Fig. 2: Real-time Prediction and Latent State Dynamics During Electrocorticogram (ECoG) Data Assimilation.

a Prediction and latent state sequences in data assimilation. This figure illustrates the sequence of predictions and the mean (μ) of posterior distributions of latent states z⁽¹⁾, z⁽²⁾, z⁽³⁾, generated in real-time in response to observed ECoG data. The Observed ECoG combines 2500 time steps (2.5 s) of awake ECoG observations followed by 1500 time steps (1.5 s) of anesthetized ECoG observations from a test individual (an individual not included in the training dataset). This figure plots values after data assimilation is complete across all time steps of the observed ECoG sequence. The vertical solid line shows the moment of transition between awake and anesthetized conditions in observed ECoG data, with dashed lines showing the time window range when reaching this moment. In this range, updates of latent states are made in response to the transition in the observed ECoG. Predicted and observed ECoG signals are visualized by selecting 5 channels from the original 20 channels for visual clarity. Specifically, these five channels are chosen from the frontal pole (FP), primary motor cortex (M1), primary somatosensory cortex (S1), intraparietal sulcus (IPS), and primary visual cortex (V1) regions. Similarly, the z⁽¹⁾µ sequences were also visualized by selecting these five regions. b Example of the distribution of the mean (µ) of z⁽³⁾ obtained during training and data assimilation phases. In this study, the latent unit z⁽³⁾ is prepared in two dimensions, and the x- and y-axes represent the z⁽³⁾µ values in the first and second dimensions, respectively. The gray arrow indicates the direction in which z⁽³⁾ changed over time during data assimilation.