Merge pull request #274 from MIT-LCP/edf_convert

Improves EDF speed / fixes bugs

Author: briangow

docs/conf.py

@@ -30,7 +30,7 @@ def __getattr__(cls, name):
             return MagicMock()
 
 MOCK_MODULES = ['numpy', 'matplotlib', 'matplotlib.pyplot', 'pandas', 'scipy',
-                'sklearn', 'sklearn.preprocessing', 'mne']
+                'sklearn', 'sklearn.preprocessing']
 sys.modules.update((mod_name, Mock()) for mod_name in MOCK_MODULES)
 
 # -- General configuration ------------------------------------------------

requirements.txt

@@ -5,7 +5,6 @@ idna==2.9
 joblib==0.14.1
 kiwisolver==1.2.0
 matplotlib==3.2.1
-mne==0.20.5
 nose==1.3.7
 numpy==1.18.5
 pandas==1.0.3

sample-data/SC4001E0-PSG.edf

@@ -0,0 +1,8 @@
+wave_4 7 1 3659 0:00 01/01/2001
+wave_4.dat 16x1000 1(1)/uV 15 0 5393 -12624 0 ECG
+wave_4.dat 16x100 2(1)/mg 13 0 1371 18209 0 Accelerometer_X
+wave_4.dat 16x100 2(1)/mg 13 0 1273 29556 0 Accelerometer_Y
+wave_4.dat 16x100 2(1)/mg 13 0 -473 -31584 0 Accelerometer_Z
+wave_4.dat 16 1(1) 15 0 32766 -29151 0 Marker
+wave_4.dat 16x5 1(-32767)/ms 15 0 -32768 13833 0 HRV
+wave_4.dat 16 -19.3447902992(3064)/degC 11 2048 2461 27731 0 DEV_Temperature

sample-data/SC4001E0_PSG.dat

@@ -63,7 +63,6 @@
         'joblib>=0.11',
         'kiwisolver>=1.1.0',
         'matplotlib>=2.0.0',
-        'mne>=0.18.0',
         'numpy>=1.10.1',
         'pandas>=0.17.0',
         'pyparsing>=2.0.4',

sample-data/SC4001E0_PSG.hea

@@ -256,7 +256,7 @@ def test_2f(self):
         """
         # Uniform sample rates
         record_MIT = wfdb.rdrecord('sample-data/n16').__dict__
-        record_EDF = wfdb.rdrecord('sample-data/n16.edf').__dict__
+        record_EDF = wfdb.edf2mit('sample-data/n16.edf').__dict__
 
         fields = list(record_MIT.keys())
         # Original MIT format method of checksum is outdated, sometimes
@@ -270,8 +270,13 @@ def test_2f(self):
         for field in fields:
             # Signal value will be slightly off due to C to Python type conversion
             if field == 'p_signal':
-                true_array = np.array(record_MIT[field])
-                pred_array = np.array(record_EDF[field])
+                true_array = np.array(record_MIT[field]).flatten()
+                pred_array = np.array(record_EDF[field]).flatten()
+                # Prevent divide by zero warning
+                for i,v in enumerate(true_array):
+                    if v == 0:
+                        true_array[i] = 1
+                        pred_array[i] = 1
                 sig_diff = np.abs((pred_array - true_array) / true_array)
                 sig_diff[sig_diff == -np.inf] = 0
                 sig_diff[sig_diff == np.inf] = 0
@@ -299,8 +304,8 @@ def test_2g(self):
 
         """
         # Non-uniform sample rates
-        record_MIT = wfdb.rdrecord('sample-data/SC4001E0_PSG').__dict__
-        record_EDF = wfdb.rdrecord('sample-data/SC4001E0-PSG.edf').__dict__
+        record_MIT = wfdb.rdrecord('sample-data/wave_4').__dict__
+        record_EDF = wfdb.edf2mit('sample-data/wave_4.edf').__dict__
 
         fields = list(record_MIT.keys())
         # Original MIT format method of checksum is outdated, sometimes
@@ -309,19 +314,18 @@ def test_2g(self):
         # Original MIT format units are less comprehensive since they
         # default to mV if unknown.. therefore added more default labels
         fields.remove('units')
-        # Initial value of signal will be off due to resampling done by
-        # MNE in the EDF reading phase
-        fields.remove('init_value')
-        # Samples per frame will be off due to resampling done by MNE in
-        # the EDF reading phase... I should probably fix this later
-        fields.remove('samps_per_frame')
 
         test_results = []
         for field in fields:
             # Signal value will be slightly off due to C to Python type conversion
             if field == 'p_signal':
-                true_array = np.array(record_MIT[field])
-                pred_array = np.array(record_EDF[field])
+                true_array = np.array(record_MIT[field]).flatten()
+                pred_array = np.array(record_EDF[field]).flatten()
+                # Prevent divide by zero warning
+                for i,v in enumerate(true_array):
+                    if v == 0:
+                        true_array[i] = 1
+                        pred_array[i] = 1
                 sig_diff = np.abs((pred_array - true_array) / true_array)
                 sig_diff[sig_diff == -np.inf] = 0
                 sig_diff[sig_diff == np.inf] = 0