ECG Autoencoder PheWAS in the Model Zoo (#514)

* categorical and continuous composite TensorMaps, bump version, add ecg phewas to model zoo

.gitattributes

@@ -6,3 +6,4 @@ notebooks/**.ipynb filter=nbstripout
 *.linux filter=lfs diff=lfs merge=lfs -text
 *.osx filter=lfs diff=lfs merge=lfs -text
 *.genes filter=lfs diff=lfs merge=lfs -text
+model_zoo/ECG_PheWAS/*.h5 filter=lfs diff=lfs merge=lfs -text

.github/workflows/docker-publish.yml

@@ -16,7 +16,7 @@ on:
 #      - v*
 
   # Run tests for any PRs.
-  pull_request:
+  # pull_request:
 
 env:
   IMAGE_NAME: ml4h_terra

docker/vm_boot_images/config/tensorflow-requirements.txt

@@ -34,8 +34,8 @@ plotnine
 vega
 ipycanvas>=0.7.0
 ipyannotations>=0.2.1
-torch
+torch==1.12.1
 opencv-python
 blosc
 boto3
-ml4ht==0.0.9
+ml4ht==0.0.10

ml4h/defines.py

@@ -58,7 +58,6 @@ def __str__(self):
     'anterior_papillary': 9, 'LV_cavity': 10, 'LA_cavity': 11, 'body': 12,
 }
 LAX_2CH_HEART_LABELS = {
-    'aortic_arch': 1, 'left_pulmonary_artery_wall': 2, 'left_pulmonary_artery': 3,
     'LA_appendage': 4, 'LA_free_wall': 5, 'LV_posterior_wall': 6, 'LV_anterior_wall': 7, 'posterior_papillary': 8,
     'anterior_papillary': 9, 'LV_cavity': 10, 'LA_cavity': 11,
 }
@@ -88,10 +87,9 @@ def __str__(self):
     'RV_cavity': 5, 'thoracic_cavity': 6, 'liver': 7, 'stomach': 8, 'spleen': 9, 'kidney': 11, 'body': 10,
     'left_atrium': 12, 'right_atrium': 13, 'aorta': 14, 'pulmonary_artery': 15,
 }
-MRI_SAX_SEGMENTED_CHANNEL_MAP = {
-    'background': 0, 'RV_free_wall': 1, 'interventricular_septum': 2, 'LV_free_wall': 3, 'LV_cavity': 4,
-    'RV_cavity': 5, 'thoracic_cavity': 6, 'liver': 7, 'stomach': 8, 'spleen': 9, 'kidney': 11, 'body': 10,
-    'left_atrium': 12, 'right_atrium': 13, 'aorta': 14, 'pulmonary_artery': 15,
+SAX_HEART_LABELS = {
+    'RV_free_wall': 1, 'interventricular_septum': 2, 'LV_free_wall': 3, 'LV_cavity': 4,
+    'RV_cavity': 5, 'left_atrium': 12, 'right_atrium': 13,
 }
 MRI_AO_SEGMENTED_CHANNEL_MAP = {
     'ao_background': 0, 'ao_superior_vena_cava': 1, 'ao_pulmonary_artery': 2, 'ao_ascending_aortic_wall': 3, 'ao_ascending_aorta': 4,

ml4h/explorations.py

@@ -1045,7 +1045,13 @@ def explore(args):
 
 def latent_space_dataframe(infer_hidden_tsv, explore_csv):
     df = pd.read_csv(explore_csv)
-    df['sample_id'] = pd.to_numeric(df['sample_id'], errors='coerce')
+    if 'sample_id' in df.columns:
+        id_col = 'sample_id'
+    elif 'fpath' in df.columns:
+        id_col = 'fpath'
+    else:
+        raise ValueError(f'Could not find a sample ID column in explore CSV:{explore_csv}')
+    df['sample_id'] = pd.to_numeric(df[id_col], errors='coerce')
     df2 = pd.read_csv(infer_hidden_tsv, sep='\t', engine='python')
     df2['sample_id'] = pd.to_numeric(df2['sample_id'], errors='coerce')
     latent_df = pd.merge(df, df2, on='sample_id', how='inner')

ml4h/models/pretrained_blocks.py

@@ -80,8 +80,8 @@ def __init__(
             *,
             tensor_map: TensorMap,
             pretrain_trainable: bool,
-            base_model = "https://tfhub.dev/jeongukjae/roberta_en_cased_L-24_H-1024_A-16/1",
-            preprocess_model="https://tfhub.dev/jeongukjae/roberta_en_cased_preprocess/1",
+            base_model="https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/3",
+            preprocess_model="https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3",
             **kwargs,
     ):
         self.tensor_map = tensor_map

ml4h/plots.py

@@ -1032,6 +1032,9 @@ def plot_survivorship(
     :param title: Title for the plot
     :param prefix: Path prefix where plot will be saved
     :param days_window: Maximum days of follow up
+    :param dpi: Dots per inch of the figure
+    :param width: Width in inches of the figure
+    :param height: Height in inches of the figure
     """
     plt.figure(figsize=(width, height), dpi=dpi)
     days_sorted_index = np.argsort(days_follow_up)
@@ -1100,6 +1103,9 @@ def plot_survival(
     :param title: Title for the plot
     :param days_window: Maximum days of follow up
     :param prefix: Path prefix where plot will be saved
+    :param dpi: Dots per inch of the figure
+    :param width: Width in inches of the figure
+    :param height: Height in inches of the figure
 
     :return: Dictionary mapping metric names to their floating point values
     """
@@ -1109,7 +1115,7 @@ def plot_survival(
     plt.figure(figsize=(width, height), dpi=dpi)
 
     cumulative_sick = np.cumsum(np.sum(truth[:, intervals:], axis=0))
-    cumulative_censored = (truth.shape[0]-np.sum(truth[:, :intervals], axis=0))-cumulative_sick
+    cumulative_censored = (truth.shape[0]-np.sum(truth[:, :intervals], axis=0)) - cumulative_sick
     alive_per_step = np.sum(truth[:, :intervals], axis=0)
     sick_per_step = np.sum(truth[:, intervals:], axis=0)
     survivorship = np.cumprod(1 - (sick_per_step / alive_per_step))

ml4h/tensorize/dataflow/bigquery_ukb_queries.py

@@ -77,11 +77,14 @@ def write_tensor_from_sql(sampleid_to_rows, output_path, tensor_type):
                     hd5.create_dataset('icd', (1,), data=JOIN_CHAR.join(icds), dtype=h5py.special_dtype(vlen=str))
                 elif tensor_type == 'categorical':
                     for row in rows:
-                        hd5_dataset_name = dataset_name_from_meaning('categorical', [row['field'], row['meaning'], str(row['instance']), str(row['array_idx'])])
+                        fields = [str(row['fieldid']), row['field'], row['meaning'],
+                                  str(row['instance']), str(row['array_idx'])]
+                        hd5_dataset_name = dataset_name_from_meaning('categorical', fields)
                         _write_float_or_warn(sample_id, row, hd5_dataset_name, hd5)
                 elif tensor_type == 'continuous':
                     for row in rows:
-                        hd5_dataset_name = dataset_name_from_meaning('continuous', [str(row['fieldid']), row['field'], str(row['instance']), str(row['array_idx'])])
+                        fields = [str(row['fieldid']), row['field'], str(row['instance']), str(row['array_idx'])]
+                        hd5_dataset_name = dataset_name_from_meaning('continuous', fields)
                         _write_float_or_warn(sample_id, row, hd5_dataset_name, hd5)
                 elif tensor_type in ['disease', 'phecode_disease']:
                     for row in rows:

ml4h/tensorize/tensor_writer_ukbb.py

@@ -786,11 +786,9 @@ def _write_ecg_bike_tensors(ecgs, xml_field, hd5, sample_id, stats):
     for ecg in ecgs:
         root = et.parse(ecg).getroot()
         date = datetime.datetime.strptime(_date_str_from_ecg(root), '%Y-%m-%d')
-        write_to_hd5 = partial(create_tensor_in_hd5, hd5=hd5, path_prefix='ukb_ecg_bike', stats=stats, date=date)
-        logging.info('Got ECG for sample:{} XML field:{}'.format(sample_id, xml_field))
-
         instance = ecg.split(JOIN_CHAR)[-2]
-        write_to_hd5(storage_type=StorageType.STRING, name='instance', value=instance)
+        write_to_hd5 = partial(create_tensor_in_hd5, hd5=hd5, path_prefix='ukb_ecg_bike', instance=instance, stats=stats, date=date)
+        logging.info(f'Got ECG for sample:{sample_id} XML field:{xml_field}')
 
         protocol = root.findall('./Protocol/Phase')[0].find('ProtocolName').text
         write_to_hd5(storage_type=StorageType.STRING, name='protocol', value=protocol)
@@ -881,10 +879,13 @@ def _write_ecg_bike_tensors(ecgs, xml_field, hd5, sample_id, stats):
                 if field_val is False:
                     continue
                 trends[lead_field][i, lead_to_int[lead_num]] = field_val
-            trends['time'][i] = SECONDS_PER_MINUTE * int(trend_entry.find("EntryTime/Minute").text) + int(trend_entry.find("EntryTime/Second").text)
-            trends['PhaseTime'][i] = SECONDS_PER_MINUTE * int(trend_entry.find("PhaseTime/Minute").text) + int(trend_entry.find("PhaseTime/Second").text)
-            trends['PhaseName'][i] = phase_to_int[trend_entry.find('PhaseName').text]
-            trends['Artifact'][i] = float(trend_entry.find('Artifact').text.strip('%')) / 100  # Artifact is reported as a percentage
+            try:
+                trends['time'][i] = SECONDS_PER_MINUTE * int(trend_entry.find("EntryTime/Minute").text) + int(trend_entry.find("EntryTime/Second").text)
+                trends['PhaseTime'][i] = SECONDS_PER_MINUTE * int(trend_entry.find("PhaseTime/Minute").text) + int(trend_entry.find("PhaseTime/Second").text)
+                trends['PhaseName'][i] = phase_to_int[trend_entry.find('PhaseName').text]
+                trends['Artifact'][i] = float(trend_entry.find('Artifact').text.strip('%')) / 100  # Artifact is reported as a percentage
+            except AttributeError as e:
+                stats['AttributeError on Trend Data'] += 1
 
         for field, trend_list in trends.items():
             write_to_hd5(name=f'trend_{str.lower(field)}', value=trend_list)
@@ -900,12 +901,15 @@ def _write_ecg_bike_tensors(ecgs, xml_field, hd5, sample_id, stats):
             write_to_hd5(name=f'{str.lower(phase_name)}_duration', value=[phase_duration])
 
         # HR stats
-        max_hr = _xml_path_to_float(root, './ExerciseMeasurements/MaxHeartRate')
-        resting_hr = _xml_path_to_float(root, './ExerciseMeasurements/RestingStats/RestHR')
-        max_pred_hr = _xml_path_to_float(root, './ExerciseMeasurements/MaxPredictedHR')
-        write_to_hd5(name='max_hr', value=[max_hr])
-        write_to_hd5(name='resting_hr', value=[resting_hr])
-        write_to_hd5(name='max_pred_hr', value=[max_pred_hr])
+        try:
+            max_hr = _xml_path_to_float(root, './ExerciseMeasurements/MaxHeartRate')
+            write_to_hd5(name='max_hr', value=[max_hr])
+            resting_hr = _xml_path_to_float(root, './ExerciseMeasurements/RestingStats/RestHR')
+            write_to_hd5(name='resting_hr', value=[resting_hr])
+            max_pred_hr = _xml_path_to_float(root, './ExerciseMeasurements/MaxPredictedHR')
+            write_to_hd5(name='max_pred_hr', value=[max_pred_hr])
+        except AttributeError as e:
+            stats['AttributeError on HR Stats'] += 1
 
 
 def _write_tensors_from_niftis(folder: str, hd5: h5py.File, field_id: str, stats: Counter):