The city of Montreal has a strong history in epilepsy research. In the middle of the 20th century, a notable surgical procedure, the “Montreal Procedure”, was developped by Dr. Penfield at McGill university. This procedure allowed to cure nearly 50% of the patients that received it. Since then, lots of research has been accomplished regarding epilepsy. More recently, electroencephalogram (EEG) acquisition and analysis technologies have emerged as promising tools in the clinical word. Those technologies use EEG data acquired with cleverly placed electrodes on or inside the head.
Being particularly interested in that field, I have had the opportunity of meeting with Dr. Mohamad Sawan, professor and researcher at Polytechnique Montréal, along with one of his students, Elie Bou Assi. Dr. Sawan created Polystim in 1994, a world-renowned laboratory that has the mission of developing smart medical devices and applications.
Recently, Elie and Dr. Sawan have published a paper, “A Functional-genetic Scheme for Seizure Forecasting in Canine Epilepsy”, about using intracranial electrodes to predict seizures in epileptic dogs. Their research could eventually be translated to human subjects and could greatly improve intervention schemes in people whose epilepsy cannot be cured using traditional methods. This specific research was the subject of our interview.
From localizing the epileptic focus to predicting seizures
When asked about the motivation of his lab to work on this subject, Dr. Sawan said that before working on this subject, they had already been working for almost 6-7 years on localizing and disrupting epileptic seizures. “We wanted to localize where an epileptic seizure starts, and we were wondering if we could stop it before it starts”, Dr. Sawan mentions. For these works, he mentions that they were more interested in the electronic components (dedicated software and hardware) rather than the signal analysis component. “Regarding epilepsy, there is localization, there is early detection to attempt to stop it, and there is prediction”, he adds. Early works involved using non-invasive EEG electrodes to localize epileptic generators inside the brain. However, these electrodes were limited in precision, hence why they eventually decided to develop more invasive implants to further increase localization, but also to deliver electronic stimulations to prevent epilepsy from occurring when it was detected.
“After that, my team and I told ourselves that localization with EEG is all well, but could we do more? And that is when Elie came in. Could we do more than localization; prediction for example?” he told me.
Localizing epilepsy: A complex task
Obviously, predicting seizures with great accuracy cannot be achieved with simple signal analysis. The algorithm developed by Elie, which can accurately predict seizures more than 30 minutes in advance, uses both seizure localization as well as machine learning.
One of the difficulties regarding localization, Sawan says, is that “you have the impression it is here when, in reality, it is there. It is similar to forest fires – where was the original spark?”.
To accomplish all of this, they used open access data that contained intracranial EEG recordings of various epileptic dogs. Elie mentioned using concepts similar to those employed in connectomics to model neuronal pathways inside the brain. This is generally realized with MRI scans, however, in this study, something similar is achieved using a set of 16 electrodes. More precisely, they use a method called the directed transfer function (DTF), that they had worked on in previous research. Elie explains, “There are various methods to create a neuronal network: for example, bivariate methods, such as correlation and synchronization, which attempt to find similarities between two electrodes. The advantage of DTF is that we can directly fit a multivariate model, which means we don’t have to repeat the task for every pair of electrodes possible. The advantage, statistically speaking, is that when we repeat the experiment, there is a high probability of having ‘lucky’ correlations, which is not the case when using DTF since we fit only one time. DTF is also a measure of causality – we don’t see the similarity between signals, but rather we see if one signal can predict future signals”. However, there are some drawbacks when using DTF, as it requires stationary signals and it has trouble predicting when an epileptic focus is transferred to another location which itself regenerates the seizure. “We only see one generator and cannot know if it’s this one or another that is the true source of the seizure”, he says. That is why they introduced an adaptive method to increase source detection accuracy in previous works to go along with DTF.
Therefore, the DTF method is a method which allows them to find which electrodes are useful to localize and detect seizures, and it allows to remove “useless” electrodes from the analysis. More specifically, DTF enables the identification of which electrodes “receive” a lot of information from other electrodes and which electrodes “send” a lot of information to other electrodes, and therefore, which electrodes are of interest for the analysis.
Information about the transmitted and received signals from each electrode is then fed to a k-means clustering algorithm, which automatically classifies electrodes as relevant or irrelevant based on a threshold. “It is the state-of-art method in non-supervised learning, and it is adaptive”, so it finds the thresholds itself, Elie adds. He also mentions that given that the source of the seizures doesn’t change for an individual, there are advantages to performing such electrode pre-selection. They, however, haven’t looked in cases where individuals may have multiple seizure sources. He says that “they can still find out the sources even if they are multiple, and that doesn’t change the fact that they do not change either for an individual”.
The figure on the right shows an example of those electrode relations within a dog’s brain. The higher the value (red), the stronger the interaction. In that particular figure, we observe that the electrode labeled “11” has very strong causal interaction with 5 electrodes and medium causal interaction with 3 other electrodes. Electrode “11” can therefore be used to predict signals from these 8 other electrodes.
Artificial intelligence in live brains
Obviously, localizing seizure sources do not suffice to predict seizures. To achieve prediction, they use signals from the previously selected electrodes. From each signal, they extract different features. However, much like there are good and bad electrodes, there are good and bad features. These features are given to a genetic algorithm to find out which one are relevant to determining what they call the “preictal state”, that is, the period before a seizure occurs. Combined with the electrode selection, Elie says that this allows the algorithm to be very patient-specific, meaning that it considers the differences that exist between subjects. He also states that by preselecting electrodes, genetic algorithms become much less computationally heavy and therefore more efficient to use, as “genetic algorithms are like heuristic searches in that, when you reduce information they don’t need, they converge faster”.
Once the relevant features have been determined, they send their information to one final layer, a support vector machine (SVM) classifier, which seeks to separate classes with a linear hyperplane. The SVM gives a probability that, at any given time, a patient is in a state preceding a seizure. However, the algorithm can produce false alarms. “This is a big issue. False alarms (i.e. false positives) mean that you detect a seizure when there are none. Some people claim they detect 100% of the false alarms, but that’s impossible”, claims Dr. Sawan. To reduce the number of false alarms, they regularize the SVM classifier output using a technique called “firing power” (FP), which is nothing more than calculating the temporal moving average of the classifier’s output. When the FP signal is superior to a set threshold, the patient is in preictal state. The figure on the left, from their article, shows that. The false alarms would occur when the SVM output crosses the threshold, and the yellow region represents the length of the preictal state.
However, Elie mentioned that recent studies have shown that those false alarms were more like putative alarms, that is, when the brain prepares itself to generate a seizure but then restrains it.
Promising results
The algorithm that Elie developed as part of his research has shown promising results. Indeed, they achieved a sensitivity of almost 85% when using a 1-hour preictal time and a 5-minutes intervention time. That later corresponds to the time needed for intervention. “Those are conventions in the domain. We have interictal, which is far from a crisis, there is ictal, which is during the crisis, there is the preictal, so just before a crisis, and there is postictal, which is when the seizure is ending”, says Elie, adding that “this has been put in place to determine that we are truly predicting and not just detecting crisis. The intervention time is the minimal time to intervene if we predict a crisis. Here, we need to predict 5 minutes before the onset”. I asked him if 5 minutes was not too small, and he responded that this is sufficient to tell someone on the verge of a seizure to stop driving and park himself, for instance.
However, he says that “one of the most disturbing thing to tell a patient is to inform him about an incoming seizure without being sure when it will occur, because this can create stress and anxiety”. Therefore, this psychological aspect of the patient is to be considered when developing seizure predicting device. Patients should not be informed too long in advance, but just enough so that they can stop activities (such as driving) which could become dangerous during a crisis.
It is important to recall that this research was done on dogs. Dr. Sawan told me that “the idea is to build a smart implant which informs about preictal state and stimulate seizure sources to prevent epilepsy crisis”. He adds that implantable brain electronics are now safe enough to wear long term.
Next steps
When I asked him about what was to come following this research, Dr. Sawan mentioned that they were working on their own hardware, building an implantable system-on-chip that performs the required signal analysis.
Overall, it was a very interesting interview that allowed me to learn several new things, but also how things I already knew about, such as connectomics and machine learning, can inspire new technologies to hopefully help increase the quality of life of people with various medical conditions. I notably learned a lot about seizures in general and how they are triggered. I was also surprised to see that deep intracranial electrodes weren’t needed to achieve accurate seizure forecasting, as this experiment used surface intracranial electrodes.
Research in seizure forecasting allows the development of new tools to help medical staff and researchers to better understand underlying epilepsy mechanics. Advances could therefore allow the development of permanent wearable devices that can be implemented on epileptic patients to warn them about upcoming seizures and to increase treatment efficiency by delivering electrical impulses to prevent them from occuring. Excitement is real when it comes to smart medical device advancements!
Quotation and publication approval
The responses in this article have been translated, slightly edited and condensed for clarity. Professor Mohamad Sawan and Elie Bou Assi have both given their approval for the publication of this article.
References
[1] E. Bou Assi, D. K. Nguyen, S. Rihana and M. Sawan, “A Functional-genetic Scheme for Seizure Forecasting in Canine Epilepsy,” in IEEE Transactions on Biomedical Engineering. doi: 10.1109/TBME.2017.2752081