DiffLSTM-MTE: A Hybrid LSTM-Diffusion Framework for Virtual iEEG Reconstruction From MEG

Le résumé fourni par la source

Accurate non-invasive characterization of complex brain dynamics, especially aberrant network activity in epilepsy, is crucial yet challenging due to the invasiveness of iEEG and limitations of noninvasive MEG. We introduce DiffLSTM-MTE, a novel deep learning framework that reconstructs virtual iEEG (ViEEG) from MEG recordings. This framework synergistically integrates an LSTM for temporal modeling of MEG with a conditional Denoising Diffusion Probabilistic Model (DDPM) for synthesizing realistic iEEG. Our physiologically-informed training objective incorporates losses for noise prediction, spectral fidelity, and Interictal Epileptiform Discharges (IEDs) preservation. Evaluated on data from five epilepsy patients, DiffLSTM-MTE demonstrates high temporal fidelity (mean PearsonR= 0.71 ± 0.08), preserves spectral characteristics (The mean PSD RMSE is < 0.071), reconstructs spatial coherence (mean correlation 0.92 ± 0.04), and accurately characterizes IEDs (76% ± 7% alignment). These results highlight DiffLSTM-MTE’s potential as a non-invasive tool to enhance epilepsy diagnostics and pre-surgical assessment by providing richer, patient-specific neurophysiological insights.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.