How AI can improve the quality of life of patients with epilepsy
Epilepsy is a disease that is difficult to cure or treat. Epileptic seizures have so far been difficult to predict, so it is also challenging to gather relevant information about the seizures without keeping the patients in the hospital for an endless period of time. That is why the Flanders AI Research Program (FAIR) started a use case four years ago to enable mobile follow-up of epilepsy patients. This way, doctors can learn much more about the disease in a shorter period of time and can respond more quickly to treatment.

Epilepsy
Epilepsy is a neurological disease that affects nearly 65 million people worldwide. The treatment of epilepsy has improved considerably in recent years, but 35% of patients still suffer from epileptic seizures for the rest of their lives.
The unpredictability of the disease and, in particular, of the exact moments when they will have an attack, has a huge impact on their quality of life. Ordinary actions such as driving a car or spending a day away from everyone suddenly become potentially life-threatening actions.
Need for more and more accurate data
Epilepsy patients would benefit greatly from a solution that automatically sends an alert to patients, friends and family when the signs of an epileptic seizure are observed.
They would also benefit from a system that objectively and automatically measures seizures and transmits these results to the treating physician. After all, when the doctor depends on the patient's own observations, they are rarely accurate. Greater correctness could greatly increase the quality of the treatment.
In short: patients and the medical world need a system that can correctly detect and describe epileptic seizures. This is now possible when the patient is in the hospital, but until now there has been no reliable solution systematically available for daily use that can generate sufficient (reliable) data on which to base predictions or treatments.
Widely rolled out to patients
In the meantime, there are wearable sensors on the market that continuously measure relevant values, even in patients in their daily lives. FAIR’s research also uses such sensors, at a fairly advanced level, according to Maarten Devos, project leader at Stadius, KU Leuven Centre for Dynamic Systems, Signal Processing and Data Analytics: "With every large study, you first have a 'proof of concept' (can we measure and analyse something relevant), then the validation in a larger study, with research in the hospital itself (in this case comparing data from sensors in the form of stickers stuck behind the ear with the classic EEG), and thirdly the use of these sensors by patients at home and on the move (first these stickers are applied by the medical professionals, then the patients do it themselves). Nowadays we are already in phase 3: we already have several patients that we measure almost continuously for 3 to 6 months. In the meantime, there are already more than 500 patients that we have measured."
"We can give people a sensor when they are first diagnosed, so that we can immediately measure the impact of the medication. This has two major advantages. First, we can start with a low dose and, if the medication doesn't work, immediately increase the dose. Secondly, we shouldn't wait six weeks to try something new if you already know after a week that the prescribed medication doesn't work."
In the meantime, the technology is also advanced enough to monitor children with these sensors. Because children often find monitoring with an EEG too frightening, the arrival of these sensors is an enormous added value for the follow-up and treatment of children.
The research
The data collected by these sensors only becomes valuable when it can be analyzed, and leads to new insights and even predictive value. That is the purpose of this use case from the Flanders AI research program: KU Leuven, UGent and UHasselt are working together with UZ Leuven on the development of an AI method that predicts as accurately as possible when a patient will have another epileptic seizure or when the risk of a seizure increases significantly. The methodology used integrates cutting-edge noise removal techniques, automatic feature extraction, and multi-view approaches to achieve optimal performance. Also, the built-in uncertainty quantification and the "white-box" approach ensures that the physician can trust the AI and that the total solution makes the most of physician's complementary expertise and automatic modeling. With this information, the patient can monitor himself better in the long term and the treating neurologist can get a better view of the patient's medical condition.
Increasingly complex and multimodal
Maarten De Vos, project leader at Stadius, KU Leuven Centre for Dynamic Systems, Signal Processing and Data Analytics: "My predecessor (Sabine Van Huffel) was a pioneer in biomedical research, and she also worked on signal processing and decision support. But that was based on simpler models, which required much less computing power and less data. Back then, we had to rely much more on human insights and interpretation. Now we can process a lot more data, and with more complex models, and that has given a boost to our research."
"Today, we work much more multimodally. We try to integrate data from different types of sources (EEG, heart rhythms, activity, etc.) and over a longer period of time. Then you really create added value: if you can use data from six weeks of continuous measurement instead of just the data from the patient visit, that makes a big difference. At the same time, of course, this also creates such an amount of data that it can no longer be processed by humans alone. So you also need powerful AI systems to search through this data for relevant information and analyze it. Until recently, they simply weren't available."
It's not just the amount of data that's a challenge for this research, Maarten continues: "The automatic detection of epileptic seizures is not an easy task either: signals are often of poor quality, the datasets are highly unbalanced because epileptic seizures occur very irregularly, the results vary greatly from patient to patient, and there are many different types of epileptic seizures. All of this made it very difficult to make rapid progress. But with the improvement of the available infrastructure, and with the ever-larger datasets that we are building up along the way, the steps we are taking are getting bigger and bigger."
In addition to the infrastructure and the larger datasets that lead to better results, there are other factors that contribute to the success, Maarten explains about this project: "For example, we have rolled out a learning strategy that no longer depends on a complete and time-consuming annotation of the portable data. We have also been able to show that the integration of multiple types of portable data (EEG and ECG) leads to an 11% better detection. And we can build new decision support systems for this application faster thanks to automated development tools."
The entire cycle is not yet fully automatic, Maarten nuances: "So-called candidate signals - signals that may indicate an attack - are detected and presented to the doctors. They still have to validate the detected signals. Still, this is already a huge time saver. To manually analyze a classic EEG you sometimes need up to 2 hours. Now the doctors can process all relevant information in 5 to 10 minutes, partly because the candidate signals are also provided with an estimate of how (un)sure we are that it really is an attack."
The detection of signals and the validation of these signals is still an ongoing interaction between the medical and technical teams, which should lead to the improvement of the AI models used for this, Maarten notes: "And, even though it still remains a 'work in progress', we have already come a long way. In one particular analysis, we once got a sample from a patient and we didn't find a single seizure. We were already starting to doubt our solution, until it turned out that the patient had indeed not had any more seizures in the meantime. So we've come a long way."
The Importance of Collaboration in the AI Ecosystem
Maarten: "Our strength lies in the interdisciplinary: you need experts in the medical world, top engineers, and a supplier that can make good medical devices. Every 1 to 2 weeks we meet with all parties involved, and there are a lot of them. There is also a lot involved in such an interdisciplinary approach: quality of the model, quality of the different types of sensor data (brain, heart, movement, etc.), and the integration of all that data."
"We also depend on the experts' assessment, and that is not always obvious. Sometimes doctors disagree on how long an epileptic seizure lasted and when exactly it started. That makes it difficult to label the data. The interaction between our perception and the correct interpretation of the doctors should ensure that the algorithm becomes stronger and stronger."
What can we expect in the future?
"In the near future, we want to be able to validate the existing approach on much larger datasets and on more datasets from 'the real world', in realistic but more challenging home environment of the patient," says Maarten, "and we continue to look for ways to integrate additional contextual information, such as the medication or the patient's sleep patterns, so that we can make the existing algorithms even better."
This should lead to a fuller life for the patient and at the same time better insights for the treating physicians, Maarten expects: "In time, almost all on-site measurements in the hospital will be replaced by mobile measurements. This will generate huge amounts of extra data and information, and improve the algorithms at an even faster pace. For example, it is impossible to capture data about brain function when a patient participates in extreme sports with a fixed device."
"We will be able to better detect and use the 'biomarkers', the indicators that suggest that an attack is imminent, so that we can predict the arrival of an attack fairly accurately."
But predicting this is not for tomorrow, Maarten nuances: "Warning of imminent attacks is more difficult than detecting them. It's like a weather forecast: you may be able to predict that one is coming soon, but be able to predict that you're going to have an attack tomorrow at 11 a.m.? We're not there yet."
"There are so many factors that determine whether or not you have a seizure, and when exactly it will happen. In some patients, you see patterns where they always have the seizures at the same time of day, and then the probability that your prediction will come true becomes a bit higher. But you can't correctly predict every attack for everyone. Brains are complex objects and you can't map them very well yet."

For the impact of this new solution on patients, we listened to Lauren Swinnen, doctoral researcher in epilepsy at KU Leuven:
"The lives of patients with epilepsy are heavily influenced by their condition, on the one hand because of the unpredictability of the seizures and on the other hand because of comorbidities and stress. In the questionnaires of our study, everyone indicated that they would very much like to become seizure-free and to be able to do daily things, such as driving, again. As a result, they were also very motivated to use the sensors. Thanks to these sensors, patients can send accurate and objective data of their seizures and their frequency to their neurologist from the comfort of their own home. Based on this, the neurologist can adjust the medication if desired and analyze the effect on the seizures over time, with the ultimate goal of achieving seizure freedom."
"Patients have different opinions about the comfort of the solution: Most find it a big improvement compared to being in the hospital for an EEG. Some still find it too visible, which is why we offer them the autonomy to partly choose when to wear the sensor. This way, the patient's quality of life is preserved, especially compared to the uncertainty of living with epilepsy." In addition to the direct impact on the patient, the solution can also be used in clinical trials for the development of new antiepileptic drugs, where it can provide an objective picture of seizure frequency under that particular medication. This way, it also indirectly contributes to the quality of life of the patients."

The overview below offers a look at the technical aspects of the project and the partners who took care of each aspect.
The quality of the biomedical signals is improved with traditional noise removal techniques. In addition, the multimodal algorithms can make use of the various sensor modalities. The noise quantifiation models are trained on the device and adjusted with an (un)supervised transfer learning scheme. The automatic segmentation and feature extraction was greatly improved by using the multimodal structure of the data and sensor-based data fusion (KU Leuven-STADIUS).
The multimodal classification was improved with multi-level and multi-view models and multi-class classification (UGent-KERMIT), and with the addition of expert knowledge in the model (KU Leuven-STADIUS), context-awareness (which allows the model to adapt to different situations such as awake/sleeping, stress, active) (UGent-IDlab). By making the model interpretable, the patient can gain more insight into the causes and possible triggers for the attacks (such as stress or poor sleep).
All pictures (c) Laboratorium voor Epilepsie Onderzoek UZ KU Leuven


