We are pleased to announce that the paper “EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs” was accepted for presentation at the prestigious KDD 2024 Conference.
Navid Foumani is the lead author. The co-authors are Dr. Mahsa Salehi (Monash University), Dr. Geoffrey Mackellar, Dr. Soheila Ghane, Dr. Saad Irtza, and Dr. Nam Nguyen (EMOTIV Research, Pty Ltd).
EMOTIV sponsors Navid Foumani, a PhD candidate who has been working on applying deep learning methods to EEG data under the supervision of Dr. Mahsa Salehi at Monash University in Melbourne, Australia. Navid worked closely with our team to develop a novel self-supervised architecture known as EEG2Rep, which is immensely promising for modeling EEG data.
As one of 5 EEG datasets, Navid applied these methods to our Driver Attention data:18 subjects x 45 minutes of simulated driving with intermittent distractors typical of a driving experience (mobile calls, text messages, navigation, music selection, conversation, mental calculations on the fly etc.). Our Driver Attention algorithm was delivered with a 68% accuracy metric using state-of-the-art machine learning methods in 2013.
We sponsored Mahsa during her PhD at Melbourne University in 2015, providing her with the same dataset. She managed to improve the accuracy metric to 72% using ensemble methods.
The EEG2Rep model was applied to the Driver Distraction dataset and achieved the highest accuracy to date, 80.07%, a substantial improvement. Additionally, the model significantly outperformed state-of-the-art methods in each of the five public datasets, including emotional and mental state detection, multitasking, resting state EEG, and detection of medical conditions such as epilepsy and stroke.
This success opens up the possibility of developing a foundational model for EEG data that can generalize across various tasks and applications, pushing the boundaries of what can be achieved in the field of EEG analysis.
