Multi-centre validation
Evaluation designed around real variation between clinical sites.
01 · Clinical translation
Clinical neural AI
Our clinical work focuses on robust, interpretable models for routine EEG. Rather than optimizing on a single benchmark, we test how models behave across hospitals, patient groups, recording protocols, and equipment. Current work includes interictal biomarkers for epilepsy and rigorous tests of whether published systems are ready for real clinical data.
Interpretable biomarkers
Features and models that support clinical reasoning, not only prediction.
Clinical EEGEpilepsyExternal validationBiomarkersTrustworthy AI
02 · Signal quality
Efficient neural signal restoration
Eye movements, muscle activity, and environmental interference routinely obscure EEG. We build compact denoising systems that exploit the structure of these artifacts while protecting diagnostically important neural activity. The goal is practical signal restoration for both laboratory pipelines and resource-constrained devices.
Femtomodels
Exploring the lower limits of model size for generalisable EOG denoising.
Wavelet attention
Multi-scale, frequency-aware networks for diverse EEG artifacts.
EEG denoisingEOG artifactsWaveletsCompact networksEdge AI
03 · Neural principles
Bio-inspired computation
Biological systems achieve useful computation with striking efficiency. We study recurrent excitatory–inhibitory systems, attractor-like dynamics, and connectome-constrained networks as compact testbeds for sequential inference. The emphasis is on controlled experiments that separate the effect of topology, dynamics, and learning rules.
Balanced recurrent systems
Excitatory–inhibitory architectures for anytime and sequential prediction.
Connectome models
Testing what biological wiring contributes under carefully matched controls.
E/I balanceAttractor networksConnectomicsNeuromorphic AISequential inference
04 · Theory
Neural dynamics & information
EEG is a multivariate observation of a nonlinear dynamical system. We use tools from time-series analysis, covariance geometry, information theory, and network science to characterise state changes and ask which aspects of a signal representation remain identifiable and useful.
Nonlinear dynamics
Interpretable descriptions of brain-state trajectories and transitions.
Information fidelity
Principled limits on what signal representations preserve or discard.
Multivariate time seriesCovariance geometryInformation theoryNetwork dynamicsState transitions
05 · Neurotechnology
Brain–computer interfaces
We investigate signal acquisition, conditioning, feature extraction, and adaptive decoding for real-time neural interfaces. This work connects our foundational research on robust signal processing with systems for communication, motor rehabilitation, and assistive neurotechnology.
Real-time EEGAdaptive decodingAssistive systemsEMGProsthetics