Scientists have developed a new artificial intelligence (AI) system that can analyze brain interactions to more accurately predict epileptic seizures, including rare types that are caused by bursts of abnormal electrical activity in the brain.
According to Cure Epilepsy, approximately 65 million people worldwide suffer from the neurological disorder epilepsy, with 3.4 million people living with the condition in the United States alone. The most serious complication is injury or death from seizures, with approximately 1 in 1,000 people with epilepsy dying suddenly and unexpectedly each year.
Key to preventing these deaths is better predicting the onset of epilepsy seizures, and the World Health Organisation (WHO) predicts that with better diagnosis and treatment, seven in 10 people with epilepsy could live seizure-free.
How AI can help predict epileptic seizures
AI and machine learning are already being deployed to model, detect and predict epileptic seizures through electroencephalography (EEG) data collected with electrodes on a patient's scalp. While these systems are better at finding patterns in the data than humans, they fall short in detecting rare seizures.
In a new paper published in the journal Advances in knowledge discovery and data mining Researchers from the University of Southern California (USC) describe an AI system that could overcome the challenges faced by current machine learning platforms.
“Our AI system is designed to enhance seizure detection and classification from electroencephalography EEG data,” said lead author Arash Hajisafi, a USC computer science doctoral student. “Our system utilizes a dynamic graph neural network (GNN) architecture that is carefully trained on labeled EEG recordings.”
The team's GNN framework is built using spatial relationships and descriptions of each part of the brain, including higher brain activity such as the prefrontal cortex, and these features are used to model each type of seizure.
This additional context means that the USC team's AI can learn effectively even when provided with just a few seizure samples, which is advantageous for identifying seizure types with little available data and overcomes the need for huge labeled datasets for training other deep learning architectures.
“By integrating this diverse context, our framework is able to more accurately characterize EEG recordings, which enables our models to extract meaningful patterns even with very few training samples,” says Hajisafi. “As a result, our system can generalize well to unknown cases and provides consistent performance across both common and rare seizure types.”
AI can spot rare epileptic seizures
When a 60-second EEG clip is fed into the USC team's new system, NeuroGNN, it can accurately determine whether the clip contains a seizure. If a seizure is detected, the system can classify the type of seizure, even if it is a rare type.
“Deep learning models perform well at learning patterns from large amounts of data. But when the data is imbalanced, the model tends to perform well for classes with abundant samples, but underperform for rare classes,” Hajisafi said. “This is because the model is optimized to minimize overall error, which ends up favoring more common patterns in the data. As a result, it often performs poorly for rare seizure types, where the model does not have enough samples to learn their unique characteristics.”
To test NeuroGNN, the research team took advantage of a well-known public seizure dataset called the Temple University Seizure Corpus (TUSZ), which is widely recognized in the medical field and contains separate EEG recordings for training and testing.
The research team examined 3,050 annotated seizure events from over 300 patients, covering eight different seizure types, recorded using 19 electrodes on a standard 10-20 EEG system. For their analysis, the researchers classified the seizures into four classes:
“Our approach consistently outperforms previous state-of-the-art models in both seizure detection and seizure classification tasks,” Hajisafi said.
On a seizure detection task, determining whether a 60-second EEG clip contains a seizure, the team's model achieved up to a 5.3% improvement over existing deep learning approaches. Hajisafi says this is a notable but relatively modest improvement, as other deep learning models also perform well when large numbers of training samples are available.
“The real strength of our approach becomes apparent in the seizure classification task, especially the rare seizure class, where the number of training samples is very limited,” he added. “In this context, our model delivers a significant overall improvement of 12.7%, which is quite significant for a deep learning task. Furthermore, when we carried out additional analyses in scenarios with even more severe data scarcity, we found that our approach performs surprisingly well in rare seizure classification, outperforming previous models by a whopping 29%.”
A highly adaptable model with a bright future
Hajisafi added that the team expected that adding situational factors would improve seizure detection, but they were surprised by how powerful the factors were together.
“Their combined impact was much greater than the sum of their individual parts,” the researchers continued. “Furthermore, while the performance of previous models dropped significantly with fewer training samples, our approach remained consistent even after training on a fraction of the available training data. This robustness demonstrates that our framework is particularly suitable for applications with limited data availability and is highly adaptable to a variety of medical and non-medical domains.”
Hajisafi said there are many potential avenues the technology could pursue in the future, but the team's next step is to integrate it into wearable technology, such as an EEG cap, that can continuously monitor brain activity outside of a clinical setting.
“Our system is suitable for continuous diagnostics and can be integrated into a wearable device to enable real-time seizure detection/classification,” Hajisafi said. “This would allow an alert to be sent to the patient's smartphone and to the healthcare provider as soon as a seizure is detected.”
This system could be adopted in wearable EEG technology as well as potentially applied to other wearable electronic devices that could enable epileptic seizure detection beyond EEG.
“We hope to explore the possibility of using existing wearable technology, such as smartwatches, to leverage measurements of physical activity and physiological responses, such as heart rate and movement patterns, as proxies for direct recordings of the brain's electrical activity,” Hajisafi concluded. “Using these data sources, we could develop similar methods for detecting and classifying seizures without relying on traditional electroencephalography.”
Reference: A. Hajisafi., et al., Dynamic GNNs for Accurate Seizure Detection and Classification from EEG Data, Advances in Knowledge Discovery and Data Mining [2024] 28th Pacific Asia Knowledge Discovery and Data Mining Conference, PAKDD 2024, Taipei, Taiwan, May 7-10, 2024, Proceedings, Part IV
Featured Image: Google DeepMind on Unsplash
