Foundation Models for Epileptogenic Zone Identification in Drug-Resistant Epilepsy
*Thomas Lai, *Thi Kieu Khanh Ho, Petr Klimes, Jan Cimbalnik, Martin Pail, Milan Brazdil, Birgit Frauscher, and Narges Armanfard (*Equal Contribution)
arXiv Preprint, June 2026
Description: We developed EpiiSLM, a dual foundation model system for identifying the epileptogenic zone (EZ) in drug-resistant epilepsy using stereo-electroencephalography (sEEG). A signal foundation model is trained on over 104,000 minutes of sEEG recordings from the Montreal Neurological Institute & Hospital, anchoring EZ biomarker extraction on non-epileptic signals. A language foundation model then integrates sEEG-derived outputs with multimodal clinical information to produce interpretable predictions. Under leave-one-patient-out evaluation, EpiiSLM achieved 0.978 contact-level positive predictive value (PPV) and 100% region-level accuracy. Notably, EpiiSLM requires only one night of interictal sleep data, suggesting the potential to reduce invasive sEEG monitoring duration and improve surgical outcomes.
Thomas Lai, Thi Kieu Khanh Ho, and Narges Armanfard
ECAI 2024 (Oral)
Description: We propose MOSAD, a multi-head framework for open-set time-series anomaly detection (TSAD). While most TSAD methods are unsupervised (trained on normal data only) or supervised (prone to overfitting on seen anomaly types), MOSAD addresses the open-set scenario where only a few labeled anomalies from a limited set of classes are available during training. The framework uses a shared representation space with three specialized heads — generative, discriminative, and anomaly-aware contrastive — to achieve state-of-the-art performance on three real-world datasets, demonstrating superior adaptability and robustness to unseen anomaly types.
Age, Breed, and Individual Dog Classification Based on Canine Bark Samples
Thomas Lai aka WhatDoesTheDogSay
IJCAI 2025 Competition (Winner)
Description: Won the Barkopedia competition at IJCAI 2025, a large-scale challenge for canine vocalization analysis. The competition featured eight tasks including dog age group classification, breed classification, individual dog recognition, emotion classification, and vocal separation/detection. Using deep learning techniques adapted from speech and audio processing, I competed across the demographic and identity classification tasks (age, breed, and individual recognition) on a diverse benchmark dataset of real-world dog barks.
Deep Reinforcement Learning to Control Feeding in Cell Therapy Manufacturing
Thomas Lai, Katherine MacDonald, Ali Chamkalani, Christopher Sherwood, Nathan Lawrence, Nikesh Patel, James Piret, and Bhushan Gopaluni
Undergraduate research project report (NSERC USRA) @ UBC Michael Smith Laboratories
Description: Developed a model-based framework to optimize feeding strategies in allogeneic regulatory T-cell (Treg) therapy manufacturing. Current protocols feed cell cultures more nutrients at a higher frequency than necessary, increasing contamination risk and media cost. The approach uses logistic models to simulate stochastic culture dynamics, combined with deep Q-learning (DQN) agents for optimization and online control. The best RL agent scored 85% more reward than standard clinical trial protocols, demonstrating the efficacy of RL controllers in human cell culture systems and addressing challenges of scarce data and high donor-to-donor variability in allogeneic therapies.
The Effect of Cerebral Blood on MRI T1, T2, and T2* in the Grey and White Matter
*Thomas Lai, *Stefano Bambace, *Efe Grant-Oyeye, *Jeanine Looman, Sethu K. Boopathy Jegathambal, Amir Schmuel (*Equal Contribution)
Undergraduate capstone project @ McGill Neuro
Description: Investigated the contribution of cerebral blood to MRI relaxation parameters (T1, T2, and T2*) in grey and white matter brain tissue. This capstone project combined whole-brain MRI data with high-resolution imaging techniques to study how the presence of blood influences quantitative MRI measurements, which is relevant for understanding brain organization and improving the accuracy of tissue characterization in neuroimaging.
