Sieci neuronowe oparte na fizyce do niezawodnego odkrywania biomarkerów w przenośnych systemach EEG niskiej gęstości
Physics-Informed Neural Networks for Robust Digital Biomarker Discovery in Low-Density Wearable EEG Systems
W skrócie
[Preprint - wstępne wyniki] Badacze opracowali nową metodę analizy zapisów mózgu (EEG) z przenośnych urządzeń, które są znacznie prostsze niż sprzęt szpitalny, ale trudniejsze w interpretacji. System wykorzystuje sztuczną inteligencję wzmocnioną zasadami fizyki, aby dokładnie wykrywać napady padaczki nawet gdy sygnał jest mocno zniekształcony szumem i ruchami pacjenta. W testach metoda ta wykazała 88% dokładność w najtrudniejszych warunkach (porównując do 74% dla dotychczasowych podejść), co daje nadzieję na szybszą diagnozę padaczki w warunkach domowych i ambulatoryjnych.
Oryginalny abstract (angielski)
Abstract Background: Wearable electroencephalography (EEG) holds transformative potential for community-based neurological diagnosis, yet the translation from hospital-grade 23-channel recordings to low-density, motion-corrupted wearable signals introduces severe domain shift that invalidates conventional machine-learning biomarkers. Thousands of patients with suspected seizures face diagnostic waits exceeding twelve months, with up to 30% subsequently reclassified between epilepsy, functional neurological disorder (FND), and other non-epileptic conditions. Methods: We propose a Physics-Informed Neural Network (PINN) framework for robust EEG biomarker learning that couples a graph neural network (GNN) spatial encoder with a bidirectional LSTM temporal encoder and four physiologically grounded loss components: a spectral constraint ( L spec ), a spatial diffusion constraint ( L spat ), a functional connectivity stability constraint ( L conn ), and a novel biomarker stability constraint ( L bio ). The framework is validated on the CHB-MIT Scalp EEG Database (subject chb21; four annotated seizures), with a synthetic degradation pipeline simulating electrode dropout (23 → 4 channels), Gaussian noise (σ ∈ [0.1, 2.0]×SD), motion artefacts, and downsampling (256→64 Hz). Strict leave-one-patient-out cross-validation was employed throughout. Results: Under hospital-grade conditions the PINN-GNN achieved AUROC = 0.972 and F1 = 0.954 for seizure detection. Under extreme wearable-grade degradation (σ = 2.0) the PINN-GNN retained AUROC = 0.881, compared with 0.741 for a baseline GNN,a 16.2% relative robustness improvement. The biomarker stability index (BSI) was significantly lower for the PINN-GNN across all six biomarker classes tested (band power ratio, spectral entropy, phase-locking value, graph centrality, Hjorth mobility, and theta coherence; p Conclusions: Physics-informed constraints substantially improve EEG biomarker stability under real-world degradation conditions, advancing the feasibility of community-based seizure monitoring. The proposed framework directly addresses the EPSRC-funded research mandate to develop and validate robust biomarkers for wearable EEG deployment in epilepsy, FND, and related conditions, and provides a validated methodological foundation for differentiation of epilepsy, FND, and syncope from low-density EEG data.