def _load_examples_from_file():
"""Loads the examples from a gzipped-ndjson file."""
for file in os.listdir(FLAGS.src_folder):
if not file.endswith('.gz'):
continue
with gzip.open(os.path.join(FLAGS.src_folder, file), 'r') as f:
for line in f:
yield features.build_example(json.loads(line))
def main(data, context):
"""Extracts features from a patient bundle for online prediction.
This process is broken down into a few steps:
1. Fetch the Resource we get triggered on, and fetch/extract the patient that
it is related to.
2. Fetch everything for the patient from step 1, and extract the
features we are interested in.
3. Send the features to Cloud ML for online prediction, and write the
results back to the FHIR store.
Args:
data (dict): Cloud PubSub payload. The `data` field is what we are looking
for.
context (google.cloud.functions.Context): Metadata for the event.
"""
if 'data' not in data:
LOGGER.info('`data` field is not present, skipping...')
return
resource_name = base64.b64decode(data['data']).decode('utf-8')
if (utils.CONDITION_TYPE not in resource_name and
utils.PATIENT_TYPE not in resource_name and
utils.OBSERVATION_TYPE not in resource_name):
LOGGER.info('Skipping resource %s which is irrelevant for prediction.',
resource_name)
return
credentials, _ = google.auth.default()
http = AuthorizedHttp(credentials)
resource = get_resource(http, resource_name)
if resource is None:
return
patient = get_corresponding_patient(http, resource_name, resource)
if patient is None:
LOGGER.error('Could not find corresponding patient in resource %s',
resource_name)
return
project_id, location, dataset_id, fhir_store_id, _ = _parse_resource_name(
resource_name)
patient_id = 'Patient/{}'.format(patient['id'])
patient_name = _construct_resource_name(project_id, location, dataset_id,
fhir_store_id, patient_id)
patient_bundle = get_patient_everything(http, patient_name)
if patient_bundle is None:
return
predictions = predict(features.build_example(patient_bundle))
if predictions is None:
return
action = get_action(data)
create_or_update_risk_assessment(http, patient_name, predictions, action)
def _load_examples_from_gcs():
"""Downloads the examples from a GCS bucket."""
client = storage.Client()
bucket = storage.Bucket(client, FLAGS.src_bucket)
for blob in bucket.list_blobs(prefix=FLAGS.src_folder):
if not blob.name.endswith('.gz'):
continue
print('Downloading patient record file', blob.name)
with tempfile.NamedTemporaryFile() as compressed_f:
blob.download_to_filename(compressed_f.name)
print('Building TF records')
with gzip.open(compressed_f.name, 'r') as f:
for line in f:
yield features.build_example(json.loads(line.decode('utf-8')))