Here’s an interesting article in Nature which was published in May 2018 that should be on the radar of any physician interested in Artificial Intelligence:
“Scalable and accurate deep learning with electronic health records”
This study, by analyzing the EHRs using Deep Learning, was able to significantly better predict:
1. In-Hospital Mortality.
2. Length of Hospital Stay.
3. Readmission Risk.
A couple of interesting points.
First the patient base the study used was not overly large – it looked at only 114,000 patients. However, the amount of information it extracted for its deep learning algorithm was huge – 46 billion data points. (That’s roughly 400,000 data points per patient!) .
Second, – it was atheoretical – which mean it wasn’t based on any apriori theory about what makes a high risk admission – it just looked at everything – height, weight, zip code, written words in the chart, every lab, every test. Then it powered through this data to predict what would happen for a given admission. The bottom line – it could say “this particular patient we can say with 90% certainty will die during this hospitalization” but, partly because it is based on an atheoretical model, it’s not able to give a reason why.
For practical reasons, this study only used the hospital data, but Google was involved, I suspect the next step (and least for someone like Google) is they will also want to use other input data for these models, including the patient’s online search history, phone location history, purchasing history, social media post history, credit history, etc.
Remember, the key to Deep Learning algorithms ( as opposed to expert AI learning systems, which does have some theoretical basis driven by expert opinion, such as physician input ) isn’t so much the type of data, but the amount of data.
That’s why I think companies such as Amazon, Apple, Microsoft, Google are going to change healthcare (for better or worse): Because they will see themselves as AI Companies doing some Healthcare, rather than Healthcare Companies using AI.
In some sense, these companies will consider your white count, your book purchase history, and your last Facebook post as equally valid (but differently weighted) inputs for such things as diagnosis and healthcare prediction analysis.
Anyway, this study is a good read because it gives one a sense of what Deep Learning is able to do, and what tools are just a couple of years away from being placed in a healthcare organization’s data toolkit.