April 22, 2019
A Vital Need for Better Healthcare Software: Introducing Vital Software Emergency EHR
Walkie-talkies, fax machines, DOS programs. It’s not the 80s. This is the state of healthcare software today, in 2019, in well funded hospitals around the world.
A few years ago just prior to moving overseas, I went to get my records for a surgery I had in 2013 to correct a deviated septum. I was told it would be $75 and they could not send the records over email, which they thought was insecure, but only by fax. Who the hell still has a fax number? I never got my records.
That same week, I had driven my car from California to Georgia to gift to a family member. As a courtesy, I had the oil changed at Jiffy Lube. They had all my service records years back, and on the other side of the country. There were better records - or at least more accessible records - for a $40 oil change than a $5000 surgery.
We all seem to have stories like this. Immunization records only proven by a sheet of paper in our mom’s attic. Patient portals that we’ve logged into exactly once because the interface was atrocious and the records were woefully incomplete.
Healthcare is backwards and we almost expect it. Or are used to it and expect no better.
As a patient, this led to a few bad experiences each year. But for doctors and nurses, the system, and particularly the software is soul sucking. As it turns out, the software they use day in and day out looks like it’s from Windows 98-era, typically has no mobile, no search, no AI and usually requires walking back to a desk to type notes.
In fact, it turns out that doctors spend two hours in the electronic health records (EHR) system for every hour they spend with patients, plus another 1.5 hours at home each night finishing their notes. If you could reverse that equation, you’d “double” the number of doctors in the world.
That is Vital’s ultimate goal: to double the effectiveness of healthcare through better software.
But I’m getting ahead of myself…
Vital was not my idea:
In 2016, I had just sold a business and finished a 3000 km walk across New Zealand. I had a bit of time on my hands for the first time since I started Mint in 2006. Consequently, I put the call out to family and friends: I’m available for machine learning and software help.
My sister jumped on the opportunity. She’s a professor of Epidemiology at Emory University, and studies kidney transplant statistics across populations. In the past, researchers had used numerical lab values and discrete categories to predict kidney transplant success or failure prior to surgery. Her idea was to use free-text from doctors’ notes to aid in prediction.
Having a background in natural language processing - first in categorization financial transactions at Mint.com, and later at understanding questions at Fountain - this was right in my sweet spot. Instead of simply solving the problem, however, I built classify.ai. The idea was a website for public and private datasets where you could upload a spreadsheet or CSV with mixed numeric and text data, pick a target column, and predict anything.
It worked. Doctors notes were predictive of kidney transplant success!
My brother-in-law, and now Vital Co-Founder, Dr. Justin Schrager, saw the results and wondered if they could be applied to the emergency room. He explained that he asks the same ten questions to patients over-and-over: Why are you here? What makes it better/worse? Any recent surgeries or hospital visits?
He thought: Could we ask patients these questions while they were waiting - sometimes for hours - and make a prediction about whether they would ultimately be admitted to stay overnight in the hospital?
I asked him for data, and he pointed me to the NHAMCS dataset, generated each year by the Center for Disease Control (CDC), and based on anonymized data from hundreds of emergency rooms around the country. Since I had just written a machine learning “platform” the first basic model was quick to produce. By analyzing free-text like “reason for visit” our admit-or-discharge classifier had an area-under-the-curve (AUC or C-Statistic) in the upper 80s. In other words, it worked well. We could predict, within about 10 minutes of arrival, how severe a patient was, how many resources they might require, and whether they would be going home that night.
In the next couple of weeks, I built a prototype mobile website patients could use. I’m not the best front-end developer - or designer - but it worked. We then set out to get feedback from other doctors and nurses in the emergency room.
My first visit behind-the-scenes visit to an ER was shocking. And I don’t mean in a graphic or horrific way. First, there are not the arterial blood-spurts everywhere that I had imagined. Half the patients seemed there for cold, flu, or urinary tract infections - the kind of thing that probably could or should be handled with a primary care doctor or urgent care clinic. Second, all of the software and communications technology I saw was far more antiquated than I could have possibly imagined.
Literally, walkie-talkies or pagers were being used to find doctors and nurses. To solve for lack of mobile or tablet accessibility, laptops were put on rolling carts, or COWs (computer-on-wheels). Doctors were so bogged down trying to enter 10 page notes - 9 of which were for billing or record keeping - that they hired “scribes”, assistants paid out of the doctors own salary to take dictation and fill out forms. Everything was, of course, Windows desktop based. These were the kind of places still using Internet Explorer 9 and Windows XP.
Then I saw another ER. It was the same. Then I talked to 10 doctors in New Zealand, to see if a small nation with national health IDs and a single payer was any better, or more interconnected. It was worse. Everything was cobbled together, dozens of snowflake configurations or custom software, at best communicated through a 1990s-era, pipe-delimited standard called HL7.
I saw an imaging system for storing X-rays and MRIs, built in the mid 2000s, restricted to a few tens of gigabytes of storage, that would halt frequently when storage was full necessitating manual backups. The vendor responsible for the system had since moved on and would only commit to bare minimum maintenance once a week. That meant there were times when hospital imaging could go down for six days at a time.
It turns out this kind of ad-hoc, snowflake system was par for the course.
Healthcare is Anti-Startup:
As Vital took shape as an idea, and eventually a company, I found out we were rare. There are few startups directly going after the software used to run hospitals, and the $1.3T in spending each year that goes through them.
Hospitals are the domain of huge corporations, namely: Epic, a 9000 person private company in Madison WI, and Cerner, a 20,000 person company with a $25 billion market cap, out of Kansas City. Installing and configuring their software can take 12-24 months, and cost a single hospital tens of millions of dollars. The cost to upgrade the Veterans Administration system, for example, is presently at $16 billion contract over a decade of roll-out.
All of these systems got their start in the 1970s. Meaning their code bases actually date back to then. Some of the “newer” EDIS (emergency department information systems) companies were started in the 1980s. Everything is in healthcare is decades old, in bad need of an upgrade.
So why aren’t there startups in a literal trillion dollar market? Healthcare is not the place for a minimum viable product. You cannot put something half-baked out there, and only then refine the user experience like you can with consumer internet software. It has to work from day 1, be HIPAA compliant, have millions in liability insurance, and pass some rigorous security reviews. It takes a year to sell into hospitals (hospitals are conservative, rightfully so, when patient safety is on the line). That takes more time, patience, and funding run-way then most startups can muster.
At the same time, health software has to be developed almost blindly. There’s no sandbox for healthcare development. There are no API toolkits to generate fake patients and messages, or observe the kinds of interactions you’ll see in a hospital. And because of HIPAA, you can’t use real data for development. Imagine if PayPal or Stripe had no way to test payments or refunds until the system went live? It would be a disaster.
Ultimately, we had to build a whole fake hospital system, generating realistic patients, lab results, and events before we could build our own system.
In healthcare, you’re also largely blind to the competition: you can’t really see their software. There are no free trials. Everything is a major million-dollar enterprise installation. Epic is known to threaten anyone who puts screenshots on the web. The competition is seen only over the shoulder of a doctor when doing user research, but never to be touched or controlled on one’s own. Imagine being a game developer who could never play a competitor’s game.
Healthcare is where startups face the highest level of inertia, complexity, and vested interests; some of the longest sales cycles; some of the most stringent requirements for privacy and compliance.
Naturally and naively, it’s where we’ve placed our bet.
What is Vital?
Vital is software for hospital emergency rooms and patients. Of course it has the AI we began with and more. We have algorithms for predicting patient severity, what labs for imaging might be needed, even whether a patient might become violent during a visit. We have algorithms to make patients safer in the waiting room, and lead them down the right path early for faster, better service. Algorithms to save hospitals millions in finding bottlenecks.
But Vital is also tackling the core workflow of the hospital. We are a total rethink of an EHR in the modern era: note taking on any device, voice interfaces, Google-like search within a patient’s records, or across patients. And so much more coming soon.
Healthcare is the most exciting place for a product person like me to play. It makes the fintech I worked on as Founder of Mint.com and VP of Product Innovation across Intuit look easy by comparison.
How can we show a clinician who has 30 seconds to spare a patient’s health history? Is someone in the ER simply having a bad night, or are they chronically ill? Is it best to show a history of recent visits? How much detail is needed before getting bogged down? Are abnormal lab results or vital signs most important? What if the patient has a condition where a result is expected to be a bit abnormal? Should we highlight what the last doctor concluded, or does that introduce too much bias too soon?
The product aspects of healthcare are endlessly fascinating. They require research, data, discussions with doctors and nurses, and intuition and vision for what healthcare can be. We’re committed to cracking the code. We’re committed to making healthcare more efficient. We’re for doctors and nurses. We’re for patients. We are Vital.