The College of Engineering and Computing (CEC) 3-Minute Thesis (3MT) competition challenges graduate students to present their research concisely and compellingly to a non-specialist audience within three minutes. Paras Parani participated in this prestigious competition, showcasing his innovative research on utilizing Large Language Models (LLMs) and Vision Transformers (ViTs) for epileptic seizure prediction. His latest findings demonstrate the potential of transformer-based architectures to achieve significant accuracy while maintaining computational efficiency for real-time applications. Out of 13 participants, Paras was selected among the top two and has been nominated by the college to represent CEC at the university-wide competition in January.

The American Epilepsy Society (AES) Fellowship is a reward for early career medical professionals and researchers in the area of epilepsy. Approximately 105 fellows are selected for this prestigious award each year out of thousands of applicants. The award comes with a 1-year complimentary membership to the AES, registration fee waiver, and a $575 USD award to travel to 2024 AES Annual meeting (in Los Angeles, CA this year). Attending this event and a 1-year membership will provide Dr. Umair with further resources and networking opportunities to take the early lab’s early seizure prediction research forward.

Preliminary results from the work of Paras Parani, Dr. Umair Mohammad, and Dr. Fahad Saeed on lightweight seizure prediction models were accepted at the 2024 IEEE International Conference on Big Data workshop HPC-BOD. This research introduces ESPFormer, a lightweight transformer-based model, designed to predict epileptic seizures using patient-independent EEG data. By tokenizing and adapting multi-channel EEG time-series data, the model achieves a balance between accuracy and computational efficiency, enabling near real-time prediction capabilities. The proposed architecture outperforms resource-intensive models like Vision Transformers (ViTs) and Large Language Models (LLMs) on several benchmarks. Successful implementation of ESPFormer marks a significant step forward in developing scalable and effective seizure prediction technologies.

Preliminary results from the work of Paras Parani, Dr. Umair Mohammad and Dr. Fahad Saeed on early seizure prediction using pre-trained large language models (LLMs) and vision transformers (ViTs) were accepted at the 8th International Conference on Data Science and Machine Learning (CDMA 2024). The work proposes techniques to modify multi-channel time-series electroencephalography (EEG) data and pre-trained ViTs and LLMs to predict seizures with the partial re-design and re-training of only the first and last layers. Successful execution of this strategy will be a game-changer for seizure prediction technologies. It will allow for the accurate and precise prediction of seizures using models with billions of parameters that have been trained on huge volumes of data by applying it to retroactive EEG datasets that are comparable much smaller in size.

Saeed Lab webpage (https://pcdslab.github.io/) is now up and running. This will ensure that all members of the team can contribute to the webpage for better science communication to the scientific community as well as general audience. We are very excited to be here!