Research

We introduce a geometric framework for k-means that characterizes the structure of local optima in clustering. The approach reveals the geometric relationships between solutions and provides theoretical guarantees for cluster detection under diverse data distributions.

We introduce ChatBCI, a P300 speller that integrates large language models to provide context-driven word prediction during EEG-based spelling. The system significantly reduces keystrokes and writing time (-53% and -62%) while increasing information transfer rate by 229% compared to conventional letter-by-letter BCI spellers.

We introduce ChatBCI-Assist, an intent-based P300 speller that integrates a locally-deployed large language model with an adaptive stopping strategy for real-time EEG-based communication. The system achieves record-level online performance, reaching 30.7 characters/min (105.2 bits/min ITR) and 147.1 bits/min semantic information transfer rate, enabling faster and more natural sentence composition in practical BCI communication.

We propose SpellerSSL, a self-supervised learning framework that combines P300 aggregation with cross-domain EEG pretraining to learn generalizable representations for speller BCIs. The method achieves 94% character recognition with only 7 repetitions, reaches 21.86 bits/min information transfer rate, and reduces calibration data requirements by 60%.

We introduce EEGM2, a self-supervised framework for long-sequence EEG modeling that integrates a U-shaped encoder–decoder with the Mamba-2 state-space architecture to achieve linear computational complexity and efficient long-range dependency modeling. The approach improves representation learning for EEG signals, enabling scalable training, faster inference, and strong cross-subject generalization across diverse EEG datasets.

We introduce a novel data-driven dynamic stopping strategy for LLM-based P300 speller BCIs, which adaptively determines when to finalize character selection based on real-time EEG decoding. The method achieves over 260% improvement in information transfer rate compared to traditional fixed-repetition approaches.

We present ChatBCI4ALS, the first intent-based P300 BCI communication system tailored for individuals with ALS. Online experiments achieve up to 42.16 characters/min typing speed and 128.85 bits/min information transfer rate, significantly advancing real-time BCI communication for assistive applications.

We introduce TopoEEG, an image-based framework that generates sequences of topographic maps from EEG data, preserving both temporal and spatial information for decoding neural dynamics. Leveraging a TimeSformer backbone, the method enables accurate single-trial early detection of P300 responses.

We propose P3T (P300-Transformer), a novel single-trial P300 detector and transfer model designed to optimize the information transfer rate in P300-BCI speller systems while maintaining a high character recognition rate.

We propose a deep learning framework utilizing dynamic channel selection for early classification of left and right hand motor imagery tasks, reducing data dimensionality and accelerating real-time BCI decoding.