Reducing Clinical Variation in Medicine with Artificial Intelligence
Let’s start with acknowledging that Medical Errors are now the 3rd leading cause of death in the USA, having surpassed COPD. In addition, studies have shown us that 30 percent of all medical care costs are unnecessary. While clinical variation is not the only reason for these two issues, it is a significant contributor. If eliminating medical errors and unnecessary medical care do not form the basis for a “Just Cause Initiative”, I don’t know what does.
As someone who has practiced medicine for 37 years, the last seven as a hospital CMIO, I can’t help but feel responsible for both these issues. Couple this with the fact that most of us are trying to cope with the movement of medicine from “volume to value", and many of us have a foot in each world and are trying not to fail! Some of us are also preparing for risk sharing and if we are going to do that, we must tackle the issue of Clinical Variation, or prepare to fail! In this article, I will share how we, at Flagler Hospital, are reducing clinical variation to enhance patient safety, improve our quality, while reducing the cost of the care we provide.
Hospitals have been trying to address this for a very long time, usually with energetic clinical staff. Optimistic administrations have done the best they could, with chart reviews and data extractions. But these efforts are biased by the choices of what they decided to observe, and the results are usually months old by the time the data is analyzed.
Over the past few years, we have had enormous advances in the hardware and software used to store our patient data. Advances in Artificial Intelligence have made it possible for us to look at all the data, allowing it tell us what we have done well and where we have fallen short. It can now do this in a matter of minutes, not months or years.
At Flagler Hospital we have chosen to use AI software developed by Ayasdi, which uses a particular branch of mathematics called Topology. This software allows us to look at all our patient data chronologically, creating a separately flagged event for every order, medication administration, lab result, and all patient care tasks. In other words, everything that happened to the patient. It then groups the patients into “treatment groups” based on their similar care. Next, it shows us the Direct Variable Cost, length of stay, and co-morbid conditions, allowing us to examine and identify any statistical differences in these groups. Ayasdi also allows us to create our own “User Defined Variables” for additional, customized analysis. With these, we can further analyze our treatment groups based on factors such as age, gender, lab value result or whether the patient came from home, a SNF or another medical facility.
We select the best treatment group, known as our “Goldilocks Group”. It demonstrates the best combination of lower direct variable cost, length of stay, readmission rate, and mortality. We use Ayasdi’s unsupervised learning algorhithm4to generate a CarePath from this treatment group. It tells us, “if you want future patients to be in this group, here are the things you need to do, and the timing and sequence with which to do them”. We then edit the CarePath and deploy it, use it to make changes to corresponding admission and treatment order sets and then monitor adherence of our physician to the CarePath.
While all of this is music to our CFO’s ears, it is meaningless unless we can operationalize it. We, at Flagler, have been fortunate. Seven years ago, we formed our “PIT Crew” (Physician IT Crew). It is comprised of 20 physicians, one or two from every department in the hospital, who were given authority by the hospital board to make decisions concerning EMR workflow, order sets and orders. They meet regularly and review all things “EMR” and can make decisions or recommendations to individual departments or the medical staff in general. This team has been vital to our EMR success and forms the “cornerstone” of our project to reduce clinical variation.
Two years ago, when I discovered Ayasdi, I brought a plan to my CEO who gave me the go-ahead for the Pilot. The PIT Crew was brought in at the very beginning of the process, as were several Informatics staff members with excellent SQL skills. Here is how we have operationalized our use of Ayasdi:
1. We created 2,500 lines of SQL code to extract patient data from our EMR (Allscripts), our Analytics Platform (CPM), our surgical system (SIS), our Enterprise Data Warehouse and our financial system. These queries were parameterized so we can easily move from one diagnosis to another
2. Once the data is extracted, one of our DBA’s uploads the data to Ayasdi
3. We perform several rounds of semantic and syntactic validation
4. We run the “unsupervised learning” algorithm of Ayasdi to generate the treatment groups
5. We compare the groups statistically using P-values and K-S test (Kolmogorov-Smirnov)
6. We present the groups to the PIT Crew who review the data and decide on the “Goldilocks” treatment group
7. Ayasdi is then used to generate the CarePath from the Goldilocks treatment group
8. A PIT Crew member from the appropriate medical specialty and I edit the CarePath
9. Admission and treatment order sets are changed to comply with the CarePath
10. Education goes out to the medical staff
11. Order Sets and CarePath are deployed
12. Monitoring begins with the physician adherence report generated by Ayasdi
13. Quarterly reports go the Administration, the Hospital Board, the ACO and physician staff showing changes in Direct Variable Cost, Length of Stay, Readmissions and Mortality
We completed our first CarePath, involving Pneumonia patients from 1/1/14 to 6/30/18, in nine weeks. We completed our second, Sepsis, in two weeks. We went on to complete COPD, Heart Failure and are now on our 5th, Total Hip. We plan to roll out 18 CarePaths in 18 months.
We went live with our first CarePath, Pneumonia in August of 2018. Our results show a reduction in Direct Variable Cost of $1,073 per patient, a reduction in Length of Stay by 1.38 days and a reduction in Mortality from 1.5 percent to 0.0 percent. With our efforts, we appear to be on the right path to do our part in reducing clinical variation, as well as reducing unnecessary costs and eliminating medical errors. I believe that all hospitals, even sole community hospitals like ours, can do this well. Together we can improve the way healthcare is delivered for our patients, our communities and our nation!
Clinical Informatics and the Promise of Advanced Technologies
No Wrong Door: Connecting the Dots in Health and Human Services
Cyber security- A Proactive Approach to Securing Information
Technology to Proactively Run a Healthcare Organization
By Debra Jensen, CIO, Charlotte Russe
By Phil Jordan, CIO, Telefonica
By Alberto Ruocco, CIO, American Electric Power
By Sven Gerjets, SVP-IT, DIRECTV
By Adrian Mebane, VP-Global Ethics & Compliance, The Hershey...
By Mike Fitton, Wireless Business Unit Director, Altera
By Jim Kaskade, VP and GM, Big Data & Analytics, CSC
By Graham Welch, Director-Cisco Security, Cisco
By Michael Watkins, Senior Product Director, Global Knowledge
By Nelson C. Vincent, EdD, VP for IT and CIO, University of...
By Sharon Gietl, VP-IT & CIO, The Doe Run Company
By Arnold Leap, CIO, 1-800-Flowers.com
By Gary Barlet, CIO, USPS OIG
By Mike Dieter, CTO, Transplace
By Bill Schimikowski, VP, Customer Experience, Fidelity...
By Kevin Kometer, CIO, CME Group
By John Landwehr, Public Sector CTO, Adobe
By Marc Probst, CIO & VP, Intermountain Healthcare
By Charles Koontz, President & CEO, GE Healthcare IT & Chief...
By Jeff Bauserman, VP-Information Systems & Technology,...