Poster Presentation

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

System identification from raw EEG during clinical propofol sedation with an interpretable latent SDE model

K. Seeliger¹, Jan W. Kantelhardt¹, Jakob Garbe², Nico Scherf³, Thomas Schmid¹; ¹Martin-Luther-Universität Halle-Wittenberg, ²University Hospital Halle, ³Max Planck Institute for Human Cognitive and Brain Sciences

Presenter: K. Seeliger

Current EEG-based sedation monitoring commonly uses handcrafted spectral or entropy features and uses them for retrospective sedation level prediction. This gives only limited access to the temporal dynamics of consciousness and is not designed to address the clinically relevant problem of short-term future sedation change prediction in continuous propofol administration settings. Here, a dynamical autoencoder, encoding into a latent stochastic differential equation (SDE)-based Kuramoto-like dynamical system, is trained end-to-end on large-scale raw two-channel clinical EEG recorded under propofol sedation. The data consist of 285 recordings during routine endoscopy, with expert nurse annotations of depth of consciousness on the MOAAS scale. The model maps each input window to a low-dimensional latent feature state using a temporal convolutional neural network (TCN), evolves this state forward with Kuramoto-like coupled oscillators with learned state-dependent parameters, and decodes the predicted latent trajectory back to EEG. This creates a compact dynamical system with learned interpretable internal parameters relevant to consciousness depth, such as oscillator frequencies, pairwise coupling, and phase synchrony. Although MOAAS labels are not used to shape the latent dynamics directly, the learned latent space shows graded structure with respect to sedation depth. In held-out subjects, these latent features support prediction of conscious and unconscious states. The results therefore show that sedation structure can emerge in a forecasting dynamical autoencoder trained directly on raw EEG reconstruction. More broadly, this supports end-to-end learning of interpretable dynamical systems from raw neural time series.

Topic Area: Methods, Tools, Theory & Neural Coding