Use cleaned epochs for sensor-space decoding, time-frequency analysis, and covariance calculation

docs/source/v1.5.md.inc

@@ -1,5 +1,8 @@
 ## v1.5.0 (unreleased)
 
+This release contains a number of very important bug fixes that address problems related to decoding, time-frequency analysis, and inverse modeling.
+All users are encouraged to update.
+
 ### :new: New features & enhancements
 
 - Added `deriv_root` argument to CLI (#773 by @vferat)
@@ -32,3 +35,4 @@
 - Fixed bug with parallelization across runs for Maxwell filtering (#761 by @larsoner)
 - Fixed bug where head position files were not written with a proper suffix and extension (#761 by @larsoner)
 - Fixed bug where default values for `decoding_csp_times` and `decoding_csp_freqs` were not set dynamically based on the config parameters (#779 by @larsoner)
+- A number of processing steps erroneously **always** operated on un-cleaned epochs (`sensor/decoding_full_epochs`, `sensor/decoding_time_by_time`, `sensor/decoding_csp`); or operated on un-cleaned epochs (without PTP rejection) if no ICA or SSP was requested (`sensor/ime_frequency`, `sensor/make_cov`) The bug in `sensor/make_cov` could propagate to the source level, as the covariance matrix is used for inverse modeling. (#796 by @hoechenberger)

mne_bids_pipeline/steps/sensor/_02_decoding_full_epochs.py

@@ -60,6 +60,7 @@ def get_input_fnames_epochs_decoding(
         run=None,
         recording=cfg.rec,
         space=cfg.space,
+        processing="clean",
         suffix="epo",
         extension=".fif",
         datatype=cfg.datatype,
@@ -94,7 +95,6 @@ def run_epochs_decoding(
 
     epochs = mne.read_epochs(in_files.pop("epochs"))
     _restrict_analyze_channels(epochs, cfg)
-    epochs.crop(cfg.decoding_epochs_tmin, cfg.decoding_epochs_tmax)
 
     # We define the epochs and the labels
     if isinstance(cfg.conditions, dict):
@@ -111,6 +111,10 @@ def run_epochs_decoding(
         [epochs[epochs_conds[0]], epochs[epochs_conds[1]]], verbose="error"
     )
 
+    # Crop to the desired analysis interval. Do it only after the concatenation to work
+    # around https://github.com/mne-tools/mne-python/issues/12153
+    epochs.crop(cfg.decoding_epochs_tmin, cfg.decoding_epochs_tmax)
+
     n_cond1 = len(epochs[epochs_conds[0]])
     n_cond2 = len(epochs[epochs_conds[1]])
 

mne_bids_pipeline/steps/sensor/_03_decoding_time_by_time.py

@@ -65,6 +65,7 @@ def get_input_fnames_time_decoding(
         run=None,
         recording=cfg.rec,
         space=cfg.space,
+        processing="clean",
         suffix="epo",
         extension=".fif",
         datatype=cfg.datatype,

mne_bids_pipeline/steps/sensor/_04_time_frequency.py

@@ -33,9 +33,6 @@ def get_input_fnames_time_frequency(
     subject: str,
     session: Optional[str],
 ) -> dict:
-    processing = None
-    if cfg.spatial_filter is not None:
-        processing = "clean"
     fname_epochs = BIDSPath(
         subject=subject,
         session=session,
@@ -46,7 +43,7 @@ def get_input_fnames_time_frequency(
         space=cfg.space,
         datatype=cfg.datatype,
         root=cfg.deriv_root,
-        processing=processing,
+        processing="clean",
         suffix="epo",
         extension=".fif",
         check=False,

mne_bids_pipeline/steps/sensor/_05_decoding_csp.py

@@ -121,6 +121,7 @@ def get_input_fnames_csp(
         run=None,
         recording=cfg.rec,
         space=cfg.space,
+        processing="clean",
         suffix="epo",
         extension=".fif",
         datatype=cfg.datatype,

mne_bids_pipeline/steps/sensor/_06_make_cov.py

@@ -37,7 +37,6 @@ def get_input_fnames_cov(
 ) -> dict:
     cov_type = _get_cov_type(cfg)
     in_files = dict()
-    processing = "clean" if cfg.spatial_filter is not None else None
     fname_epochs = BIDSPath(
         subject=subject,
         session=session,
@@ -48,7 +47,7 @@ def get_input_fnames_cov(
         space=cfg.space,
         extension=".fif",
         suffix="epo",
-        processing=processing,
+        processing="clean",
         datatype=cfg.datatype,
         root=cfg.deriv_root,
         check=False,

mne_bids_pipeline/tests/configs/config_ds001810.py

@@ -15,6 +15,7 @@
 conditions = ["61450", "61511"]
 contrasts = [("61450", "61511")]
 decode = True
+decoding_n_splits = 3  # only for testing, use 5 otherwise
 
 l_freq = 0.3
 

mne_bids_pipeline/tests/conftest.py

@@ -43,6 +43,8 @@ def pytest_configure(config):
     ignore:is_categorical_dtype is deprecated.*:FutureWarning
     ignore:use_inf_as_na option is deprecated.*:FutureWarning
     ignore:All-NaN axis encountered.*:RuntimeWarning
+    # sklearn class not enough samples for cv=5
+    always:The least populated class in y has only.*:UserWarning
     """
     for warning_line in warning_lines.split("\n"):
         warning_line = warning_line.strip()