Deep Learning for EHRs — Paper of the Week — June 27th

This week’s paper can be found here: Scalable and accurate deep learning with electronic health records

Note: The writing style of this post may be a little different from previous posts. I’m going through Massive Science’s Storytelling 101 course for scientists right now, which encourages scientists to talk about science as a narrative instead of as academic literature. This is my first attempt using this approach on actual science!

Welcome back to Paper of the Week!

Last week, we talked about an Android app for baby monitoring in the intensive care unit (NICU). We also read an opinion piece that I co-wrote for the Syracuse Post-Standard on why I think that AI should be incorporated into healthcare to reduce patient medical expenses. This week’s paper is a continuation of this AI in medicine trend! I don’t quite remember where I found this article, but it wasn’t Twitter (this time). I’ll update this post if it comes back to me.

For the majority of people in the US, healthcare costs have been increasing for a couple of years. It is more common than ever for personal debt to accumulate from expensive medical care, and millions of people are forced to file for bankruptcy. With drastic cuts being proposed at the federal and state levels to Medicare and Medicaid, it seems like this issue will continue to get worse in the coming years unless cost-saving measures are implemented to drive the cost of medical care back down.

While we typically look to legislators to solve these types of problems by making deals with insurance companies and drug manufacturers, another potential solution is in developing better preventative care. As of 2014, 86% of US healthcare spending goes to chronic conditions. For context, we spent $2.7 trillion on healthcare in 2014. If we could predict health trends that would lead to chronic conditions, we could treat those conditions before they become chronic, which would drive down healthcare costs. This is the economic incentive behind preventative care, which insurance companies have already supported through promotions such as reimbursing gym memberships if you go to the gym regularly, lowering rates if you quit smoking, and giving out free fitness trackers if you meet certain fitness milestones. What we don’t talk as much about is how medical technology could facilitate better preventative care.

Machine learning has already been proven to be useful in predicting medical outcomes for specific diagnostic cases, and there have been efforts to expand to more general prediction using electronic health records. The latter has had limited success. Doctors and hospitals were incentivized to use EHRs in 2011 when the Center for Medicare and Medicaid Services (which run Medicare and Medicaid at the federal level) offered reimbursements to those who did. Using EHRs standardized patient records across hospitals, where there previously was no one format that doctors and hospitals had to use. In other words, one doctor could use Post-Its, another could use Excel, and a third could use a subscription from a medical record-keeping company, but in 2011, everyone had to be on the same page. While EHRs have resulted in more consistency between patients in how diagnoses are recorded, doctors don’t always describe diagnoses in the same words. This leads to “messy” data when we try to use it for machine learning.

This paper, which comes from Google Mountain View, University of California San Francisco, University of Chicago Medicine, and Stanford University, aims to use deep learning to circumvent the messiness of EHR data. Deep learning typically means that the neural network used to develop a model for the data has many layers and can analyze many input variables. Specifically, the authors use deep learning on patient EHRs to predict patient morality (how likely a patient is to die soon), 30-day unplanned re-admission (whether the patient would have to come back within 30 days after being discharged), long length of stay (more than seven days), and final diagnosis (what the patient was diagnosed with). They predicted these outcomes at several different times, ranging from 24 hours before the patient had been admitted to 24 hours afterward. For the diagnosis and re-admission outcomes, they also predicted at the time of discharge.

Their predictions show the promise of using deep learning for messy data with a large number of variables. The predictions were best for patient mortality (ROC = 0.93–95) at 24 hours after admission, and diagnosis (ROC = 0.90) at discharge, and were fairly good for the length of stay (ROC = 0.85) at 24 hours after admission and re-admission (ROC = 0.76) at discharge. It is hard to compare the results of this approach directly to past studies, as most studies do not use the same models, medical data, or diagnosis definitions, however, the authors believe that the results are still better than past similar attempts to make predictions based on messy EHR data.

What does all this mean for reducing the cost of medical care and improving preventative care by predicting chronic conditions? Right now, it means that we are one step closer. The predictions here are based on training data with known outcomes, so we don’t know how well this would work on EHR data with unknown outcomes (i.e. if a patient hasn’t been diagnosed or re-admitted yet). The results are good in comparison to past studies but are not accurate enough to safely implement on real people. One of the most impactful takeaways (in my opinion) from this paper is that deep learning can be used on doctor’s notes — that is, written notes from doctors that do not conform to any EHR standard — to identify which factors in the notes are important in predicting the outcomes, in conjunction with more standardized EHR data. “Reading” doctors notes and replicating them has been a recent struggle of AI researchers, so this deep learning approach may contain some clues on how deep learning might be used to improve interpretation of written diagnoses.

As artificial intelligence development accelerates, implementation of deep learning for medical records in doctor’s offices and hospitals for preventative care is closer than ever, and certainly has the potential to reduce costs in the long term. We can’t afford these increasing medical costs forever.

Originally published at