medicine?

Deep Learning is a subfield of Machine Learning that uses large quantities of data to analyse rich but complex information to identify patterns, trends and correlations. Using Deep Learning with electronic health records can help doctors and other health professionals make better decisions about which societal and individual interventions would help patients most. It can also identify patterns of risk that would allow doctors to intervene early to prevent negative outcomes like hospitalisations or emergency room visits.

This technology has huge potential to support patient health at an individual and societal basis and to decrease costs for health systems by increasing preventative care and decreasing costly hospital stays and emergency treatment.

Why isn't Deep Learning being used more in medicine already?

The digitalisation of health records is a recent development. Since these records have become available, we've seen a large number of academic papers combining these data with Deep Learning to answer many different questions around personal health and risk identification.

This is a promising step, but the variability of the methods used, and the subsequent results have made it difficult to compare and contrast Deep Learning models. This uncertainty could be a cause of the limited use of these systems in clinical settings.

How does your recent research hope to help overcome this challenge?

In a recent study we applied and compared different Deep Learning methods and designs to the large UK electronic health records (UK EHR) dataset to understand which kinds of Deep Learning methods work most effectively and develop practical guidance on how to identify a “good” Deep Learning approach.

We applied different Deep Learning methods to the Clinical Practice Research Datalink – a unique UK resource representing one of the world’s largest healthcare databases - by capturing information from primary care settings as well as data from other national databases like hospitalisations and vital status. We asked each model to learn from the data what predicted emergency admissions and heart failure. From the results, we were able to identify the types of Deep Learning systems that work best with healthcare data and can provide more certain results.

What did the study find?

We identified four key types of Deep Learning networks that had been used in healthcare research papers. We found that a type of Deep Learning method called a Recurrent Neural Network (RNN) performed best out of the four. In RNN, the input data are organised as a sequence and processed by the model in a sequential manner - so it analyses the data from the first recorded symptoms or vital readings to the last recorded ones to identify patterns across a large number of patient’s health history. We believe that this type of analysis performs well because symptoms and other medical data are, by nature, sequential.

However, we also found some of the difficulties that all types of Deep Learning methods are facing with these large medical datasets. One of the key challenges was that large datasets, like the UK EHR, are highly imbalanced, which means that it’s typical to see data from many more people without a specific diagnosis than those with it, but it’s the second, smaller group that we want the model to be able to identify and learn risk patterns about.

Health data also throw up their own unique challenges like irregular patient visits, the large number of medical concepts (e.g. diagnoses, measurements, prescriptions and procedures), the diverse nature of the data and high rate of non-numerical information. Effective capturing of all these details, and applying mathematical operations to them for computational analysis, are therefore key to the success of any Deep Learning method in healthcare.

What does this mean for the future of Deep Learning in medicine?

The increasing use of Deep Learning in medicine will no doubt continue. However, before we see a world in which Deep Learning findings are central to clinical management of health risks, researchers need to address two important issues:

  • Firstly, model uncertainty, which can provide a confidence score for each prediction. For example, when the model is used for a new patient who is not similar to any patient the model has previously seen, the prediction should be uncertain and produce a low confidence score, which can tell the doctor to assign less credit to the predictions when making decisions;
  • Secondly, model interpretability, which explains why the model gives a certain prediction for a certain patient. For example, the model could show a small number of important input variables that contribute most to the prediction in the context of a certain patient, so the doctor is able to understand why the model produced such a prediction and draw on it as a reliable source to inform their clinical judgement.

In addressing these issues and increasing the transparency around how Deep Learning-derived recommendations are communicated, the confidence in Deep Learning models in a clinical setting would be greatly strengthened. Higher confidence from doctors would lead Deep Learning methods to become more applicable to patients and therefore more widely used.

We hope that other Deep Learning researchers in the healthcare space will adopt these recommendations as this technology, if properly and widely applied, could make a huge difference to the health of individuals and the effectiveness of healthcare systems overall.

This opinion piece reflects the views of the author, and does not necessarily reflect the position of the Oxford Martin School or the University of Oxford. Any errors or omissions are those of the author.