Let users run find_bads_muscle also when no sensor positions are av…

…ailable (mne-tools#12862)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Clemens Brunner <[email protected]>
Co-authored-by: Eric Larson <[email protected]>

doc/changes/devel/12862.other.rst

@@ -0,0 +1 @@
+:meth:`mne.preprocessing.ICA.find_bads_muscle` can now be run when passing an ``inst`` without sensor positions. However, it will just use the first of three criteria (slope) to find muscle-related ICA components, by `Stefan Appelhoff`_.

mne/preprocessing/ica.py

@@ -62,6 +62,7 @@
     _PCA,
     Bunch,
     _check_all_same_channel_names,
+    _check_ch_locs,
     _check_compensation_grade,
     _check_fname,
     _check_on_missing,
@@ -1955,6 +1956,9 @@ def find_bads_muscle(
         has been modified to 45 Hz as a default based on the criteria being
         more accurate in practice.
 
+        If ``inst`` is supplied without sensor positions, only the first criterion
+        (slope) is applied.
+
         Parameters
         ----------
         inst : instance of Raw, Epochs or Evoked
@@ -1992,6 +1996,8 @@ def find_bads_muscle(
         """
         _validate_type(threshold, "numeric", "threshold")
 
+        slope_score, focus_score, smoothness_score = None, None, None
+
         sources = self.get_sources(inst, start=start, stop=stop)
         components = self.get_components()
 
@@ -2002,11 +2008,32 @@ def find_bads_muscle(
             psds = psds.mean(axis=0)
         slopes = np.polyfit(np.log10(freqs), np.log10(psds).T, 1)[0]
 
+        # typical muscle slope is ~0.15, non-muscle components negative
+        # so logistic with shift -0.5 and slope 0.25 so -0.5 -> 0.5 and 0->1
+        slope_score = expit((slopes + 0.5) / 0.25)
+
+        # Need sensor positions for the criteria below, so return with only one score
+        # if no positions available
+        picks = _picks_to_idx(
+            inst.info, self.ch_names, "all", exclude=(), allow_empty=False
+        )
+        if not _check_ch_locs(inst.info, picks=picks):
+            warn(
+                "No sensor positions found. Scores for bad muscle components are only "
+                "based on the 'slope' criterion."
+            )
+            scores = slope_score
+            self.labels_["muscle"] = [
+                idx for idx, score in enumerate(scores) if score > threshold
+            ]
+            return self.labels_["muscle"], scores
+
         # compute metric #2: distance from the vertex of focus
         components_norm = abs(components) / np.max(abs(components), axis=0)
         # we need to retrieve the position from the channels that were used to
         # fit the ICA. N.B: picks in _find_topomap_coords includes bad channels
         # even if they are not provided explicitly.
+
         pos = _find_topomap_coords(
             inst.info, picks=self.ch_names, sphere=sphere, ignore_overlap=True
         )
@@ -2016,6 +2043,10 @@ def find_bads_muscle(
         dists /= dists.max()
         focus_dists = np.dot(dists, components_norm)
 
+        # focus distance is ~65% of max electrode distance with 10% slope
+        # (assumes typical head size)
+        focus_score = expit((focus_dists - 0.65) / 0.1)
+
         # compute metric #3: smoothness
         smoothnesses = np.zeros((components.shape[1],))
         dists = distance.squareform(distance.pdist(pos))
@@ -2025,20 +2056,22 @@ def find_bads_muscle(
             comp_dists /= comp_dists.max()
             smoothnesses[idx] = np.multiply(dists, comp_dists).sum()
 
-        # typical muscle slope is ~0.15, non-muscle components negative
-        # so logistic with shift -0.5 and slope 0.25 so -0.5 -> 0.5 and 0->1
-        slope_score = expit((slopes + 0.5) / 0.25)
-        # focus distance is ~65% of max electrode distance with 10% slope
-        # (assumes typical head size)
-        focus_score = expit((focus_dists - 0.65) / 0.1)
         # smoothnessness is around 150 for muscle and 450 otherwise
         # so use reversed logistic centered at 300 with 100 slope
         smoothness_score = 1 - expit((smoothnesses - 300) / 100)
-        # multiply so that all three components must be present
-        scores = slope_score * focus_score * smoothness_score
+
+        # multiply all criteria that are present
+        scores = [
+            score
+            for score in [slope_score, focus_score, smoothness_score]
+            if score is not None
+        ]
+        n_criteria = len(scores)
+        scores = np.prod(np.array(scores), axis=0)
+
         # scale the threshold by the use of three metrics
         self.labels_["muscle"] = [
-            idx for idx, score in enumerate(scores) if score > threshold**3
+            idx for idx, score in enumerate(scores) if score > threshold**n_criteria
         ]
         return self.labels_["muscle"], scores
 

mne/preprocessing/tests/test_ica.py

@@ -1520,6 +1520,13 @@ def test_ica_labels():
     ica.find_bads_muscle(raw)
     assert "muscle" in ica.labels_
 
+    # Try without sensor locations
+    raw.set_montage(None)
+    with pytest.warns(RuntimeWarning, match="based on the 'slope' criterion"):
+        labels, scores = ica.find_bads_muscle(raw, threshold=0.35)
+    assert "muscle" in ica.labels_
+    assert labels == [3]
+
 
 @testing.requires_testing_data
 @pytest.mark.parametrize(