Optimized Infusion Scheduling: What Does It Take to Do This? Webinar Transcript
SPEAKER: Good morning. Thanks for making the time. Let me quickly describe how this webinar fits in in the series of four. Some of you may have attended the first webinar where we focused a lot on, why is it that the EHR cannot do this? And why is the problem much harder than it seems? With each successive webinar, we will dig into the next topic. So today’s webinar will focus mostly on, what does it take to pull off this optimized infusion schedule? However, I will quickly cover the other topics as well. So for those of you who didn’t attend the last webinar, you’ll get a quick sneak peek. And you can download the last webinar. And we’ll repeat the same thing for the next two. OK? So what does it take to do this? To begin with, just a quick introduction in two minutes. Who are we and what do we do? LeanTaaS is a Silicon. Valley company. We’ve got now approaching 100 people with four skill sets, PhDs in Mathematics,. Software Engineers, Product Managers, and. Operations Experts, together build optimization algorithms for healthcare.
We initially have focused on unlocking capacity. So for us, a lot of magic happens when you unlock capacity. Patient access improves. It takes fewer days or weeks to get an appointment on the calendar. The wait times for patients when they actually show up to their appointments goes down by a lot. Operating costs go down because you’re better managing your facilities and your staffing to the patient load. Capital costs go down because you can defer the need to expand facilities. And finally, to the extent that more patient visits result in more revenues, it helps on the top line as well. We have proven this out. We have two commercial products right now, Infusion that we’ve spent time on, and Operating Rooms. We’re in the middle of building the next wave of products. Clinics is actually in beta trials right now, with both Memorial. Sloan Kettering in New York and MD Anderson in Houston, 20 oncologists each.
We are in the middle of recruiting eight to 10 more clinics, oncology clinics, to participate in the next wave of beta trials over the next three to six months. The product will go commercial early next year. Similarly with Imaging, at MD Anderson, we’ve done CT clusters. Inpatient beds that you see, health and labs at Emory in the Southeast, where we were able to take the wait time down from 90 minutes to 15. So that just gives you a sense of the kind of products we build. We work at about 50 health systems around the country. This is a sampling of them. And they’ve ranged from academic medical centers like Stanford and Emory, and Duke, and. UPMC, and Penn, to iconic single-service line institutions like MD Anderson and Sloan Kettering that are focused entirely on cancer, to standalone community cancer centers like Stark, as well as standalone hospitals like Boca Raton. Infusion is a real problem.
And the pace at which these centers are signing on, it seems to indicate it. At this point, there are 111 infusion centers across these 50 customers that collectively have to 2,750 chairs. Roughly about 7,500 oncology infusion treatments per day schedule on the platform and optimize on the platform. So that’s the pace at which it’s getting smarter and learning on a continuous basis. So we have dived into these next two in great detail last time. I’ll cover them in three and two minutes, just as a quick refresher. Why can’t the EHR do this? The entire mathematical foundations of every EHR out there is simply not robust to do scheduling optimization. Why is that? They all took a shortcut when it came to scheduling appointments. They said, the right way to do this is to treat appointments like a reservation on a resource. So an imaging appointment is a reservation on an MRI machine from 8:00 to 9:00 in the morning. An infusion appointment is a reservation on a chair. A doctor’s appointment is a reservation on the doctor’s calendar. So they treated everything like a reservation. And so, as a result, wherever you go in a health system, you get this grid-based approach where chairs are laid out across the top and times of the day down the left.
And as appointments are made, this grid gets filled in. This is how tennis courts and spa treatments are scheduled. And it works in those cases because they are deterministic. Meaning, the start time and the end time of the appointment is perfectly known at the time of making the appointment. That’s why you can color it and then do a grid. Infusion appointments, the start and the end time are stochastic or random. It just simply does not work to schedule it into the grid. This is why the grid looks great the night before. And it never plays out the way you hoped it would. The second reason the math doesn’t work is every health system in the country, because their scheduled on EHRs, schedule’s on a first come, first serve basis. And this looks very fair and egalitarian, and patient-centric, and polite, et cetera, et cetera. It’s also mathematically completely incorrect. Because if you thought of the need to balance arrivals and departures, then the duration matters. So, just so you get the right number of incoming and the right number of outgoing at every point in time to keep your utilization consistent, you’ve got to be able to steer appointments into the right bucket based on an optimization algorithm, and not just leave it to chance on a first come, first serve basis. The third reason the math doesn’t work is healthcare appointments invariably are connecting flights.
There’s a lab followed by a doctor’s visit, followed by an infusion treatment, followed by radiation oncology. When you stringing together three or four things, then the on-time performance of early legs of the flight has to be high, otherwise, connecting flights don’t work. And the second thing that needs to happen is, you have to optimize the system as a whole. Imagine, booking John. Smith’s travel by saying, he’s going from Chicago to. Atlanta, then Atlanta to Miami, and trying to determine when the Atlanta-Miami flight should take off. You have to optimize the system by looking at all the incoming to Atlanta, all the outgoing to the various destinations, and then figure out that right connecting windows so that the most passengers have the most convenient connect. So this is the reason the mathematics doesn’t quite work out. In two minutes, to just cover what makes infusion scheduling particularly hard. Because on the surface, one would think the scheduling should be simple. It’s a patient sitting in a chair for some amount of time, from one hour to nine hours.
And we just got to sequence the chairs right. So how hard can it be? Here’s how hard it is. If you’ve got five types of durations, a one-hour, a two-hour, a three-to-five, a six-to-eight, and a nine plus, and you’ve got roughly 70 treatments a day, this is a small to mid-sized infusion center, 20 to 25 chairs. And you’re seeing 70 patients of this duration, it makes 20 of them need a one-hour treatment, 20 of them need a two-hour treatment, and so on. You’ve got 256 possible slots, which means– imagine, you are willing to start scheduling at 7:00, 7:05, 7:10, 7:20, 7:30, and so on. And you could seat four patients at a time, 64 times 4 is 256. So you’ve got 256 possible start times. The number of ways in which you could have scheduled and sequenced the patients is a number with more than 100 zeros behind it. So expecting schedulers to just look at it and decide that John Smith should get the 9:00 appointment is just being hopeful without any real probability of success. And to put this in context, winning the mega millions is a number with only six or nine zeros behind it. So that’s the probability you’re facing. What makes it even harder is from all of these solutions, there are many constraints that need to be matched.
You’ve got to get the volume and mix right. You’ve got to be able to figure out how many nurses you’ve got, and who’s available, and when. You’ve got to figure out when your chairs are available or not. And finally, there’s a lot of variability. No matter how well you plan, some variability could have been predicted and is expected. Some variability is just variability. It was unexpected, unpredictable, just happened. And so, all of these constraints have to be forced into consideration out of the solution space that has more than 100 zeros behind it. So as a result of all of this, the way virtually every infusion center operates is delivering a profile like this. When you think of each patient as a Tetris block in length, they come in at the times of their appointments, they get seated in the part in the chair with the right nurse at the right time as best as you can. And as this unfolds, the game of Tetris is being lost. And losing the game of Tetris has two big consequences. First is, it’s a very tough question on how do you deal with nurse staffing.
Because if you had enough nurses to cover your peak, by definition, you are overstaffed before and after. Because the peak only lasts three or four hours, a nursing shift is eight or 10 hours. If your staff were the average, then right when you needed it most, you’re shorthanded on nurses, which really hurts. So whichever choice you make, you’ve got a bad outcome. But the bigger consequence is, if you had 1,000 chairs, this would not be a problem. It only becomes a problem if the peak starts to approach your chair capacity. When that happens, a system becomes mathematically unstable. And what a mathematically unstable system means is the following. Think of the freeways at rush hour. There’s no capacity to add any more cars to it. So it’s now entered a mathematically unstable state, which means every metric will go into the tank. A 20 minute drive will take 60 minutes. A fender bender that should take 10 minutes to clear will take an hour to clear. And a fender bender that should have delayed 10 people will delay 10,000 people. So every metric became two times worse, five times worse, 10 times worse, 1,000 times worse. Infusion is a series of fender benders waiting to happen. The clinical will run late. The pharmacy will back up. The lab will back up. A nurse will call in sick. A patient will show up late. A patient will react badly. If any of those happened here, it’s the fender bender at midnight. The two people exchanged information, and life goes back to normal in 10 minutes. If the fender bender happens anywhere here, it’s like the big-rig crash on the middle lane of a freeway at rush hour. Everyone’s stuck for many hours. By the time they’re digged out of the ditch, it will be “Groundhog Day,” and it will happen all over again tomorrow. So this is the mathematical consequence of why infusion scheduling needs to be optimized. The way it manifest then is most infusion centers see the following three things. Patients wait a long time, particularly in the middle of the day. The chair profile looks like that mountain we described. Often, it goes above chair capacity. What that means is the infusion suite is full, and people are sitting in the waiting room. So the waiting room chairs are acting like “State Zero” infusion chairs. And nurses tend to miss their lunch.
And in many unionized environments or environments where they really try and address this problem, the way they solve it is get extra coverage in the middle of the day by borrowing from a floor pool, or reassigning nurses, or having the nurse manager take the floor. All of those are just ways of throwing more labor at the problem rather than solving the underlying need to optimize it. OK? So this gets us through all of the background stuff. Let’s just jump into, what does it take to optimize an infusion schedule? So let’s start with what doesn’t work. Chair-based scheduling where what people do is lay out the chairs across the top, lay out the times of day, and then have a schedule of unlocking chairs. Meaning, at 8 o’clock, we’ll have four chairs available. At 8:30, we’ll open the fifth chair. At 9:30, we’ll open the sixth chair, and so on. It’s the chair-based scheduling. The reality of it, it simply does not work. For all the reasons we talked about, stochastic and deterministic, it doesn’t work. What happens is when you schedule this way, people often use half-hour intervals, 8:00, 8:30, 9:00, 9:30. It looks nice. It looks convenient.
It wastes a ton of capacity is one problem. And it creates bunch starts where five or six patients are showing up at the exact same time, which creates a bit of a gridlock problem. Now, how does it waste capacity? If you think about JFK. Airport in the ’60s, they too had flight departures every half an hour. And they could do a few hundred flights a day. Now, they have departures every 90 seconds, and they do a few thousand flights a day. So when you artificially create longer gaps than is strictly needed, it ends up being a big drain on capacity. The second thing that happens is we talked about the tennis court reservation model. It doesn’t work because of variability. And the third thing that happens is this approach, at no point in time, looks across the population of people in chairs. And says, do I have the right portfolio of people who are in the first, third of their journey in the infusion chair, to those who are in the middle of their flight, to those who are beginning to wrap up their journey? You need to get all three of those in balance in order for the thing to move smoothly. And a chair-based schedule does not even try and touch that because they’re just getting schedule as they come. Therefore, you get the triangular profile. The second approach that does not work is a nurse-based approach. Many centers try and look at which nurses are on deck, try and sort out either with acuity or with just timing and duration who should sit where, and try and balance it out.
Unfortunately, this doesn’t work either for a bunch of reasons. One is, when nurse two got assigned this set of patients, there was an implicit assumption that patient two would arrive half an hour after patient one. If patient one is now 25 minutes late, the domino effect has started. Nurse two now has a 5-minute gap between her two patients, and therefore, is obviously not going to be able to deal with it. Second is, appointments don’t end on plan. They sometimes need more hydration. They sometimes have a reaction. They sometimes, it took longer to get the meds formulated. Whatever the reason is, if an appointment runs long, it impinges on the ability of that same nurse to start the next appointment. And the third thing, which again, is the fallacy of averages, is while the aggregate allocation by acuity seems fair and balanced, of course, it depends on the timing. Two nurses could each have 30 acuity points. But if one of them has three acuity points per hour over a 10-hour shift, and the other one has 20 of the 30 acuity points showing up in a two-hour window, they have wildly different experiences. And therefore, even though they both have 30 acuity points, they did not have the same experience. And so, this attempt to balance by nurses doesn’t work. And you will also get a triangular profile. The only way that works is an optimizing approach that balances durations. Why does that work? First of, the duration is the factor that figures out when you should start it. And it can create a flat profile by making sure that your starts, and your mid-flights, and your wrap-ups are roughly in sync through the middle of the day once you’re fully ramped-up. The second thing that happens is, implicitly, you are delaying the binding of the patient to the chair and the nurse as far as possible, right? In order for an infusion to happen, four things need to come together. The patient, at that time, the chair, the nurse, and the time slot. So, the time slot, the patient, the chair, and the nurse. To the extent, you first bind the patient and the time slot. And then delay the chair and the nurse as much as you can. You end up with a more optimal profile.
This is why when you book a hotel room, they’ve got the date of your arrival and your name. They do not book you to a room-type or a specific room, room number 706 or whatever. They’d find it delayed because at the last minute, based on housekeeping check-in, check-out, they can manage the optimization. So delayed binding is your friend. And the third thing is you design the volumes and durations so that you can play this winning game of Tetris and organize how the slots should be arranged, so that it flows smoothly. It ramps up smoothly, stays flat, and ramps up smoothly. So this is why duration-based approach is the only way it works. So now, let’s get into the details. How do you make it work? What you have to do mathematically is take this profile and make it this profile. And in order to do that, there are five math problems to be solved. The first one is predicting the volume. How many patients will you get on a Monday? How many will you get on a Tuesday? How confident are you in that forecast of the volume? Does it cover half your Mondays or does it cover 95% of your Mondays? And sure, there’ll be some egregious cases where it didn’t fit. But by and large, you can take that forecast to the bank. So that’s the first hard math problem to solve.
The second hard math problem to solve is, what’s the mix? If you’re going to get 60 patients on a Monday, how many of them need a one-hour treatment? How many of them need a two-hour treatment? How many need a three? And can you be confident in that mix with a fairly high degree of certainty Monday after Monday after Monday?
The third is an honest self-assessment as to how accurate your system is in predicting the duration. So if you said this is a three-hour treatment, how accurate are you? Is it always between 2:45 and 3:15, in which case, you’re very, very accurate? Or is it a jump ball? It could be a one-hour appointment, could be a nine-hour appointment, you’ve just logged it as a three. And the distribution is all over the map. That’s the third problem. The fourth problem is, now that we know what the hour profile is for each duration in your forecasting, and they’ll be different. Your accuracy in forecasting one-hour treatments might be much, much better than your accuracy in forecasting five-hour treatments. So once you understand how accurately or not you predict each duration bucket, how do you factor that in into creating a little bit of flex in the length of the. Tetris block, if you will? And finally, what is the supercomputer winning Tetris strategy, given this collection of Tetris blocks with the mix and the duration of the volume? So once you figure that out, you’re then able to lay out the Tetris blocks in a way that’s going to make that Monday go right.
And it’s a different layout of the Tetris block to make Tuesdays go right, and Wednesdays. So each Tetris block needs– each Tetris game needs its own configuration of blocks, such that that day of the week goes well. Why is this magic? It’s magic for four reasons. The first and most important one, it unlocks capacity. Right when you needed in the middle of the day, you will have chairs available. And if you have chairs available, lots of good things happen. Add-on patients can show up. Someone can run late. Someone at the clinic, someone can run long in their duration treatment at the infusion center. You’ve got the shock absorbers to absorb that. It’s a bit like having a few extra bucks in the bank to deal with an unplanned expense, right? That’s kind of what the first one is. The second one is, patients have choice. If you’ve played this. Tetris game correctly, there are many trains leaving the station throughout the day.
So it’s not as if, boy, if don’t get the 9:00 slot, I’m stuck. There are no more three-hour appointments for that day. If there are lots of three-hour appointments that need to be given, you have to mathematically spread them and sprinkle them throughout the day so that it becomes like New York City. You never look at the timetable when you’re in New York. Just go to the subway, stand on the platform. And in the next five minutes, a train will show up. So you don’t need to look at the schedule because the next train is along any minute now. That’s kind of what you have to push for. It flattens the workload for nurses. Infusion nurses have a horrendous workload. It’s very like early morning and late afternoon. And it’s a hair-on-fire time in the middle of the day. So leveling the workload for nurses creates a lot of sanity and a lot of staff satisfaction for that. And finally, it fits the nursing schedule. Nurses who show up first leave first. Nurses who show up last leave last. So suddenly, this notion of a nurse’s schedule to get off at 6:00, but at 5 o’clock, is told, hey, can you stay until 8:00 today because we’re running late.
Sure, that’s fine once, twice, three times. But after a while, it becomes very difficult to manage your home life, and kids, and school, and dinner, and all of that if it’s unpredictable when you can get off of work. And that’s a big source of nursing dissatisfaction. So getting the math on this right takes all that away. Now, how do you actually do it? Remember, we said scheduling to a chair is wrong. Scheduling to a nurse is wrong. You don’t schedule to those things. You use those things as constraints. Because when you understand how many chairs you’ve got and when they’re available, how many nurses you’ve got, when do they start, when do they end, those become mathematical constraints in the quest for an objective and optimization answer. So what do we do? First, you start out by saying, all right, what’s the timing envelope in which this infusion center operates? On Mondays,. Tuesdays, Wednesdays, each day of the week, what’s the start time? What’s the end time? What is their tolerance for running up until the end of the day? Are they OK with the last patient scheduled to get out of the chair at 8 o’clock when their close time is 8 o’clock? Or do they say, nah, I want a 30 to 45-minute buffer.
So let’s plan this to the last patient that’s out of there at 7:30, that we have got 30 minutes to clean up, wrap up, and have everyone out the door at 8:00. So you’ve got to understand the buffer. How do they treat holidays? Different parts of the country and different holidays behave differently. Some places, the day after Thanksgiving operates a lot like a Sunday. Other places, it operates a lot like a Saturday. Christmas often operates like a Sunday. And so, being able to figure out that, yes, Christmas may have been on a Wednesday. But from a template and a Tetris block layout, it behaves like a Sunday. It’s a very useful thing to know. And then what happens to the pharmacy? When do they start? When do they end? Do they have any constraints? Like, we don’t mix at lunch or we stop mixing at four, which would govern when you bring patients and how you bring patients. So you have to understand all the constraints around the operating hours envelope in which you run. Having done that, you then need to understand all the constraints around the nursing schedule, right? Which is, each day of the week, what’s the start time? What’s the end time? Therefore, the shift length of each nurse. How many nurses of that type? And what type of nurses are they? Are they infusion nurses? Are they simply dedicated to injections and blood draws? So you’ve got to understand all of that.
Having understood that, you also have to take nurse capacity into account about how you deal with lunch. Are you just saying, hey, at any point in time, no more than two nurses go out at lunch or go out for lunch at a time. Is it one at a time? Is it a 30-minute lunch? Is it in a one-hour lunch? What are you doing about it? Because you are effectively taking down your nursing capacity in various ways, any time between 11:00 and 2:00. And so, your strategy for dealing with lunch affects the nursing capacity you’ve got available, which is a mathematical constraint into the optimization. So if you don’t get that right, being able to lay the Tetris blocks out right becomes impossible. Having done that, you then get into the duration buckets. This is now saying, how many Tetris– how many types of. Tetris blocks do I have? Do I need a one-hour, two-hour, three-hour, and a four-hour? Or can I do one and two as one block, and three and four as another block? What needs to be done before you make that determination is not to make it in a vacuum.
What needs to happen is, you have to lay out all the 10,000 appointments you’ve done over the last several years. Let’s say, if you’ve got a historical data dump, lay them out from shortest to longest and see where the clusters naturally form. So if you do a lot of one-hour appointments, the number of things in the one-hour bucket will be very, very high. If you do little ones, little twos, little threes, then the one to three might be a reasonable bucketing strategy. It’s a bit like looking up at the sky at night and seeing where the natural clusters of stars are. What we do is take the data and see where the natural clusters form. When those natural clusters form, we can come back and say, we think you should segment your appointment durations into the following six buckets. A zero to two bucket, or two to three bucket, or three to four, four to five, six to eight, and a nine plus. Those buckets form natural balancing clusters that gives you enough choice to play the Tetris game well. Having done that, we can then lay out additional constraints, right? Which is, for each bucket, what is the min and the max, what you expect the duration to be? And then, this is very important. Because just the duration and the chair doesn’t matter. The nurse matters as well. And the nurse time spent with the patient is entirely dependent on how you deliver clinical care. So some centers may say,. I want a nurse one-on-one with the patient for the first 30 minutes and for the last 30 minutes as well, or for the first 30 and the last 10, or 20/20. So you describe how much sole focus attention you want putting the nurse and the patient at the start of the infusion, at the end of the infusion, and what the coverage is in mid-flight. Some people say, I want a nurse to cover no more than three patients at a time. Or I want to nurse to cover no more than five patients at a time. So different centers do different things.
The optimization algorithm should not opine on your clinical standards and how you deliver care. The optimization algorithm needs to accept your clinical standards and optimize against that. Now, if, mathematically, it turns out to be an infeasible solution, the optimization algorithm needs to tell you. So for example, in an extreme case, just to make it clear, if I were to say, I want my nurse one-on-one with the patient for the entire duration of the treatment. And I’ve only got three nurses. Then it’s unlikely. I’m going to be able to deal with 30 patients in that center. Because all my three patients– all my three nurses are locked down with one patient at a time. So that day, you can create clinical constraints that make it impossible to find a solution. And so, the optimization algorithm needs to come back and challenge you on those, rather than opine for you. Having done that, the next set of things you need are, what are the time windows? Implicitly, if you’ve got a six-hour appointment and you wanted to wrap up by 7:30, then you should not start that later than 1:30, right? And so, each duration bucket had a last time to start, which is all gated by the time you said was the end point. So this also sets up rules for the Tetris scheme.
So you will find long appointments tend to bias last on the Tetris game. Short appointments tend to sprinkle evenly. So this is the way the constraint plays into the placement of the Tetris blocks. Having got the time window constraints, what needs to happen then is to get the historical data. Because this is a deep-pattern recognition algorithm. So everything needs to go back historically. So you can see in some cases, we’ve got 150,000 records. This infusion center sees only 100 patients a day. So it’s 1,500 days almost of data, right? So it’s four or five years of data. What this does is provides the pattern of the volume by day, the mix as a function of that volume, how the starts and stops work, what the delays are. And from this, you can create a starting pattern of how to win that historic game of Tetris by solving those five hard math problems I laid out earlier. So once you run the optimization, to build the templates, here’s what you’re aiming for. You’re aiming for a green line like this with a slow and smooth start, getting to the near chair peak but not crossing, staying flat for as long a portion of time as you can, and then ramping down smoothly towards the end of the day. OK? Now, why don’t we start more abruptly? It’s difficult to start more abruptly.
In this place, they’ve got 36 chairs. If you said, I want to start at 36 patients, well, 36 patients starting at 7:00 AM would require that you had at least 20 or 25 nurses on deck to get started. So that’s not feasible. So you need to kind of invariably have a ramp and a flat portion. So in order to do that, this is where those math problems I talked about coming in. You have to figure out the volume. You have to figure out the mix. You have to figure out the Tetris blocks and lay them out this way. It’s not enough to do just that. You have to simulate reality. Because life never happens the way the mathematical model said it would. How do you simulate reality? There are a thousand reasons why an intrusion treatment goes late. Everything from couldn’t find a parking spot to construction is making the commute crazy, to labs, to pharmacy, to doctors, to et cetera. All the thousand reasons finally manifest in only one of two variables. What is the punctuality of the patient? Do they show up early, on time, or late? And two, what is the duration of the treatment? Was it shorter than you expected, equal to what you expected, or longer than you expected? The reason why it happened doesn’t matter. Those are the only two parameters. What you can do in a simulation algorithm is go back over the tens of thousands of appointments you’ve had, figure out the lateness and the leakage on duration, and say, this is the probability of lateness or leakage occurring by time of day, by day of week. OK? Once you do that, you can run a discrete event simulation algorithm that says, yes, I know, theoretically, you’re saying this is the template. When I run it through a simulation, this is what I think will really happen. OK? And it will happen for a bunch of reasons, even though they had slots on the calendar, they went unfilled in the morning, or that afternoon. It didn’t turn out to be as bad as they thought. Each center has its own profile about how the theoretical templates will manifest as a simulation output. But this gives the center a lot of confidence that the Tetris configuration being suggested is a robust configuration that will work. It will withstand the shocks that will inevitably occur. So finally, after all of this drama, the output is a stunningly simple output. And the way I like to think about it is, if you look at the Google screen, it says Google. It’s a blank white screen. And it’s a rectangle that’s says search. That’s it. They’ve spent $5 billion making that. And to continue to spend $2 billion a year making it better. But that’s all it does. So similarly, for all of the crazy math, and the simulation algorithms, and et cetera, et cetera, what we’ve got is a template that lays out like this.
Across the top are durations, right? Zero to two hours, two to three hours, three to four hours, four to five, five to six, et cetera. And down the side are times of the day on Mondays when you should be offering up slots. Think of these as the number of stocks of that patient. So 8:00 AM on Monday, zero to two hours, you can start two patients at that time. But don’t start anyone else. OK? So let me make this a little bigger so it’s easier to follow along. Yeah. Let me highlight four or five pieces of this. One is the duration. Again, we do the clustering to get them into the right buckets. That matters because they’ve got to stay somewhat balanced. Second is the time. You can start at 10-minute intervals or 15-minute intervals. Generally, tighter intervals are better. 30 minutes is not great. An hour is definitely not great. You don’t need to be crazy about this and start at 7:00, 7:01, 7:02. The reason we converge on 10 is 10 is a reasonable window of time for a nurse to walk up to the front desk, greet the patient, walk them back, get them seated, and have the initial conversation with them. So 10 minutes of a heartbeat feels like a pretty reasonable heartbeat.
Third is, you have to figure out how many simultaneous starts you can accommodate. This center, between 8:00 and 8:30, all the nurses were on deck and the patients were not. So they said, hey, we can deal with three or four starts. Once it gets past 9 o’clock, life’s a bit crazy. I don’t want to start more than two at a time. So across the road, you’ll never see more than two stops at a time. All right. The fourth thing that happens is the duration accounts for that error we talked about, which is, how predictable are you when you say it’s six to eight? Based on that, we have to flex it. And the final piece of magic is, how many starts of that duration you should do it that time? This ties back to why. I said first come, first scheduled models are flat out wrong. You can still be patient-centric. The conversation and Mrs.. Jones goes something like this. Rather than saying, we’re open 8:00 to 5:00. When would you like to come? Or how about 8 o’clock, or how about 9 o’clock, or any random time? The conversation between a scheduler and Mrs. Jones goes like this. Mrs. Jones, I see you need a three to five-hour treatment on Monday. I can seat you at 8 o’clock, 9:20, 9:40, 10:00, 10:20, or 10:40. Would any of those work? And if Mrs. Jones says, yes, one of those work, great. You’ve got a match. You’ve put the square peg in the square hole, and it’s been perfect. So your jigsaw puzzle is forming correctly. If Mrs. Jones is stubborn and says, nope, I need to be seen at 8:30. And that’s the only time I want to do it. Fine, do it on overbook. Some small number of overbooks won’t kill you.
Because mathematically, remember, we’ve done the simulations and made it robust. Now, if you overbook all of them, it’s as if you didn’t have an answer key, and you were just putting them down wherever you wanted. Anyway, in which case, it won’t work. Over time, you can teach your schedulers to become very strategic. So when they see adjacent slots like this, a 3:00 to 5:00 next to a 6:00 to 8:00, and they try and get Mrs. Jones to take this spot and this spot is gone, they can offer up this one. So they can put things into the wrong duration bucket, overbook, underbook. As long as these discrepancies are small and under control, the system, as a whole, performs perfectly. OK? OK. Now, what does this do? The reason it unlocks a massive amount of productivity improvement is when you go from this shape Tetris to the shape Tetris, what you’ve done is you’ve driven out the white spaces. The white spaces, in essence, are wasted capacity. It’s leaks in the bucket of either chair capacity on nurse capacity. When you tighten it, you essentially improve the productivity of the system as a whole. The simplest way to measure productivity is to think that the numerator is patient hours, which is, how many patients times the durations. So if you had each patients and they had one-hour, two-hour, three-hour treatments, et cetera, you add up all the hours they needed, that was the total number of patient hours in a day. The nursing hours are the number of FTE hours of nurses who care for patients, not the supervising nurses or the management nurses, who actually had their hands on treating the patients. When you measure this ratio, you can, with this optimization, make it 20%, 25% better. When you make it better, how do you monetize it? You’ve got different ways of monetizing it. Let’s say you’re a growing center or you’re finding you’re pushing patients out or too long.
You could increase the number of patient treatments you see in a day or you could absorb the future growth. Say, you absorb 15% growth by only adding 5% staff, and not 15% staff. This is one way of monetizing the productivity. A second way of monetizing the productivity, if you’re not in a high-growth setting, is reduce your costs. You could reduce your hours of operation because you can see the same patient workload and close at 6:00, instead of at 8:00. In which case, your shift costs and your overhead costs are lower. Or you could consolidate. If you’re forced to keep open a little satellite center because your main center is creaking under the strain, you could potentially not need the smaller center anymore. Or you could say, I’m going to apply all of this productivity to offering superb service to patients, right? So same nursing load, same patient hours. I’m just here to reduce the wait time. I’m going to give them appointments in a shorter window. I’m going to have more nursing one-on-one time with them so that the care experience is better. The way you think about this is, what this productivity lift does is it gives you. Disney Dollars to spend. There’s three ways of how you choose to spend the Disney. Dollars, on rides, on food, or on merchandise, right? So it’s entirely your choice. So clearly separating the creation of the productivity from the monetization of the productivity is a very important concept in how you get this. So once you get down to the, I want to tighten the. Tetris blocks, there are various things you can do. Your optimal shape is this trapezoid.
And you can start to see wherever there’s leakage and how I cut it. If you want to get faster on the ramp up, it’s a steeper trapezoid. If you wanted to stay later in the afternoon, you can make the descent steeper. Finally, when you run out, you add chairs that lifts the top of the bar or you stay longer– you stay open longer, either on weekdays or start opening on weekends. It increases the hours. So this is the way you think about increasing the size of the football field on which you play. But for each decision, it has to be backed by mathematical simulation, right? What if I did the following? How many more patients could I see if I change my nurse shift? So each question has a mathematical set of answers behind it. So that was the double click on, what does it take to do it? Again, the next two topics will be covered in great detail at the third and the fourth webinar. But let me do the two-minute version just so you get a sense of what’s coming. How can planning be done in advance? If you build this right, you should be able to forecast a day in advance of the day with a very high degree of precision. In this case, it’s at 10-minute windows where gray means you’re running right on track. Yellow means you’re running slightly lighter than planned. Orange would have meant you’re running slightly harder than planned. Red would mean you’re running above chair capacity, which means your waiting room is beginning to have chairs on it. Once you plan it one day at a time, you can then get the weather system for the next 30 days, also at 10-minute precision.
This is incredibly powerful. If I can tell 20 days in advance that I’m going to get a weather system forming in the middle of the day,. I could double click on it and say what exactly is forming and start to say, oh, I get it. There’s a one-hour period where. I’m really running flat out. Maybe I need to bring in one more nurse from the floor pool or see if I can get two patients to move earlier. I know John Smith and Jane Doe are flexible on their commute. And they don’t have a leading clinic appointment before it. Let me call them and see if they can come in earlier. Obviously, it’s a last resort. No one likes shuffling patients around. But being armed with this level of precision is incredibly powerful. Rather than the fate of most infusion centers where the nurses show up to work knowing that the tsunami will hit sometime during the day, just not knowing when and how it’ll hit and how bad it’ll be. And they suffer through that and come back again the next day. So having this precision is very powerful. We’ll dive more into what planning means in webinar three. Webinar four, we’ll talk about how you do very robust diagnostics. There’s a lot of data already captured in EHR. You don’t need new data. But there are ways to slice the volume where you’re hitting the peak, how often you’re flirting with the edge of the peak, what your add-on and cancellation rates are. Do they compensate for each other? Do they not? Do they aggravate each other? What are your arrival and signature of a patient are? Are patients arriving chronically early or are they arriving chronically late? And helping you avoid make the mistake of– just because patients arrive early and the nurse looks like she’s available, and a chair looks like it’s available, to go ahead and seat the patient early because that’s a good patient-centric thing to do. It turns out it’s mathematically the exactly wrong thing to do. So we’ll dive into all of those things.
And finally, it boils down to, what is the profile of your ideal Tetris game? How well are you scheduling against that Tetris game? Which means, are you following the rules of the game and listening to the recommendations? And three, then how well are you executing against your own schedule? Meaning, is there truth in scheduling? When you put out a schedule, is that how the movie plays out or does it play out completely differently? So we’ll talk about all of those things. We are at the end now. We’ve got time for Q&A. I don’t know if any questions are coming in yet. So go ahead and type any questions you’ve got.
We’ve got one question in– oh, a second and a third. Let me start with the first one. Does your software interferes with common EHRs? Yes, we are EHR agnostic. It works. We have, so far, customers, little over half of them are Epic. About a quarter of them are Cerner. And the rest are distributed across a variety of things, ARIA, Flatiron, MOSAIQ. We’ve seen a whole bunch of them. The reason it’s EHR agnostic is we don’t need to integrate with the EHR. There’s no IT project to connect. We don’t write to the EHR. We don’t read directly from the EHR. What we work with our customers to do is have them pull the data out of the EHR in the form of standard report and push it to us like CSC files. And we help them get it to where it’s automated. Every night, we get a push. And that lets us run without getting caught up in the specific IT-like integration of EHRs. As a result from kickoff to go live, it tends to be an eight to 10-week window. It’s very, very quick.
Second question. We’ve made many attempts at process improvement. Things get better for a while and then slide back into the way they were before. How is this different? That’s a great question. What happens with process improvement, not knocking it, the great thing to do, is they’ll tend to focus on the various steps in the process that cause bottlenecks. We didn’t get authorizations, or the order was incorrect, or the meds weren’t validated. So they’ll work their way through the life of the patient from the time they enter the infusion suite to when they leave, right? And try and knock these things off one at a time with a good robust root-cause problem solving. Which says, OK, if you’re not getting authorizations right, let’s make sure from now on, we put a check to get the authorizations confirmed the night before. So lots of things like that are done. The reason they don’t fix the problem is, think of it as the jigsaw puzzle is the way in which the appointments were allocated. Now, on the morning off, you’ve got what you got. And the patients are going to show up according to the appointments they were given.
So if the jigsaw puzzle was constructed in an incorrect manner, there’s very little you can do by process improvement to solve it. So the way to get the jigsaw puzzle done correctly is to have played that winning. Tetris with the forecast by day, by mix, et cetera, et cetera, and shuffle them. And then to stay tuned from a machine-learning basis to continuously learn when the jigsaw puzzle needs to be tweaked. Because the volumes have changed, the mix has changed, the patient profiles have changed, new oncologists have been added, new hours of operation now in effect. All those things caused the jigsaw puzzle and the Tetris– winning Tetris hand to be shuffled. And so process improvement cannot automatically stay up with all of that. And that’s the reason, no matter how successful it is for the first 30 days, 60 days, 90 days, 120 days, it will 100% degenerate back to the 10:00 to 2:00 peak at the end of time. OK? How does schedulers actually use the software? Do schedulers or admin need to do detailed calculations? No. None of it. Our model on this is, all of us know how to drive a car. Very few of us understand how a fuel injector works. And so the software is built with that philosophy. And you can think of. Apple’s philosophy. And iPad is a very sophisticated, very complicated device. A two-year-old can also use it perfectly fine. And so, schedulers do not touch any of the math. In fact, schedulers do not even touch the software application. The optimization algorithm gets fed back as EHR complex. So today, that. EHR has templates. Let’s do two appointments at 8 o’clock, four appointments at 8:30, et cetera. That was just informed by gut feeling and common sense. We’re replacing that with math, all right? And so what we do inside the EHR is say, let’s create five new types of appointments. An infusion one-hour, an infusion two-hour, an infusion three-hour, and so on. Now, on Mondays, offer up infusion one hours at 8:00, 10:00, 9:20, 10:30, 11:40, et cetera.
We’ll figure out the math of that for Mondays, Tuesdays, Wednesdays. And so once that gets baked back into the EHR, all that the scheduler is doing is looking for an infusion of their appointment of the right type and awarding it much like they do today. We’ve just spent 20 minutes training them on how to make intelligent trade offs. If you can’t find exactly what you’re looking for, find a slightly bigger container. If you can’t find a slightly bigger container, find a slightly smaller container. If you can’t find that, overbook. But try not to do it too often. Try and overbook in a domain where either right before or right after, there’s an unbooked slot, and hold onto that slot. So we teach them the four or five tricks. And the schedulers can do it just fine. Another question now. Are infusions are set up by office after the doctor’s appointments? So the infusion area does not have the ability to set up appointments. How could we do this? This is the classic linked appointment problem. And the linked appointment problem is, think of it as a connecting flight, Boston to Chicago, Chicago to Seattle. Where the first leg of the flight, Boston-Chicago, is the clinic appointment.
The second leg of the flight, Chicago-Seattle, is the infusion appointment. When you make it as a connecting flight, it places a lot of stress on the system. And the reason it places stress in the system is, how long should the connection be in Chicago? If it’s 20 minutes, you’re guaranteed to miss the flight. If it’s six hours, you’ll make the flight. But you’ll curse for five and a half of the six hours because you wasted your day. And the third source of stress is, if your flight pushes back from Boston a little bit late, you’ll worry you’re going to miss your connecting flight. What we do mathematically is turn that model on its head. Now, imagine I will tell you, instead of a connecting flight to Chicago, I’m going to book you on a nonstop from Boston to Seattle. Guess what happens? You’re no longer stressed about the connecting flight when your flight pushes back late. When you get to Seattle, that’s the clinic appointment, the long flight. When you get back– get to Seattle, step onto the curb and take a taxi. That’s your infusion appointment.
Now, pretend you’re the fleet manager of the taxi service. How many cabs should you have waiting curbside on Mondays at 7:00, Tuesdays at 8:00, et cetera? How many of them should be minivans? Because families are traveling with lots of luggage. Initially, as the fleet manager, you start with the flight schedule. United’s bringing in a 747. I’m going to have 200 taxis waiting. Delta’s bringing in a regional jet, I’m going to have 50 taxis waiting. Over time, you realize that the flight schedules are unreliable. So you build your own data. For six months, at 7:00. AM, 7:15, 7:20 on a Monday, on a Tuesday, on a. Wednesday, you count noses. How many passengers have stepped onto the curb and what kind of a taxi did they take? Once you understand that, you’ve built your model of how many cabs you need every minute of every day. Guess what has happened once you have built that model? You suddenly earned independence from the flight schedule. Now, when a passenger steps on the curb, you no longer care if they are an on-time passenger on a United Flight or a delayed passenger from a Delta flight. It doesn’t matter. Your model said a passenger is going to step on the curb now. We have done the same thing. We built all our models by counting noses as patients step across the foyer into the infusion suite. Once you get the models right for Mondays, Tuesdays, Wednesdays, we no longer care whether the patient is arriving on time after an appointment with the breast clinic or they’re arriving late after an appointment with the sarcoma clinic.
It no longer matters to us. And so, this is how you de-link it. You can never get an oncologist to behave differently because of the chair. The tail cannot wag the dog. So you have to tell the oncologist, practice exactly as you practice. See whomever you see. Schedule whomever you want. We have understood how to manage it. Then the clinic appointment, if the same person, we’ve got customers who do the same person, books both the clinic appointment and the infusion appointment. They know how to make the connecting flight. And they know if they run late, it doesn’t matter. They just do their best efforts. And if you’ve got separate people scheduling it, that works out too. Because we know when they were seen. So this is how you decouple without doing a real decouple. Because some centers tried to do the clinic appointment in the previous day and the infusion appointment on the next day. That’s a terrible idea. That would be like if United tells you, drop your bags off one day before your vacation. Then take the flight. And then come back one day after you land to pick up your bags. There are certain aspects of journeys that just should not be decoupled in ways that make it so grossly inconvenient. And so, that’s the way we would deal with that. Part of this affect nurse and patient experience and satisfaction. Actually, enormously well, in many cases, we run a nurse satisfaction survey before and after and ask pointed questions.
Your personal satisfaction, satisfaction as a team, the flow of current patient arrivals, the workload for RNs, patient wait times, and patient’s stat, right? And the typical results we see is nurse satisfaction improves by 30 and 40 points. So in many of our cases, they’ve improved from the mid-20s to the mid-60s, specifically on things like workload and stuff where the level of current satisfaction is simply not good. Because nurses are aware that they feel less than fully utilized early morning and late afternoon. And they feel overwhelmed and overworked, and prone to potentially making mistakes in the middle of the day. OK. We’re down to five more minutes. That covers the questions we’ve seen so far. Any new ones coming in?
MODERATOR: It doesn’t look like it. But if you have any questions post-webinar, you’d see our contact info on the console. Feel free to reach out to us. I want to thank Mohan for presenting today’s webinar. And I want to thank all of you for joining us. One reminder, keep an eye on your inbox. You should be receiving a link to a recording of this webinar that also includes the slide deck. You should get that within 24 hours. And thanks again to all of you for joining us. Thank you.