• Politics
  • Diversity, equity and inclusion
  • Financial Decision Making
  • Telehealth
  • Patient Experience
  • Leadership
  • Point of Care Tools
  • Product Solutions
  • Management
  • Technology
  • Healthcare Transformation
  • Data + Technology
  • Safer Hospitals
  • Business
  • Providers in Practice
  • Mergers and Acquisitions
  • AI & Data Analytics
  • Cybersecurity
  • Interoperability & EHRs
  • Medical Devices
  • Pop Health Tech
  • Precision Medicine
  • Virtual Care
  • Health equity

The Hype and the Upside of AI in Healthcare: A Conversation

Article

Janae Sharp talks machine learning with Don Woodlock.

Artificial intelligence (AI) in healthcare is long on aspiration and short on return. Healthcare AI is predicted to reach a $6.6 billion industry by 2021. This week I spoke with Don Woodlock, head of the HealthShare business unit at InterSystems, about the future of AI - and especially to separate the hype from the upside when it comes to its adoption.

My supposition: whether there are use cases for AI in healthcare, or whether this is just something people like to sink money into at the moment? Where is the real value?

Here’s an excerpt from our conversation, where Don Woodlock answered some of my questions.

Don Woodlock: When we talk about AI, there are a lot of people in the field who are working on “out doctoring the doctor.” They want to build an algorithm that is smarter than physicians. I don’t think that is the best use of our time. AI does a good job at “knowledge work.” Being a physician is one of the most complex knowledge worker jobs on the planet. Maybe from an academic standpoint, the endeavor is a good one - but from a practical point, aiming at the most critical life and death jobs seems like it isn’t the best way forward. There are better use cases for AI in healthcare that are still in the realm of knowledge work, but are more achievable. They are less important than life and death. Why are we trying to sell physicians on an endeavor that will get rid of them, especially since that endeavor falls so far short of the hype?

Janae Sharp: Where do you see a practical return on investment in artificial intelligence?

D.W.: One place where there is a great use case for AI is something that the ONC is taking on right now, patient matching. With patient matching, you get records from multiple organizations, which sometimes means you get slightly different names and addresses. We have a master patient index product that matches them. The product is really good. It is a regular old algorithm that tries to match things in a specific way. We have been using machine learning to improve the information. We give a yes or a maybe -so some human works through the matches that are a “maybe.” The input loop is from watching the manual matches, then the learning model predicts which people match. Since they have encountered so many records, we can reduce the work list size by 40% with no errors in terms of patients matching. One of our customers had 18 people working on the patient matching list. It is a lot of effort to do patient matching manually. So, that application of AI represents a significant time and cost savings.

J.S.: That could have a huge impact on the rising administrative costs. So many people that look at healthcare say the administrative waste is burying us.

D.W.: Yes, generally the administrative costs of healthcare are enormous. Thirty percent of our total cost is too much to pay for manual work such as claims adjudication or improving referrals. When you take time consuming manual work and reduce it by 40%--that is a huge deal. So AI has important potential and current return on investment for reducing administrative burden.

J.S.: So that is cost reduction. There is a category of artificial intelligence that is focused on predicting the future, too. Are forward predictive algorithms able to show a real return on investment? Some of it seems like common sense. For instance, it’s fairly easy to predict that if a patient doesn’t have shelter, they won’t recover from an illness. Can AI help there?

D.W.: One of the most important use cases for “predicting the future” is something that avoids costs and saves lives. Hospitals can have financial penalties for emergency room readmission, but knowing which patients are at risk of readmission can also help save their lives.

If you are running a hospital and you are charged with discharge - if that patient comes back in 30 days, you won’t get paid. But who will come back? Manually, they use a LACE Score. We can look at more factors and understand their impact more completely, and in some cases predict who is going to get diabetes or even attempt suicide in the next 12 months.

J.S.: Can you tell us more about the LACE Score? I'm not familiar with that acronym.

D.W.: Right now readmission to the hospital after a stay is a challenge. We want to help people when they visit the emergency department, and help them get care. The cost to intervene and transition patients after an emergency room visit is lower than the cost of having them come back. Typically a few things are considered when predicting if patients will heal: L, for Length of stay; A, for acuity of admission; C, for comorbidities; and E, for emergency department visits in the last 6 months. That’s how we get the acronym “LACE."

J.S.: How does the LACE Score work?

D.W.: You have a total number of points, and patients that have a higher score are more likely to be admitted. Everyone has a score, and discharge people focus on people with higher scores to look at the highest risk. We have done work that involved ignoring this score, but using more data, to predict admission. Machine learning does a much better job at predicting which patients will be readmitted than the LACE score. It can take more things into account. Things like medication, allergies, abnormal lab values, clinical notes, zip code, gender. It can look at things like social history, like smoking or alcoholism. There are 2 advantages to using AI for this: One, you can pump more data into it--this would take too much time for a nurse. LACE are the top 4 things, but ML can be more accurate. Two, with LACE, the factors are simplified: If you stay 1 day, you get 1 point, etc. Machine learning can figure out the actual impact factor. It can handle a more complex algorithm If you want to look at the top 25 from a risk point of view, if you use LACE, you will find some of them, but if you use machine learning you will find more. I use the factors around cost of readmission, cost of intervention (like telehealth or a visit), chance that intervention will help, etc.

J.S.: So how much money could this approach save?

D.W.: For a large customer on an annual basis, you would save $5.5 million by using this model and focusing on the right patients, rather than using the old model. Healthcare resources are scarce, so you must focus your time and attention on the right folks. Artificial intelligence is great at processing large amounts of data for proactive interventions in health.

J.S.: I know CareCognitics uses algorithms to predict future risks and as a result to have less costly interventions. They also report better motivation on the part of patients to come to appointments.

D.W.: Yes - administrative predictions are great, and machine learning can do a good job of predicting those things. If you are running a billing office, you want to know which claims aren’t going to be paid by the insurance company before two months. Machine learning can also do a good job of predicting those “no shows” - that’s a huge area where artificial intelligence can save real money. You want to know which patients aren’t going to show up today. Fourteen percent of your patients on any given day are not going to show up. That is standard. That is the normal course of a physician’s life - so, you overbook. Many organizations utilize reminders and texts. But even with those in place, the rate of no shows is above zero, so it is smart to overbook to account for the realities of life.

J.S.: Yes -I have missed appointments before, if I’m sick or have a flat tire or forget.

D.W.: Yes - that is human, so the strategy that health care offices or hospitals use to account for life varies. One strategy is overbook randomly. If you have 30 patients, then four won’t show up, so you overbook by four people RANDOMLY or evenly throughout the day. With machine learning you learn to overbook the slots that have the highest likelihood of not showing. You can learn about specific patients and past behavior. You can also look at people like them, including people who live the same distance from the office, or people in their families. This model is much more accurate at predicting no shows than other techniques--in one study they looked at the different methods, but they also looked at physician idle time and patient wait time.

My conversation with Don Woodlock excited me about the potential for artificial intelligence to reduce the cost of care and reduce wasted time.

Separate from the hype of futuristic concepts like robot doctors, it remains important to look at what is working in AI in healthcare. Using prediction to improve efficiency is a step in the right direction. We can use machine learning to reduce administrative waste and efficient operations based on a better knowledge of the future and of patient behavior.

Get the best insights in digital health directly to your inbox.

Related

Mount Sinai and Hasso Plattner Institute Launch Digital Health Institute

UVA Launches Evaluation of Clinical Decision Support Tools

Mount Sinai & LabCorp to Launch Digital Pathology Center

Related Videos
Image: Ron Southwick, Chief Healthcare Executive
George Van Antwerp, MBA
Edmondo Robinson, MD
Craig Newman
Related Content
© 2024 MJH Life Sciences

All rights reserved.