Sandip Vasavada: Thanks. Thanks, Alan, and thanks for inviting me. This is certainly a wonderful opportunity to present some of this experience we've had. So, this is, like you mentioned, our opportunity to say, can we better predict or prognosticate who would go on to advanced therapies and perhaps which one is best for that individual patient? So, many people who are watching in would know that overactive bladder therapy and outcomes are something we deal with on a regular basis, no matter what our specialty is. Per our AUA and SUFU guidelines, certainly the first three versions, which I was fortunate to be part of, we came up with the simple stuff, behavioral modifications, medical or pharmacological management. And then we knew that we used to call them third-line therapies, but really neuromodulation and chemodenervation were the mainstays of those. So, the most recent version of the guidelines, which I actually moved to stress guidelines, but my very close friends and colleagues are running the OAB guidelines with this, but they really emphasized shared decision-making. And so, that's something that really we use to eliminate the step therapy, but really emphasize talking to a patient, going over what is best in their particular situation. So, it does put the onus on us to make some form of a decision assistance with these patients to really help them choose the best outcomes and the best choices for them.
Many of us know that we use, in some of our therapies, almost a 50% reduction in incontinence episodes that allow us to best assess their ability to empty and improve their symptoms in the best way possible. So, it's some of the measures that we've used. We've used it in both sacral neuromodulation most certainly, and we also use it selectively in onabotulinum toxin, obviously keeping in mind to balance efficacy, but also side effects. So, we want to do that. So, there's a few things that the guidelines don't really tell us. They don't really prognosticate who would actually fare best with the actual therapy itself. And then guidance to those patients who'd be at higher risk for certain complications or perhaps adverse events. And other algorithms that we've used in one way, shape, or form don't really have the ability to take any of the external datasets that we've had to prognosticate who would do best. So, mind you, this is not a concept that we just came up pontificating. This is something that was even mentioned well over 10 years ago. And if you read this editorial comment, it was listed. It would be a step forward to be able to predict in which patient's therapy for various symptoms of overactive bladder is destined to fail. This depends, of course, on the definition written by the one and only Alan Wein.
So, again, nothing we came up with, but well, 10 years ago or more, Alan very eloquently said this, and I think fast-forward 10 years, here we are still trying to answer that very simple and perhaps basic question. So, asking ourselves with modern technology, can we use, in this case, machine learning to help predict who would respond to advanced or as we now call them minimally invasive therapies? Can we show those patients who would be at risk for complications to help educate them as to perhaps what's best as far as an option going forward for them? And can these AI techniques help us do this more accurately and faster? So, the only actual dataset that we could really lean on was that of the ROSETTA trial. So, the ROSETTA trial was a JAMA publication. I was one of the lead authors on that as well. This is, again, almost 10 years ago. And this was a comparison study of onabotulinum toxin for sacral neuromodulation in the refractory urgent incontinent idiopathic patient population. So, this is one of the only RCTs of those two advanced therapies. So, the only head-to-head comparison that we really had, it has a complete dataset and that helps us from a teaching perspective to these machine learning algorithms. We then, as many of you perhaps are aware with this, randomized these idiopathic overactive bladder patients to onabotulinum toxin or sacral neuromodulation. Initially had a 6-month data time point, and we did one-month data points at each, so with voiding diaries and questionnaires on a monthly basis. And then we also had a second paper, which was a 24-month outcome as well. And this was basically the basis for our machine learning and data. So, again, taking good data, using that, and then training that for the algorithm going forward. So, what was important is that we used the ROSETTA for the training data. So, anyone who knows a little bit of the nuances of this data knows that we used a higher dose of Botox, and that's sort of a separate discussion point, but it's a higher dose of the Botox than we use typically in the idiopathic patient population.
Everything else is pretty much standard. This is a fairly refractory subset, had a fair amount of leaks per episode. We actually also did urodynamics on that because we used that to teach the protocol and teach the algorithm for the machine learning so then we can then apply that. So, then we ultimately validated that on the ABC trial. The ABC trial dataset, which is one, the anticholinergic versus Botox dataset used as we would use idiopathic patients with a hundred units. They also had urodynamics very minimally. In other words, what we do on a typical daily basis, we don't need to rely on urodynamics for the average overactive bladder patients. So, this is more consistent with, as we would say, real world. So, we use the really detailed dataset, urodynamics, a little different with 200 units, and then use that to train the machine learning and then apply that to more real world, which would be the ABC trial data. And so, then from that point, we use something called operator splitting. So, this is actually some people from the quantitative finance. These are actually hedge fund managers and people who do their quantitative data using it for things like money and finances and very important stuff and saying, "Hey, can we use some of their approaches and use that to our data?"
And what you can see on the right is using this green line is our atom W with operator splitting, and it actually is learning the data with all these little scattered data points that are all throughout, which are all the data points themselves, and it can learn and mimic the actual data to show that we can do this in a faster and more efficient fashion and more accurately than just random data points alone versus even comparable training data with other forms of artificial intelligence and learning. So, we can do this by using this operator splitting method more accurately and faster compared to other networks. And the data analysis that we can get actually counts for missing data because we know we don't always have every single data point that's measured in every patient or every center. And so, having a method that can account for missing data is quite important. And most importantly is the transfer learning that we've been talking about. So, this means we can take a dataset that's quite detailed. So, that would be the ROSETTA trial data and transfer those learnings to a real-world dataset, which would be much more scalable because, again, routinely we're not getting urodynamics, et cetera, we're getting more limited data sets on our regular patients. So, transferring those learnings into a real-world dataset is incredibly important. So, when we ultimately did this, and we can look at this objectively, so asking ourselves, "Can we predict subjective response?" which would be like the PGI-I, 50% reduction in objective urge incontinence episodes, and even the BFLUTS subscale.
So, we look at very different subscales, patient-reported outcomes versus objective outcomes as well. And ultimately, both on sacral neuromodulation and botulinum toxin therapy, we were able to accurately predict those who would respond with about a 0.66 area under the curve. So, we're quite good with the AI algorithms when we look at this in every which way. We look at also side effects. So, side effects that most of us know for onabotulinum toxin injections is something we recorded with the ROSETTA trial was UTIs, because it's unfortunately a common situation that the AI algorithm was able to predict this with almost a 0.9 area under the curve in patients who had botulinum toxin. So, using our deep-learning methods, we can really accurately predict patients doing well and who's going to do well. So, in very simple terms, how would we do this? This is something like a counseling tool when we embody the entire algorithms themselves. So, we can give this patient an actual data point. So, your estimated probability based on a bunch of data points that we would input urge incontinence episodes per day, how many pads they may wear, how old they are, estrogen status, number of anticholinergics they tried and failed.
And then we can take all these types of data points, estimate their probability of responding after botulinum toxin would be X, estimated probability a leakage improvement would be X, so on and so forth. Giving them ultimately in the bottom box that Mr. Smith, 19 out of 20 patients with your sort of situation would be likely to experience clinical response to, in this case, sacral neuromodulation. So, helping them inform their decision with their next step to be able to do that. So, final take-home messages on this using this operator splitting methodology with AI, we were able to show that that'll outperform experts, and I'll hopefully call myself in the expert category as well as my colleague, and then both in the objective and subjective responses and refractory OAB cases, and we've actually validated this now on a more real-world dataset. Thank you very much, and I'm happy to answer some questions and go into more details with that.
Alan Wein: So, response, does that just mean greater than 50% or how much on the PGI-I does that mean? Does that mean excellent or does that mean... In other words, what were the gradations on the PGI-I that would determine whether somebody was a responder or not?
Sandip Vasavada: Yeah, to my recollection, Alan, this is based on the ROSETTA trial. We moved them up two points to show that there were response on the PGI-I. Otherwise, to your point about greater than 50%, yeah, unfortunately that was the metric used. It's still the metric used in what we call success for overactive bladder management in the refractory state, right? So, if I do a sacral neuromodulation trial, I want to see at least a 50% improvement to go onto implant. So, that was the data. I think we have some data on 75% or greater because that was also looked at, but for the most part, the yay or nay success or failure was based on 50% improvement.
Alan Wein: Does that pretty much parallel the number of pad reduction count or the number of leaks per day?
Sandip Vasavada: Yeah, very much for a number of leaks per day. Pad reduction count, not necessarily because as you know, there's patients who come in and they just feel more comfortable wearing pads because they've worn pads for 10 years and I'm just not going to go out in public without it. So, we've earlier used number of leaks per day. And so, that was the harder core objective measure that we tended to use for demonstration of success or not.
Alan Wein: So, it looked as though on those graphs that it looked like the non-invasive urodynamics were actually better than the full urodynamics than predicting what was going to happen or was that a misinterpretation of mine?
Sandip Vasavada: Yeah. I mean, again, I think real world, non-urodynamics, in other words, not doing urodynamics as we would do in an ABC trial or again, real world. We don't find that addition of the urodynamics data really made a difference. As you are well aware, out of the UK, they performed a randomized trial on getting urodynamics or not in this refractory state. Something we'd baked into our data and our guidelines for years, they actually did the next step doing a randomized trial, punchline being it didn't show any additional improvement by adding urodynamics, so I guess didn't see any differences that way. So, this is very consistent with that using that. And I don't think we feel that that additional data even is giving much more information. There are a couple data points of which I didn't show that it could make some differences, but as such, in aggregate, there's no need for urodynamics and we're not here to say otherwise.
Alan Wein: So, for the practitioner who's not at the Home Mecca in Cleveland who's out in Chillicothe, Ohio, let's say, but who knows how to do Botox and who knows how to do sacral neuromodulation, I mean, what are the important points for that person to consider without artificial intelligence or without typing in GPT with some parameters? I mean, what does that person best consider in their head to make this decision between SNM and Botox?
Sandip Vasavada: Yeah, I mean, obviously it's a little bit more than a coin toss. Even when we recruited patients for this study years ago, we didn't use it as a coin toss because there's always some subtlety, right? I mean, even today, there's a subtlety to it. If someone has perhaps dual incontinence, someone who... So, fecal incontinence and urinary, we may steer them more to sacral neuromodulation. Maybe they have more emptying dysfunction. We may steer them more to sacral neuromodulation. Maybe they have a neurogenic background or feel we need to improve their increase their bladder capacity, we may steer them more to botulinum toxin injections. So, there's still that element that I think you can't take away. That said, using something like our algorithm, and as we put this together to ultimately use for public use and dispensation, that would be, give me a few data points, give me patient's age, give me the number of incontinence episodes they have. So, a handful of things, very simple things that we would otherwise in a clinic visit ask for, and then we can have you put that into that second to last slide I showed you. And then it basically spits something out. "Mr. Smith, you are at this much likelihood of responding to this. You have this much percentage if you choose botulinum toxin of getting a bladder infection, so on and so forth to help them make the most educated choice." So, really simplifying all that with minimal additional data points, minimal change to the workflow that you would have in clinic.
Alan Wein: Anything new coming along that you know of that would substitute for either Botox or SNM with about the same efficacy, but minimal issues in terms of complications or risks?
Sandip Vasavada: Yeah, well, clearly there's a lot of interest in these implantable tibial nerve devices. I think it's exciting because it's perhaps less invasive, has maybe more staying power than a botulinum toxin injection, which has to be repeated. There's still some nuances with billing and reimbursements that still be hopefully flushed out in the next year or so, but it does give an additional opportunity for patients to choose another form of management for them. Interestingly, I don't think we could really bake anything into our algorithm directly because from a teaching and learning standpoint for these algorithms, we don't have that kind of a dataset to pull in, which would've been nice. And so, that's really why we had to lean on ROSETTA and pull that as a learning dataset to be ultimately able to teach and then go on to the applicability.
Alan Wein: Listen, thanks so much because I think that's really useful information and I think everybody is looking forward to the day when you guys publish something so that we all can come out with those little black boxes, that would be great.
Sandip Vasavada: Look forward to it.
Alan Wein: Appreciate it.
Sandip Vasavada: Thanks. Thanks. Thanks for having me.