Junior Awardee: Can cortical signals tell us what is happening in the deep brain? | Congress 2026

So this is a translational study exploring whether activity recorded from the cortex of the brain can be used to reconstruct what is happening in the deeper brain structures, and this might potentially address some of the limitation of sensing in current systems of adaptive deep brain stimulation.

Zixiao, congratulations on the award, and welcome to the podcast.

Dr. Zixiao Yin: Thank you. Thank you for having me.

Dr. Michele Matarazzo: I would like to [00:01:00] start with the clinical problem that this work wants to solve. Adaptive DBS is increasingly moving us from simply delivering simulation towards system that can both listen to the brain and respond to the brain and to its activity, of course. But your study starts from the premise that directly recording from the deep parts of the brain is not always as reliable as we would like.

What are currently the main limitation of subcortical sensing with current DBS systems, and what made you think that cortical activity might provide an alternative window into what is happening also in the deep parts of the brain?

Dr. Zixiao Yin: Yeah, that's a very good question. So as you said, deep brain stimulation in current days is moving from a mode of always on to the adaptive mode. So the device can listen to the brain and adjust stimulation in real time. But, our work is trying to address the problem of the listening side.

 So the listening [00:02:00] can be the fragile part. The same electrode as we know that delivers 130 hertz stimulation has also to record those very tiny microvolt scale signals, and the stimulation artifacts can distort it and the hardware can sometimes fail. The current system give you only one or two bipolar channels to record the signals from the deep brain.

So those are all limitations of the current sensing systems. That's the reason why we want to improve it, to make it more robust, to make the closed loop system. And cortical sensing is one way that we wanna propose.

Dr. Michele Matarazzo: And at least in your experience, in practical terms, how often do these artifacts or signal dropout, or the limitation, h-how often they are a meaningful problem for adaptive DBS currently?

Dr. Zixiao Yin: I think in the clinic, as, for adaptive DBS the Medtronic Percept device is entering the clinic for formally as a commercial device. And [00:03:00] in China, we have also some other brand like PINS. I would say that stimulation can oftentimes cause like distortion in signal quality and some other like movement artifacts can also be problematic.

And I think those are all problems that will reduce the quality of the signal. 

Dr. Michele Matarazzo: Now, one interesting aspect of your study is also the dataset, the scale of the dataset. We're talking about 723 hours of simultaneous cortical and subcortical recordings. If I'm not wrong, it was 49 patients across three centers Berlin, Beijing, and San Francisco.

And importantly at least what I read from your abstract these recordings cover different behavioral states, treatment condition, brain targets, and even different movement disorders. Can you walk us through this dataset? What exactly were you recording from the cortex, and what were you recording from the deep brain, and under [00:04:00] what conditions?

Dr. Zixiao Yin: Yeah, that's correct. First, I would like to appreciate my mentor, Wolf-Julian Neumann, who's now in Switzerland, but he was in Berlin, and he organized such a huge data set including patients in three continents, like in the US, in Europe, and in China. If we really want this cortical decoding deep brain signals to work, this decoder to work, we really need all sorts of behaviors and all sorts of patients, even d-different diagnosis.

So here in this data set, we record patients with Parkinson's disease and patients with dystonia, and even patients with Tourette syndrome, and we record them while they are doing hand rotations, that is movement, and while they were asleep. So we have awake and asleep state. And also, since patients with deep brain stimulation device they have therapies, for example, medications and stimulation.

So we also explored whether our decoder can [00:05:00] work in different therapeutic states like on/off STN and on/off medication. So this is how our data set covers.

Dr. Michele Matarazzo: Pretty impressive. Now, with all this information and this data, you developed a convolutional neural network called CTXNet which takes cortical signals and predicts subcortical activity. Can you explain so that we can all understand this, knowing that we're not all expert in this kind of modeling?

What information does the model receive, and what are you asking the model to predict?

Dr. Zixiao Yin: The model receives information from the cortex. We use this ECoG, electrocorticography, to sensing the cortical signals. Basically those are brain waveforms computed after some pre-processing, and then we get all sorts of information like spectral band and waveform features that then we can feed those features into this decoder.

And [00:06:00] a-as you said for the convolutional neural network part we try to rebuild the spectrum features of the deep brain. So we're at that stage, we will not re-regenerate the full signal of the deep brain, but we just wanna reconstruct the, for example, beta power along time in the deep brain, in the subthalamic nucleus.

That's what was the input and output of that data.

Dr. Michele Matarazzo: And you report that this convolutional neural network approach performed approximately 40% better than the conventional spectral feature approaches, at least for decoding STN band activity. What does this tell us? What is this 40% improvement? Is that enough? Is that accurate? 

Dr. Zixiao Yin: Yeah, that's an interesting question. So 40%, what we are comparing against is the conventional like machine learning models, because first we wanna try to build the mapping relationship between the cortex and the deep brain. So first we use those machine learning [00:07:00] models, taking conventional time series or spectrum features into the model and then try to predict the beta power in the deep brain.

But the problem with that model is that not everyone is an expert in this field, so we don't always know what's the most in-informative features that should be input into the model. So the power of convolutional neural network is that it takes raw signals from the cortex, and then it will automatically pick the most informative features that it generated from those raw signals and adjust their weights to the neural network and then give output.

So that's, I think that's how this 40% of improvement comes out. And in terms of how high it is because overall I have to admit that the overall Pearson correlation between the reconstructed beta power and the actual beta power the correlation coefficient is around 0.3 to 0.5 range.

That's not really high, but we [00:08:00] can also confirm that from this set of accuracy, we can also reliably detect like beta burst the pathological burst activities in the deep brain of of Parkinson's disease, we say around 70% of accuracy. There are still rooms for improvement, but this is more like a proof of concept.

Dr. Michele Matarazzo: Great. Now the other thing that you did in your study, because, predicting a single biomarker such as the beta activity is one thing, but you also used the denoising diffusion probabilistic models to reconstruct the raw deep brain signal itself. Can you explain what this mean and how this works?

Dr. Zixiao Yin: I think that's the most exciting part of this work. Previously we only wish that if we can rebuild the beta, that's already amazing. That's already quite strong decoder. And then we want one step further to see whether it can reconstruct the full signal recorded from the deep brain.

The [00:09:00] problem with decoding only one band is that we need one separate decoder for one band. That means if we want to decode beta, then we need a beta decoder. If we want to decode alpha we then need an alpha decoder. But if we can reconstruct the raw signal, then we can take that signal and compute whatever bands that we want.

So the DDPM, the denoising diffusion probabilistic model that you mentioned is like the model behind the current ages of image generator. So it doesn't like a regression model generate one point after another. It generate a series of numbers, of values.

So that's the raw STN signals. 

Dr. Michele Matarazzo: Great. And as you were mentioning, you went a step further also in the sense that you were not just comparing this with the actual deep brain activity, but you also studied whether this could predict motor severity, right?

Dr. Zixiao Yin: Like we take the predicted STN power, STN raw signal, [00:10:00] then computed beta power from those im-imputed signals. And then we use those features to predict the patient's UPDRS motor severity score, and we can see that the actual UPDRS value versus our predicted UPDRS value the correlation is around 0.7.

That's quite high. I need to clarify something here that as I said, that the beta power correlation between the real signal and the imputed one is around 0.3 to 0.5, but that's within patient. So the patient's beta power oscillate, but it's still around one a level, a certain level of the subject.

But if we do this across subject, each subject's baseline level is more or less it's different. And the DDPM and the neural network models can more accurately capture those cross-subject differences better than the within-subject differences. That's the point.

Dr. Michele Matarazzo: Okay. [00:11:00] Now going back from when we started this interview talking about what is the clinical problem that we want to solve. Now that we know a little more about your data let's try to imagine a realistic clinical application. In the abstract, you suggest that this could eventually contribute to non-invasive closed loop neuromodulation using scalp EEG.

And there are really two different translational steps here. First, using implanted cortical recording to augment or even, or maybe rescue subcortical sensing, and second, potentially replacing implanted cortical sensing with something less invasive, something, let's say, that you can have outside of the skull.

Which of those two application do you think is realistically closest to clinical implementation or which of these two different approaches do you think would be more valuable in clinical practice?

Dr. Zixiao Yin: Yeah that's a very good question. To be honest, I think both of them are very valuable. And in terms of which one is closer to the [00:12:00] clinical application, I think it's the first one to rescue the deep brain sensing. Because for example, if we can have a an alg- algorithm first, it detects the artifacts in the deep brain, and when it sense that there is a strong artifact that may hamper the deep brain sensing as a feedback signal for closed loop system, then it can switch from the deep brain sensing to the cortical sensing, and then the cortical sensing itself is already valuable, and then it can also impute deep brain signals and a-add that as a feedback signal.

And altogether, this will make the closed loop system more robust. And your second point is that maybe we can use for example, scalp EEG to guide those non-invasive stimulation therapies. I think that's also a good point. For example, now we have those temporal interferers or focused ultrasound that can apply stimulation to the deep brain areas.

If [00:13:00] we really also want to make those therapies closed loop we also need a feedback signal. And in many cases, for example, for Parkinson's disease, the pathological oscillation, the robust one, a-as we established now is the beta power in the STN. So if we can put a, not necessarily scalp EEG, but can also be ECoG, and then sensing the cortical signals and use that to also impute signals to get information about the deep brain state, and then we can use that to guide the delivery of non-invasive stimulation.

Then we don't need to implant an electrode into the deep brain only for the sensing reason.

Dr. Michele Matarazzo: Great. Zixiao, let me end with the last question, which is: What are you planning to do now? I guess you're still working in the area. You're planning to study further how to use this in clinical practice. What is the future-- your future as a young investigator, and what is the future of this technology?

Dr. Zixiao Yin: Yeah, thanks for the question. I [00:14:00] think there are still several limitations of this technique. As I said, I think currently, its current stage is more like a proof of concept, but if we really want to build it in a real-time system with very low delay, and when the implantable device, they also have very low super tiny memory. We cannot put in a large model in it, so we need to minimize the model and also to reduce the delay of the computation.

I think that's the next steps we will try to put this technique into the clinical setting. In addition, my current work is also on-- So this is the sensing side. As I said, this technique can be used to augment state detection. For example we found that with this imputation from the cortex the movement state or sleep state in some s- patients the decoding accuracy can be improved from around 60% to around 80% for example for the sleep stage classification problem.

So I think this is one side [00:15:00] of the closed loop system, the listening side, and also we are working currently on the stimulation side that is how to best stimulate the brain when we detect a certain state. Yeah like how should we best apply stimulation when the patient is during sleep, non-REM sleep.

But that's another story. 

Dr. Michele Matarazzo: We'll look forward to listen to that part in-- of the story in the future. Zixiao, congratulations again on receiving the Junior Award, and thank you for taking us through this very interesting work.

Dr. Zixiao Yin: Thank you. Thank you. 

Dr. Michele Matarazzo: So for those of you attending the MDS Congress in Seoul, I encourage you to look out for the poster and also this will be presented in a plenary session. Remember the poster is titled "Decoding the Brain Activity from Cortical Signals Using Deep Learning Across Movement Disorders."

Thank you all for joining.

 [00:16:00]