An AI model predicts epileptic seizure fast and accurately

Currently, there are roughly 50 million people worldwide with epileptic disorders. On the other hand, these days, you can hear about artificial intelligence (AI) everywhere. It is making its way through many realms, providing previously unbelievable solutions to various problems. This is to say that AI is now going to find its way to assist people suffering from epileptic seizures as well.

Sometimes, the passage of electrical signals in the brains of people with epilepsy goes turbulent. That is why a seizure happens with no prior alarm. The researchers at the University of Louisiana at Lafayette have developed a new AI-based model that can anticipate epileptic seizures. What’s even more surprising is that their model can predict the seizure up to one hour prior to onset, and with 99.6 percent accuracy.

Since an epileptic seizure is uncontrollable, it can drastically affect the patient’s psychological and social life. The prediction capability not only does increase a patient’s chance to hinder the seizure, but it will also help them cope with the psychological and social issues they sadly encounter in their normal life.

Previous works

These researchers are not the first ones trying to explore methods to predict seizures. There have been previous works, mostly focusing on analyzing brain activity using electroencephalogram (EEG) tests. Those models were suffering from complexity. That is because of two reasons:

· Firstly, each person has a unique brain pattern.

· Secondly, they had to extract one’s brain pattern manually and then apply a classification pattern to do the job.

The new model combines the pattern extraction and classification steps into one single automated system, hence enabling it to predict the seizure faster and more accurately.

Furthermore, the researchers also used another classification approach to further increase the accuracy of the process. As a matter of fact, they used a deep learning algorithm coupled with different electrode locations that extract and analyze the spatial-temporal features of the patient’s brain activity.

Ultimately, the researchers used the EEG readings and applied an additional algorithm to identify the most appropriate predictive channels of electrical activity inside those readings. Therefore, they speeded-up the process this way.

The model in action

The researchers then tested their model using long-term EEG data from 22 patients at the Boston Children’s Hospital. The results sounded appealing, though not in a big sample size. Not only is their model 99.6 percent accurate, but it also has a very low tendency of 0.004 false positive alarms per hour.

The model needs to be trained

The new model needs to be first set up to be able to provide accurate results. This is to say that the researchers need to train the model to help it achieve accurate and fast results for each specific patient.

The training process includes a few hours of non-invasive EEG monitoring around the seizure time. The process needs to be performed during the seizure time itself too. Furthermore, the process can be applied outside of the clinic by means of affordable EEG wearable electrodes.

How to use the model in real life

The researchers of this project say the next step is to develop a customized computer chip to process the algorithm. They are already trying to design efficient hardware that can empower the algorithm. To do so, they are taking into account the system size, power consumption, and latency to make the hardware practical.