The electroencephalogram is an imperfect tool for diagnosing epilepsy. Enter artificial intelligence, which enhances the test’s sensitivity. For people living with this condition, the technique means a faster diagnosis and introduction of the right treatment earlier in their care journey. This is what Dr. Élie Bou Assi and his research team at the Centre hospitalier de l’Université de Montréal (CHUM) in Montréal are currently working on.

The WHO estimates that 50 million people worldwide have epilepsy, a neurological disease that affects their quality of life. One of the fundamental tools in epilepsy diagnosis today is visual interpretation of an electroencephalogram (EEG) by a neurologist, who looks for interictal epileptiform discharges (IEDs), i.e., electrical signals that can be observed between epileptic seizures.

Unfortunately, more than half of all routine 30-minute EEGs fail to detect these sporadic signals. In the absence of these biomarkers, clinicians may either be reassured by a “false negative” or face a diagnostic dilemma, delaying appropriate decisions on treatment. To overcome this problem, my research project aims to explore new biomarkers of brain connectivity hidden in EEGs that offer greater sensitivity.

Theories of epilepsy pathophysiology suggest that measures of brain connectivity are probably stable biomarkers enabling differentiation between the brain of a healthy subject and that of someone with epilepsy. An epileptic seizure is the result of abnormal electrical activity triggered spontaneously and synchronously between different areas of the brain. The more connections there are between neurons in the brain, the easier it is for nerve impulses to propagate rapidly throughout the entire neural network, which favours the onset of epileptic seizures. Dr. Bou Assi’s team, of which I am a member, is working from a hypothesis which states that the functional network of an epileptic’s brain is altered compared to a healthy brain. Therefore, cerebral connectivity is a characteristic property of people affected by epilepsy. That connectivity can be calculated from EEGs.

An EEG is a recording of electrical activity in the brain using a series of electrodes affixed to the skull. Each electrode records the variation in intensity of electrical discharges in a particular region of the brain over time. Functional connectivity is defined by the extent to which electrical activity between two cortical regions is synchronous. To construct a functional connectivity network, each electrode will be represented by a node, and the extent of synchronization between each pair of electrodes, by a vertex. This will produce a functional connectivity graph from which we can extract connectivity biomarkers, which are useful for epilepsy diagnosis.

Multiple connectivity biomarkers can be calculated based on graph theory (a branch of discrete mathematics that studies the properties of graphs). These biomarkers can be grouped into categories, illustrating various network properties such as density (the number of connections per node), segregation (whether neighbours of neighbours are also neighbours, i.e., by forming a triangle) and integration (how easy it is to move from one node to any other node).

The more vertices there are, the more the nodes are connected and the greater the density. Segregation measures in a network, such as average clustering coefficient, illustrate the extent to which that network is composed of small groups with greatly interconnected nodes. The higher the clustering coefficient of a node, the more that node’s neighbours are also interconnected, forming closed circuits. Integration measures in a network, such as global efficiency, give us information about the capacity to travel from one node to any other node on the graph. In the context of the cerebral network, this corresponds to ease of communication between one point in the brain and another.

Previous studies have shown that biomarkers of connectivity derived from graph theory can be used to distinguish EEGs of people with epilepsy from those from non-epileptics. Most of those studies, however, involved sample sizes of fewer than 50 patients, which is not enough to reach a statistically significant conclusion that can be generalized to the wider population. Furthermore, they did not consider clinical covariates (e.g., age, medications, comorbidity) as confounding factors, which, according to a 2019 study by Kenézy Gyula University Hospital neurologist Béla Clemens, “Inter-ictal network of focal epilepsy and effects of clinical factors on network activity,” directly affects connectivity measures.

Dr. Bou Assi’s team at the CHUM has access to a database of more than 900 EEGs accompanied by patient files (including files of people with epilepsy and, as a control, those of non-epileptics who underwent EEG for other reasons). From each EEG, we extract connectivity biomarkers according to frequency bands and 10-second segments. This generates millions of data points: more than 900 EEGs multiplied by 100 segments multiplied by 5 frequency bands multiplied by around 20 calculated biomarkers. We then group all the biomarker values of the same category together and use them to train an artificial intelligence classifier to differentiate between people with and without epilepsy. The AI allows us to address neural network complexity and high individual variability.

We are continually enriching the database and optimizing every stage of EEG signal analysis to improve classifier performance and produce an increasingly reliable diagnostic tool.