AI in Epilepsy: A Game Changer for Seizure Diagnosis, Monitoring and Seizure Detection.

With Artificial Intelligence (AI) rapidly evolving, its potential to improve the diagnosis and personalised care for epilepsy patients has become increasingly evident. By enhancing accuracy and efficiency, AI can play a crucial role in addressing the complexities involved in diagnosing epileptic seizures.

Machine Learning (ML) can be applied to analyse EEGs, neuroimaging (e.g., MRI scans), wearable data, and seizure videos. This article will provide an overview of AI and ML, their applications in healthcare for better outcomes in epilepsy management, and their limitations.

What is Artificial Intelligence (AI) & Machine Learning (ML)?

AI is a field of computer science associated with a machine's ability to use algorithms to perform tasks that are usually require the human mind.

ML is a sub-classification of AI, which allows computers to learn from large amounts of data without needing specific instructions. One of the key techniques used in ML is called deep neural networks (DNNs). DNNs have become popular because they can process and make sense of complex data, like images, which were difficult to analyse with traditional methods.

Deep Learning (DL) is an area within ML that is able to extract complex signals from observations.

EEG Interpretation

Electroencephalogram (EEG) is one of the most common tests used to detect abnormal electrical brain activity that may be causing seizures. However, scalp EEG (electrodes placed on the scalp) has its limitations, including low sensitivity and being time-consuming to interpret.

To improve EEG interpretation, deep learning models like SpikeNet and DeepSpike have been developed to detect spikes, or abnormal brain activity patterns, in EEGs. These models have shown better accuracy than older systems, though challenges remain in balancing sensitivity and specificity. Newer models, such as SCORE-AI, are performing at levels similar to human experts in differentiating between epileptic and non-epileptic seizures.

AI models have also been applied to categorising EEGs as either normal or abnormal, with success rates around 80-90% in identifying epileptiform discharges - which are distinct electrical patterns associated with epilepsy. Despite AI's promising outcomes so far, human experts are still essential for reviewing results, and efforts are ongoing to improve AI accuracy further.

Wearable Devices

Wearable devices are increasingly being used for long-term, non-invasive monitoring and automated seizure detection. These devices allow patients to be monitored outside clinical settings. Common examples include wrist devices equipped with accelerometers, heart rate monitors, and electrodermal sensors, as well as surface EMG patches and scalp EEG systems. The current focus is mainly on detecting tonic-clonic seizures and alerting caregivers for patient safety.

Many wearables use simple, threshold-based methods to detect seizures. However, ML techniques, such as Support Vector Machines (SVM), are being applied for more precise seizure detection. SVM models are particularly effective in wearables, achieving 88-95% sensitivity for tonic-clonic seizures, with low false alarm rates and quick detection times. These models can also distinguish between epileptic and non-epileptic events.

Automated detection of Seizures in Videos

Video recordings, often taken on smartphones or security cameras, are a valuable tool for diagnosing seizures and differentiating between epileptic and non-epileptic events. While doctors and neurology experts can interpret these videos, AI and ML techniques are being explored to automate the process - i.e. increase its efficiency.

Automated video analysis could be crucial for detecting and monitoring seizures in situations like sleep or in newborns (neonates), where movements are limited. SVM models have been successful in identifying myoclonic seizures during sleep, and AI tools are being used to detect major motor seizures, such as tonic-clonic seizures. Studies in newborns, who have more restricted movements, have shown promising results in detecting motor seizures via video analysis.

More advanced video techniques, such as 3D camera footage with noise reduction, have even been able to differentiate between different types of seizures, including frontal lobe and temporal lobe seizures, as well as non-epileptic events.

AI in interpreting neuroimaging studies

Artificial intelligence (AI) has gained considerable attention in medical imaging analysis, particularly in epilepsy. Its primary applications involve pinpointing the origin of seizures for potential surgical interventions and predicting disease progression. AI and machine learning (ML) techniques are frequently used in analysing structural MRI scans to locate epileptogenic foci (areas responsible for seizures) or other diagnostic indicators.

Multiple Convolutional Neural Networks (CNNs), a type of ML model mainly used to interpret images and audio signals, have been developed to detect brain abnormalities, such as Hippocampal Sclerosis (HS), also known as Mesial Temporal Sclerosis, and Focal Cortical Dysplasias (FCDs). HS is characterized by the loss of neurons and scarring in the hippocampus, the brain region crucial for memory and emotional processing. In contrast, FCDs involve abnormal organisation and development of brain cells in specific areas.

In one study, MRI scans successfully identified FCDs with a sensitivity of 90% and specificity of 85%, although the sample size was relatively small, including only 30 patients, of whom 10 had FCDs. Another CNN model focused on unilateral temporal lobe epilepsy - epilepsy originating only in the left side of the brain - achieved a sensitivity of 82%, specificity ranging from 87% to 91%, and an accuracy of 85% to 86%. This model performed well across multiple centres and showed good accuracy even for MRI-negative cases.

Limitations of AI in Epilepsy

Artificial intelligence and machine learning (AI/ML) in medicine are still developing, with many studies focusing on proving their feasibility rather than providing fully validated solutions. Most research uses data from a single centre, which limits its general applicability. Current AI models, like SCORE-AI, aim to assist experts rather than replace them, and automated detection tools for EEG spikes and seizures often do not achieve the accuracy of human clinicians. A major challenge is the lack of large, high-quality datasets - as the existing public datasets are small and may introduce bias- and so therefore, future work is needed to standardise AI techniques and reporting in medical research.

Takeaway

AI and ML hold significant promise for enhancing the diagnosis and management of epilepsy through improved interpretation of EEGs, neuroimaging, wearable data, and video analysis. While progress has been made, challenges remain in ensuring accuracy and general applicability. Collaboration between AI technologies and human expertise is crucial for advancing epilepsy care and achieving better outcomes for patients.

References

Epilepsy Ecosystem, 2024. My Seizure Gauge. [online] Available at: https://www.epilepsyecosystem.org/my-seizure-gauge-1 [Accessed 12 October 2024].

Epilepsy Foundation, 2024. Focal Cortical Dysplasia. [online] Available at: https://www.epilepsy.com/causes/structural/focal-cortical-dysplasia#:~:text=Focal%20Cortical%20Dysplasia%20(FCD)%20is,outermost%20part%20of%20the%20brain . [Accessed 12 October 2024].

Epilepsy Foundation, 2024. Mesial Temporal Sclerosis. [online] Available at: https://www.epilepsy.com/causes/structural/mesial-temporal-sclerosis [Accessed 12 October 2024].

Fink, M., et al., 2024. Utilising machine learning in the diagnosis of epilepsy: Challenges and opportunities. Seizure, 108, pp.30-38. Available at: https://www.sciencedirect.com/science/article/pii/S1525505024001173 [Accessed 12 October 2024].

McKinsey & Company, 2023. What is AI? [online] Available at: https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-ai [Accessed 12 October 2024].

Radiopaedia, 2024. Focal Cortical Dysplasia. [online] Available at: https://radiopaedia.org/articles/focal-cortical-dysplasia [Accessed 12 October 2024].

Radiopaedia.org , 2024. Mesial temporal sclerosis. [online] Available at: https://radiopaedia.org/articles/mesial-temporal-sclerosis [Accessed 12 Oct. 2024].

Taha, L., et al., 2022. A comprehensive deep learning approach for detection of seizures in video data. Scientific Reports, 12(1), pp.1-12. Available at: https://www.nature.com/articles/s41598-022-14894-4 [Accessed 12 October 2024].