Next speaker, Dr. Mark Richardson from University of Pittsburgh is going to talk to us about computational modeling to predict functional neurosurgery outcomes. Okay, thanks, Samir. Thanks, guys, for organizing this session. I appreciate the opportunity to talk. Yeah, it's a really interesting time with the, there's a real influx of technological and computational methods, like you just mentioned, the tractography, and Ellen gave us a nice overview. And computational modeling is another kind of hot area, which a lot of us don't know a whole lot about, including me. But I've gotten an education in this in Pittsburgh from people that have worked with me in the lab, and two people in particular, Nathan Sisserson, a med student who's here and who presented yesterday, previous career in data science, and Tom Wozni, who's an incoming resident in our program. So I just acknowledge these guys. So what is computational modeling? Very simply, it's a combination of techniques from math, physics, and computer science to simulate the behavior of complex systems. So this is sometimes called running experiments in silico. But what we're talking about is simulation. And simulation makes various assumptions, and one thing we'll talk about is the level of assumptions in different types of models. It's something to appreciate. So these models contain variables. They contain a lot of variables, which is why you need a computer to run the simulations. And we use these techniques to test, in some cases, how each possible combination might produce an outcome. And through literally thousands and thousands of these computer simulated experiments, we have the opportunity to predict physical experiments that can lead to the desired outcome, like running simulations on DBS parameters and then implementing those clinically. All right. So let's talk about brain simulation specifically, DBS or epilepsy, to understand this in a little more detail. So these are steps that we might take for modeling outcomes. So I was asked to talk about outcomes, but the extent to which this is done successfully may depend on the ability to model the underlying brain state. And Ellen talked about this. So the first thing one could do would be to identify neurophysiologic variables. So these are things like the inherent information in the neural signal. It would be things like the location of the recording, the condition of the patient during the recording, et cetera. Then one can build a model for this physiology. Then it's necessary to identify therapeutic variables. So what configuration was the electrode used for stimulation? What were the stimulation parameters? Again, the location, et cetera, et cetera. What was the temporal nature of any changes? And the variable field expands really broadly if you want to include all the possibilities. Then you have to identify the outcomes. And there are a number of different things you can categorize for the outcomes. And then the idea is to simulate all possible combinations to see what happens. And ultimately, we want to be able to predict a response to therapy. So I have the loop going in this direction, but there are different ways to address this. One is to look from the level of outcomes and go backwards. And the other is to start from the physiology and go forward. Or you could start in the middle and go either way. So there are different ways to interpret this. There are also parts that you learn from modeling that then feed back into a stage prior. I think it's important to talk about this, which is modeling at different scales. So the level of baseline neurophysiology, it's important to understand the normal physiology. It's important to understand the pathophysiology. These are both very difficult things to do, depending on the level of understanding that you think is needed for modeling a potential therapeutic outcome. One can model the physical distribution of the therapy. So this could include something relatively more simple, like the distribution of an infusate in the brain for drug delivery, or the electrical field distribution, so-called volume of tissue activation in brain stimulation. The individual neurophysiologic response can be modeled. And there are different scales here. So one can think about doing this on a per-stimulation event to understand what happens to neural activity as a result of each specific simulation. Or one can think of modeling the chronic effects and levels in between. And then, of course, there's group-level outcomes data. So what we're used to thinking about with predictive, you know, statistical methods that don't require computer modeling is just looking at the outcomes and pulling as much data as we can from the medical record and trying to find a correlation. So computational modeling is a much different approach that seeks to really draw in as much information as possible and build a model to predict outcome. Let's take this example, closed-loop simulation for epilepsy, to just think about this again in a little more detail. So this is a project we've been working on in the lab. Nathan talked a little bit about this yesterday. So we want to understand the baseline physiology. In this case, we do want to understand the physiological response of simulation, again, at these specific levels per-stimulation event and over the course of chronic stimulation. We want to know what the clinical responses are. We want to know when they occur. And we want to know what neurophysiological changes correlate with things like seizure reduction, improved cognition, worsening. So how do you do this? We're currently focused just at this level. And why do we need to do this? Here's a good example. So this is a schematic of the detectors and possible simulation parameters, one example, from the RNS device. So you can see there are four channels on each lead. There are different patterns that can be set for detection. There are different patterns that can be set for simulation. And so if you tabulate all the possible combinations, you get a massive number. So there's no way, for instance, the clinician or anyone can really assess all of the possibilities. So we set up a software platform to do this. And currently what we're doing is modeling how the device works. And when we build that model, then we can work on modeling what happens to brain function in response to stimulation. And ultimately we want to tie this into outcomes. All right, so again, it's data complexity and data extrapolation. So the RNS device or any closed loop DBS system that stores data has limited amount of memory. And so it's constantly overriding data. So the data that you get from the device may or may not be a reflection of what's happening on average. So there are different techniques that one can use to try to account for the bias. So that is an important part of the modeling. And it also just brings up the point again that there are different assumptions that have to be made in all these models. And it's important to note when you're doing that and what the potential outcomes are, especially when you're using this potentially to affect therapy, therapeutic decision. So I thought I'd just pull a couple examples from the literature. This is epilepsy work from patients at Mayo and MGH. And the idea here was to predict outcome following specific functional anatomic resection. So what they did was to combine data from ECOG and surgical outcome data where they knew the result of the surgery and they knew exactly what was resected. And the goal was to predict surgical outcomes by simulating removal of network nodes. And then also to suggest alternative approaches in cases where there's poor outcome. So here are two examples from their paper. And in black are the areas that were resected. And the rainbow scale, or red, is the area that according to their model had high likelihood for seizure. And they set up a model where essentially each node was either in a resting state or in a pro-seizure state. And when enough nodes flipped, reached their threshold. So this is an example of a model that might be able to predict an alternative approach in cases of failed surgery. And this is done retrospectively. And I'll talk about this in a minute, what the limitations of this type of work are. Here's an example from the DBS literature. And many of you know that Cameron McIntyre has been working on this kind of thing for a long time in his lab. And this is a paper of theirs from 2015 that talks about this approach to model outcome that incorporates kind of all parts of the process of DBS. And they've called this clinical decision support system. So this is a schematic from their paper. And it starts with the patient, of course. Here's a database where you have all of the important variables. This is a pretty simple concept. You feed this into some sort of model that then can inform the treating clinicians. Once the model has accumulated enough data, it has the ability to potentially predict treatment. They've set this up and described this with an approach to just randomly sample all the metadata that they have access to to try to see what pops up as a predicted response. That's one approach. And in this case, what they show is for they set up a way to view the parameter space. And when you have these areas, you may not be able to see them well, that are red, with predicted motor improvement over 65%. If you take the center of that, you end up in this parameter space. So their method then has a way to report what the simulation settings, recommended simulation settings would be to land in this efficacy area. All right. So these are two or three examples of how you might think about computational modeling and how it's being used to potentially affect outcomes. And I want to take a minute to talk about the data that goes into these models. And so both of these papers, and this is nice work. It's hard to do. And we're not really sure what the value is, you know, maybe within our own practice. We don't know how easily we might, you know, put our data into the model. And these are on just a few number of patients, and so you have to take it with a grain of salt. So how do we potentially think about solving this problem? And we've been thinking about this a lot within our own group, in just terms of how we organize our own data. There's been a lot of emphasis with recent NIH funding through the BRAIN Initiative on data sharing and data management. There have been specific RFAs to encourage people to work on standards for data formatting and storage. So I think it's useful to think about this and think about how we do this in our own groups, research programs. So we're talking about organization of information. And I want to just talk about this concept of a data warehouse, which is different than a database. So a database is something that we put information into. It's static. And we use it in retrospective fashion to figure out what we did. A data warehouse is different. It's a different type of organization scheme that constantly accepts data and in an ongoing fashion has different types of outcome that can be produced. So the sources include the electronic medical record, imaging, research data, behavioral data, et cetera. Through this extraction transformation loading process, the data is loaded into the data warehouse. So, for instance, if you think about taking data from Epic, each variable is coded in a certain way and that code is completely arbitrary. So there has to be a process to understand how the information is labeled in Epic or another part of the electronic medical record and transform that into data that's usable in the data warehouse. So this data lives together simultaneously raw data, metadata, summary data, and also processed data and preprocessed data. So data at each stage can live here. And then the architecture allows you to have ongoing reporting on what's happening. And it allows you to feed the data into analysis tools. You can get dashboards that show you the level of data accrual, reporting, and then also data mining. So there are more than two but a couple different ways to think about how the data warehouse works. One is with hypothesis-driven work where we would put data into specific analysis tools to answer specific hypothesis-driven questions. You can also use this architecture for data mining. And finally, just again to mention this concept of standardization of data format. This is a real problem and issue that all of us address. We try to share data between sites. And I think this is something that we're going to hear more and more about. And I think I'll just leave it at that. I'd be happy to take any questions. Thank you.

computational modeling

neurosurgery outcomes

physiology model

data complexity

clinical decisions