Hello everyone and welcome to a new episode of the MDS podcast, the official podcast of the International Parkinson and Movement Disorder Society. Today we're gonna discuss the paper titled, the Role of Levodopa Challenge in Predicting the Outcome of Subthalamic Deep Brain Stimulation, which was recently published in Movement Disorders Clinical Practice.
I have the pleasure to have two of the authors here the junior author, Dr. Robin Wolke and the senior author professor Günther Deuschl, both from the University in Kiel Germany. Welcome and thank you very much for taking part in the podcast.
View complete transcript
[00:00:41] Robin Wolke: Thank you very much for having us.
[00:00:43] Eduardo Fernandez: For those listeners that are not familiar with the deep brain stimulation of the subthalamic nucleus, could you give us an overview of what are the main indications for this procedure and what are the main selection criteria for patients, the main inclusion and exclusion criteria?
[00:00:59] Günther Deuschl: Well, [00:01:00] DBS is not a new therapy. The first patient who operated 30 years ago almost, it's a therapy that has been investigated indeed over these 30 years. There were hundreds of controlled studies, and the first double-blind just appeared in 2020 on this treatment and they converge all on the conclusion that several patients have a very good outcome or can have a very good outcome. These are patients with advanced Parkinson's disease with fluctuations. So the fluctuations are one of the most important criteria to include a patient. The patients should have a good on.
So in their on state that is about what can be achieved with deep brain stimulation should be in the order of stage two about of Parkinson's disease. The treatment does not only help with the motor part, it's also non-motor symptoms are improved to a certain [00:02:00] extent.
And it's well known that quality of life can improve dramatically with deep brain stimulation. And as a rule of thumb, a patient who receives deep brain stimulation and is well qualified for this treatment can gain about five years in his disease course when he or she are carefully selected.
Therefore, selection plays a major role.
[00:02:25] Eduardo Fernandez: I mean, as you said, selection is key in the outcome of deep brain stimulation and, there has been for many years that idea that levodopa responsiveness is gonna determine the outcome of DBS. But this idea has been questioned and your manuscript raised questions about the best selection or how determine levodopa responsiveness in candidates for DBS surgery.
So in your manuscript, you have analyzed data from three different centers from Kiel and Berlin in Germany and from Toronto in Canada, assessing [00:03:00] how Levodopa challenge can predict the outcome in people with subthalamic nucleus DBS. Can you give us an overview of the main results of the study and also touching on the methods as well discussing a bit the sophisticated statistical methods that you have used to reach this conclusions.
[00:03:18] Günther Deuschl: The Levodopa test has become a standard test since the first publication in 2002 by Charles and colleagues from the Grand Noble Group. I think worldwide, almost every center uses this test as an orientation. Whether a patient qualifies for this, or not.
There needs to be an objective proof. And patients can tell you about their levodopa responsiveness during the, taking the medical history. And this is already in a very good orientation, but when you want to have closer data, clear data that can be documented than you need to have the patient assessed. This is done [00:04:00] almost in every DBS center and therefore data are available. And we joined together with Andrea Kühn from Berlin and Alfonso Fasano from Toronto, and also Hagai Bergman, who has been involved in these questions since the late eighties.
Whether the finding that was found so often that cohorts show a very close correlation when you look at cohorts, a very close correlation between the Levodopa test and the final outcome of the DBS surgery, there is a high correlation. Okay, but what does this say about an individual patient?
And that was the question. So we would like to see a response that tells us, oh, this patient with a statistical reliability and outcome that lies between this and that. And that would be excellent for advising the patient.
We had these many data from [00:05:00] Berlin, Toronto, and Kiel and that was the basis for these assessments. And I think how the transfer from the cohort of patients to an individual patient can be done statistically. That's probably something where you can comment on for bit.
[00:05:19] Robin Wolke: We have this very big data set consisting of 429 patients that underwent surgery STNDBS surgery in three centers. And these patients also had early follow up UPDRS testing available. Early follow up means there was a mean of 9.15 months in between the operation and the follow up.
And what we did first is exploring the dataset. First you look at the data and obviously what you wanna do is try to reproduce the correlation that was found so many times before. So we correlated the UPDRS-III reduction during the levodopa challenge with [00:06:00] the absolute UPDRS-III reduction due to the DBS stimulation at early follow up. And we found significant correlation that actually almost exactly matches that one that was found by Charlotte Zar. However, we were correlating further variables. So what do we have available?
We had available also the UPDRS-III without medication before the operation. So the UPDRS-III in the off stage at baseline. And this score also correlates significantly with the outcome. So the UPDRS-III reduction after surgery in the early follow up so now we have three variables it seem to correlate, with each other actually.
And in order to investigate which variable is more important in this triangle, we fitted a multivariate linear model just to describe the data set first, there's [00:07:00] nothing about prediction yet. So what we found is that the preoperative UPDRS-III in the off state seems to be more important for explaining this data set we have then the UPDRS-III improvements due to levodopa. That's quite notable and if you look then at the relative levodopa improvement and compared to the relative improvement to DBS stimulation, relative means relative to the UPDRS-III in mid off at baseline.
And then this correlation that was found earlier actually shrinks very much, it's still significant, but it's very slight.
[00:07:41] Eduardo Fernandez: So if I understand well, that means that this disease severity measured by the UPDRS part three at baseline before the DBS is what it determines mostly the outcome after surgery. Is that correct?
[00:07:55] Robin Wolke: That's a correct conclusion that the same conclusion we also had. [00:08:00] So how can we imagine this? Probably those patients which suffer more severely from the disease have higher UPDRS-III at baseline also have more room for improvement due to stimulation. This was actually the idea we had looking at this data.
[00:08:16] Eduardo Fernandez: And you mentioned now moving into the prediction model. How can we predict in an individual how much he's gonna improve with deep brain stimulation based on the assessment with the Levodopa challenge.
[00:08:29] Robin Wolke: Should clarify what's the difference between this explanatory analysis we just explained and the predictive analysis? So the explanatory analysis is solely centered around the data set that we have. There's no new data that comes in and that we somehow can forecast.
So prediction has a goal to forecast data that is unseen, that is not known yet. So patients come to the clinic and I want to predict how will the DBS outcome be? So for this, [00:09:00] there are several approaches. Either I train a model like we just did and predict the outcome of the new patient coming into our department.
However, a validation like this would take a lot of time and it might eventually fail. So it's quite unethical. So what we did is we used cross validation to determine the predictive power of our models. Cross validation means that the whole data set was cut into 10 equal parts, and this 10 times.
So ending up with a hundred chunks of data and each time we trained a model on nine parts of this data and validated it, on the tense part. So like this if you then regard the model it fits, you get an estimate on how your model performs on unseen data on which it has not been trained on.
So which models did we actually train? So there were several options. You could either try to [00:10:00] predict the continuous variable of UPDRS III reduction due to SDNDBS. So in this case, we trained general linear model, that was trying to predict the continuous variable.
We also trained more sophisticated algorithms, which is a gradient boosting algorithm called X Extreme boost and support vector machine with supporting nominal cannel. All those different models actually perform quite likewise resulting when trying to predict the UPDRS III reduction in absolute terms.
Due to operation and a mean r squared of 0.41 was quite a large range of this R squared from 0.35 to 0.51. And likewise, we tried to predict the relative stimulation relative to the UPDRS III baseline and this model, all performed very badly. So they drop down to [00:11:00] 0.14.
And another thing you could do with trying to improve the prediction for a person. The patient's not gonna ask you, I want to know how my UPDRS III improves. They wanna know, will I improve or will I not improve?
What we did, we defined for the UPDRS III in total that a reduction of greater than 33% of one third would be a reasonable cutoff or the minimal clinically important change of five points in the UPDRS III score. This we also extended by regarding different sub scores by regarding the tremor.
But maybe this goes a bit far here. So when we tried the same models on predicting the digitalized outcome also the results were quite mediocre, so to say. So the goodness of a classification model that predicts itemized outcomes, so good response, bad response can be measured with [00:12:00] different measures.
For example, accuracy. How many of the predicted patients were classified correctly, or you could use the receiver operating curve and the area under the curve. That's what we did, and we resulted in area under the curve of 0.66 for predicting the response greater than 33%.
And now what is also important to see whom do you actually want to identify? So in STNDBS luckily most patients do respond quite well. That's why it's so valuable and successful. But you do want to predict who's not going to respond sufficiently due to the risks related to the operation procedure itself.
And in our case, we looked at the median specificity of our models. This means how many non-responders were correctly identified using our model. [00:13:00] And here we only came up with a result of. 0.47. So basically we could not identify the non-responders using our models. And also this more sophisticated models or the non-linear predictive models like the gradient boosting algorithm, like the support vector machines, did not improve this results.
[00:13:24] Eduardo Fernandez: Excellent. Thank you for explaining so clearly the sophisticated analysis of the results. So those conclusions raise questions and have important clinical implications about the use of the Levodopa challenge in predicting or selecting patients for SD and DBS in the discussion, do you still think that assessing Levodopa responsiveness is useful in the clinic?
Or some may argue that probably that test is not needed, that we can determine how responsive a patient is based on the anamnesis, maybe we shouldn't use the UPDRS. Maybe we [00:14:00] should use all the tools or focus on some certain phenomenological aspects of the UPDRS to determine the outcome.
What clinical implications do you take from this result?
[00:14:11] Robin Wolke: Maybe before we answer this question, a little disclaimer. So in our data set, that's very important to me to say, there were no Levodopa non-responders. So people that did not respond to levodopa at all, or that did not have an indication for STNDBS, so we can actually not comment on this group of people, that they were not included in our data set. What we can comment on those people that had an indication for the STNDBS. So Just like a little disclaimer before we go into further discussion.
[00:14:42] Günther Deuschl: Right. But in essence, it's Levodopa challenge should be continued to be performed just because we need to exclude non-responders.
The data that we have on non-responders are limited, but I think they are sufficient to say [00:15:00] that levodopa non responders do not respond to DBS. So this is the most important point, first of all. Then we have to acknowledge that the prediction for an individual patient still remains very dubious.
So how good a patient responds is not very clearly to determine. This is already known by the DBS Centris and they have more than this. We mentioned only the levodopa test, but we said about the best on which is an important determining factor, whether you asked the patient for, or which you check. And in the future there may be more specific tests. For example, if a patient suffers, particularly from freezing, you have to look in a very detailed way how the freezing is influenced. And we know already from freezing sub studies that we have, that this is a relatively good predictor for this particular symptom.
[00:16:00] And in the future, we probably will have more of these specific subtests that we use to make a prediction of which symptom will improve or not. And that is actually the way into a more personalized medicine that we are in.
[00:16:18] Eduardo Fernandez: So exactly. I think Parkinson's disease is very heterogeneous now and we're hopefully trying to disentangle that heterogeneity and then move towards a more personalized medicine. I don't have any other questions. I dunno if you want to discuss anything else or add anything about the paper.
[00:16:36] Günther Deuschl: The conclusion is important for the people who are doing DBS. So for the physician working in clinical practice and not doing DBS, it's important that he or she cannot treat the patient adequately. And that he knows, he responds to levodopa and the best on [00:17:00] is important. For the centers, it's certainly a challenge to find better ways to assess levodopa sensitivity in the future. And further studies we'll follow to do that.
[00:17:11] Eduardo Fernandez: I will definitely have this conclusions in mind next time I see a patient and consider a referral to the DBS team. Thank you very much for your time and your discussion about the results. And thank you to the listeners and I will encourage them to read the full paper in the Movement Disorders Clinical Practice Journal. [00:18:00]