Simulating the M/EEG data

Obtaining the data

Up to this point, we defined a bunch of sources and set up several coupling edges. However, no data was generated yet, and it’s time to fix that now.

First, you need to run the SourceSimulator.simulate() method of the sim object to actually simulate the waveforms of all previously defined sources:

sfreq = 250     # in Hz
duration = 30   # in seconds
sc = sim.simulate(sfreq, duration)

The result of this function call is a SourceConfiguration object that contains all simulated sources and their waveforms (with desired SNR and coupling).

Now you can use the SourceConfiguration.to_stc() and SourceConfiguration.to_raw() methods to obtain source time courses and sensor-space data, respectively. The projection to sensor space requires a forward model (mne.Forward) and an mne.Info object describing the sensor layout:

stc = sc.to_stc()

raw = sc.to_raw(fwd, info)

Global adjustment of the SNR

If necessary, the global SNR can be adjusted when simulating the data. To achieve this, you need to specify the desired value of SNR (snr_global) and the limits of the frequency band that should be considered during the adjustment (fmin and fmax keys in the snr_params dictionary). As a result, the mean sensor-space variance of all point and patch sources will be adjusted relative to the mean variance of all noise sources to achieve the desired SNR:

sfreq = 250     # in Hz
duration = 30   # in seconds
sc = sim.simulate(
    sfreq,
    duration,
    fwd=fwd,    # Forward model is required for the adjustment of SNR
    snr_global=3.0,
    snr_params=dict(fmin=8, fmax=12)
)

Adding sensor noise

When projecting the simulated source activity to sensor space, it is also possible to add a desired amount of sensor noise (currently modeled as white noise):

raw = sc.to_raw(fwd, info, sensor_noise_level=0.01)

The sensor_noise_level controls the ratio of noise to total power in sensor space. In the example above, 1% of total sensor-space power will originate from noise.

Reproducibility

By design, the result of SourceSimulator.simulate() will differ every time you call the method, making it very easy to simulate multiple datasets under the same settings.

However, it is always possible to obtain a reproducible result if you provide a specific value of the random_state as shown below:

sc = sim.simulate(sfreq, duration, random_state=123)

The resulting source configuration, including the locations and waveforms of all sources, will be the same on every call.

Accessing the simulated sources

The simulated sources can be quickly accessed by their name ("source" in the example below). This might be helpful in case ground-truth waveforms were generated randomly:

s = sc["source"]
gt = s.waveform