09 Jul Just Say No to a $100M EMR implementation
By Timothy Chou
One of the joys of teaching at Stanford is the quality of the students. A few years ago, I met Dr. Anthony Chang, who was coming back to school to earn a master’s degree in bioinformatics after having already earned his MBA, MD and MPH. It took him 3 1/2 years to complete, as he was still on-call as chief of pediatric cardiology at Children’s Hospital of Orange County, didn’t know how to program, and as a life-long bachelor had decided to adopt two children under the age of two.
Among his many accomplishments is starting the AIMed conference, which as the name implies focuses on AI in medicine. It’s held annually at the Ritz-Carlton Laguna Nigel in mid-December. Anthony attracts an amazing group of doctors who can both talk about pediatric endocrinology and graph databases. Since the conference is held near Christmas I often call Anthony “The Tree” and all the guest speakers are the ornaments. This year, I was asked to speak about the future of AI in medicine.
But, before we talk about the future let’s talk about the past. I was struck by one of the doctors talking about an $80M EMR application implementation. Having experience implementing enterprise ERP applications I was amazed at the number. It turns out this is not the high water mark with examples extending to north of $1B. Seriously?
Can an EMR application be the foundation for the future of AI in medicine? They are largely based on software from the 80s. If you were to think of cars it’s like trying to build an autonomous car using technology from a Model T parts bin. Furthermore, these applications were architected to serve billing applications, not patients. As a result there is no way to deliver personalized healthcare. After all, why should your bill look different than mine? And finally rather than being designed to collect and learn from exabytes of global data from healthcare machines they are built to archive notes from a set of isolated doctors who spend valuable time as typist. Maybe you should spend $10M to feed a billing application, but not $100M.
The future of AI in medicine depends on data. The more data, the more accuracy. Where is that data? Not in the EMR. It’s in the healthcare machines: the MRI, ultrasound, CT, immunoanalyzer, X-Ray, blood analysis, mass spectrometer, cytometer, and gene sequencer. Unfortunately the world of medicine lives in a disconnected state. My informal survey suggests that less than 10% of the healthcare machines in a hospital are connected. For those in computing, it looks like the 1990s when we had NetWare, Windows, Unix, and AS/400 machines that couldn’t talk to each other — until the Internet.
It turns out in 1994 when the Internet reached 1,000,000 connected machines the first generation of Internet companies like NetScape and eBay took off. And as the number of machines connected grew we ended up with even more innovations. Who could imagine NetFlix, Amazon, Google and Lyft before the Internet?
It turns out if you connected all the healthcare machines in all the children’s hospitals in the world we’d get to 500,000 machines, very close to the 1,000,000 machines that transformed the Internet. What would this enable? To begin with, we could get rid of CD-ROMs and the US Mail as the mechanism for doctors sharing data across the country. The Chexnet pneumonia digital assistant was developed with only 420 X-rays, what if they had 4,200,000 images. But, I’m sure this is just scratching the surface of what will be possible.
It’s clear the world of medicine where we pour knowledge into an individual’s head and let them, their machines and their patients operate in isolation is at an end. The challenges of connecting healthcare machines, collecting data and learning from that data are immense, but the benefit might actually change the world and it could cost a lot less than $100M.