The host had asked ChatGPT which decision in a patient's journey it would redesign and read the answer back to her; Orfanoudaki agreed with it, then said it described the destination without describing the difficulty. She had just finished walking through a before-and-after trial of a vertical-care pathway at Mayo Clinic Arizona. The line is her constraint on the whole field: a protocol has to be usable mid-shift, without new software, by a clinician who trusts it enough to actually follow it.
we ran it as a before and after study in the new Mayo Clinic Arizona emergency department between February and April 2024. We had five weeks under the old ad hoc practice where the vertical processing pathway was open, but people could use it the way they see fit. Then, we had three weeks of education and training for the physicians, the nurses, and the advanced care practitioners in the clinic. And then we had five weeks under the protocol. In total, that's about 11,000 patient visits once we restricted to physicians who were on staff across both periods. The total length of the stay fell by about 11 minutes, and that corresponds to a 4.2 reduction. The time from arrival to clinical disposition fell by about eight minutes. That's a 4.5% reduction. Now, those effects held up even when we controlled for other kinds of characteristics, including kind of like patient population, physician characteristics, procedures order, how busy the department was. And actually, even picking the period was not trivial because we had to kind of ensure that the pre-trial period and the post-trial period was kind of like equivalent and we didn't have like a surge in crisis. Like, for example, what happens when you are around Thanksgiving, around like the New Year's holidays, they see a surge in the type of cases. So, we had to avoid these time periods. And the result that I care about just as much is the one that didn't move, which is the 72-hour return rates with and without admissions. Those were unchanged. Now, that meant for us that we were not actually buying speed by discharging people that we shouldn't have. And that was really important because quality of care is either something that you don't want to jeopardize or something that you want to strictly improve. Efficiency improvements cannot come at the expense of quality of care. The other anecdote was the qualitative feedback that we received, which was equally interesting. The clinical team pointed out that the protocol is a resource neutral in terms of physician hours, which matters enormously for adoption. They did find moving between the vertical room and the main department to be part of an interruption in the workflow, but not one that kept them from seeing higher acuity patients. In fact, they noticed the opposite, like that low-acuity patients who arrived early in the day were no longer occupying rooms, and so sicker patients were getting into the bed sooner. That was one of the core hypotheses of our flow model. And it was really gratifying to hear kind of coming back to us that, like, actually from the people doing. The work. The other anecdote from the implementation was the operations team of the Central Mayo Clinic called halfway through the implementation trial and said, Well, we noticed that your guys are going faster. Can you tell us what's going on? Because we are seeing an improvement and we don't know where this is coming because we haven't invested more money into the clinic.
That's amazing. What a confirmation of the success of your work. Wow. All right. Now it's time for our Ask AI segment. Before our conversation, I asked ChatGPT in this case to consider the challenge at the heart of your research and answer this question for me. So if you could redesign one decision made during a patient's journey through the emergency department using analytics, which decision would you focus on and why? And the response ChatGPT came back with was: I'd focus on the earliest routing decisions, determining what type of care environment a patient actually needs rather than automatically treating an emergency department bed as the default destination. So what do you think? How did AI do? And based on what you've learned through your research, is there anything you would add or change?
Honestly, that's a very good answer. And it's the same decision we chose to focus on. So credit where it's due. What I add is that it describes the destination without describing the difficulty. And framing it as what type of care environment does this patient need makes it sound like a pure prediction problem. And the single biggest thing we learned is that it isn't. The right routing decision for an identical patient may be different at three in the morning in a quiet department than it is during the midday search. Prediction tells you who could be treated vertically, but it's actually the operational model that tells you when you should actually do it and how much classification risk is worth taking given how congested you are. If you get that wrong in either direction, you lose the benefit of whatever patient streaming protocol you're implementing. And routing too few patients could lead to a pathway that sits empty or routing too many patients and you're actually sending people back to the waiting room again and again, or you're creating a disruption between who gets allocated the bed and who doesn't. So I would say this is one of the core questions that matters. The other thing I would add is the constraint that the answer requires. Like you need a design that has to be usable by a clinician with shift without the screen, without necessarily new software, and it has to be something they trust enough to follow. An optimal decision that nobody adopts is actually worth exactly zero minutes. So having worked in healthcare for multiple years now, that has become a big component of how I think about research and how I think about kind of not only the modeling part, but how do we translate that in practice?
So back to something we discussed a little bit earlier. An 11 minute reduction in an emergency room visit time might sound relatively modest, but when you multiply that across thousands of patients, I imagine it becomes pretty significant. What does an 11 minute improvement actually mean operationally for an emergency department?
And that's an excellent point. And, you know, kind of when I first saw the result, briefly, I thought of the same thing. But let me actually do the calculation with you. So for a medium-sized department, seeing about 40,000 patients a year, right, an 11-minute reduction per visit adds up to roughly 6,800 bed hours recovered annually. Now, that's enough capacity to treat around 2,000 additional patients, which could translate to something in the region of 3 million in additional reimbursement with no new beds, no new staff, and no new technology. Now, the other part I'd like to highlight kind of when doing this kind of calculation is the queuing, right? So it's less obvious, but more powerful. Systems operating close to capacity behave highly non-linearly. When you're at 95% utilization, a small reduction in how long each patient occupies a resource produces a disproportionately large reduction in how long everyone else waits. So freeing a room at 11 in the morning doesn't just help that one patient that you have at that time in the clinic. It changes the cue that builds through the afternoon. And so I would like to point that this 11-minute reduction was an average across every patient in the department, including the ones the protocol never touches. So for the patients actually routed through the vertical pathway, the difference is not that 11 minute. That 11 minute applies to everybody and that's how I think we should be viewing these kinds of interventions because even though you apply them for part of the population, you need to kind of like think about the effect overall.
So I think there's a larger idea here that I really love and I think it's kind of been the theme of our conversation today that sometimes innovation isn't about adding more resources or more technology. It's making better decisions with what you already have. What do you hope healthcare leaders and perhaps even leaders outside healthcare take away from your research?