Fig. 5.

Results of the stochastic HMM inference on task MEG data from 52 MEG subjects, modelling the activity of the two motor cortices during a button-press task (1.2s inter-trial interval). To describe each state, we used a multivariate autoregressive model (MAR) distribution that can capture the spectral properties of the data. (a) The HMM, blind to the task information during the inference of the model, can discover states that reflect the underlying task dynamics; correspondingly, the state evoked response (top left panel) differentiates between event-related components that increase in probability (e.g. state blue, representing the evoked response) and others that show a decrease (e.g. state red, representing the event-related desynchronisation). When combining each state's temporal (top left panel) and spectral information (top right panel), the approach provides a time-frequency description of the data (bottom panel). Notably, the HMM captures the rhythmicity of the task (i.e. hand movements every 1.2s). (b) State onset probability, indicating, for each state and time point, the proportion of trials for which the state becomes active at this time point; each thin line represents a subject and the thick line is the group average; statistical significance of whether the dominant state has significantly larger onset probability than any of the other states is shown on top. (c) State evoked response for data where no task is performed. (d) Correlation of the state time courses across different runs of the algorithm, proving that the runs are extremely consistent. (e) Correlation of the power spectral densities between estimations obtained from separate half-splits of the data set, averaged across 5 random splits, and with the states ordered from less to more correlated. (f) Fractional occupancy and dwell times of the four states. (g) Transition probability matrix.