Healthcare NLP v2.7.6 Release Notes

Highlights:

• New pretrained Radiology Assertion Status model to assign Confirmed, Suspected, Negative assertion scopes to imaging findings or any clinical tests.

• Obfuscating the same sensitive information (patient or doctor name) with the same fake names across the same clinical note.

• Version compatibility checker for the pretrained clinical models and builds to keep up with the latest development efforts in production.

• Adding more English names to faker module in Deidentification.

• Updated & improved clinical SentenceDetectorDL model.

• New upgrades on ner_deid_large and ner_deid_enriched NER models to cover more use cases with better resolutions.

• Adding more examples to workshop repo for Scala users to practice more on healthcare annotators.

• Bug fixes & general improvements.

1. Radiology Assertion Status Model

We trained a new assertion model to assign Confirmed, Suspected, Negative assertion scopes to imaging findings or any clinical tests. It will try to assign these statuses to any named entity you would feed to the assertion annotater in the same pipeline.

 radiology_assertion = AssertionDLModel.pretrained("assertion_dl_radiology", "en", "clinical/models")\
 .setInputCols(["sentence", "ner_chunk", "embeddings"])\
 .setOutputCol("assertion")

text = Blunting of the left costophrenic angle on the lateral view posteriorly suggests a small left pleural effusion. No right-sided pleural effusion or pneumothorax is definitively seen. There are mildly displaced fractures of the left lateral 8th and likely 9th ribs.

You can also use this with AssertionFilterer to return clinical findings from a note only when it is i.e. confirmed or suspected.

 assertion_filterer = AssertionFilterer()\
 .setInputCols("sentence","ner_chunk","assertion")\
 .setOutputCol("assertion_filtered")\
 .setWhiteList(["confirmed","suspected"])

 >> ["displaced fractures", "effusion"]

2. Obfuscating with the same fake name across the same note:

 obfuscation = DeIdentification()\
  .setInputCols(["sentence", "token", "ner_chunk"]) \
  .setOutputCol("deidentified") \
  .setMode("obfuscate")\
  .setObfuscateDate(True)\
  .setSameEntityThreshold(0.8)\
  .setObfuscateRefSource("faker")


text =''' Provider: David Hale, M.D.
          Pt: Jessica Parker
          David told  Jessica that she will need to visit the clinic next month.'''

3. Library Version Compatibility Table :

We are releasing the version compatibility table to help users get to see which Spark NLP licensed version is built against which core (open source) version. We are going to release a detailed one after running some tests across the jars from each library.

4. Pretrained Models Version Control :

Due to active release cycle, we are adding & training new pretrained models at each release and it might be tricky to maintain the backward compatibility or keep up with the latest models, especially for the users using our models locally in air-gapped networks.

We are releasing a new utility class to help you check your local & existing models with the latest version of everything we have up to date. This is an highly experimental feature of which we plan to improve and add more capability later on.

from sparknlp_jsl.check_compatibility import Compatibility

 checker = sparknlp_jsl.Compatibility()

 result = checker.find_version(aws_access_key_id=license_keys['AWS_ACCESS_KEY_ID'],
                        aws_secret_access_key=license_keys['AWS_SECRET_ACCESS_KEY'],
                        metadata_path=None,
                        model = 'all' , # or a specific model name
                        target_version='all',
                        cache_pretrained_path='/home/ubuntu/cache_pretrained')

 >> result['outdated_models']

  [{'model_name': 'clinical_ner_assertion',
    'current_version': '2.4.0',
    'latest_version': '2.6.4'},
   {'model_name': 'jsl_rd_ner_wip_greedy_clinical',
    'current_version': '2.6.1',
    'latest_version': '2.6.2'},
   {'model_name': 'ner_anatomy',
    'current_version': '2.4.2',
    'latest_version': '2.6.4'},
   {'model_name': 'ner_aspect_based_sentiment',
    'current_version': '2.6.2',
    'latest_version': '2.7.2'},
   {'model_name': 'ner_bionlp',
    'current_version': '2.4.0',
    'latest_version': '2.7.0'},
   {'model_name': 'ner_cellular',
    'current_version': '2.4.2',
    'latest_version': '2.5.0'}]

  >> result['version_comparison_dict']

  [{'clinical_ner_assertion': {'current_version': '2.4.0', 'latest_version': '2.6.4'}}, {'jsl_ner_wip_clinical': {'current_version': '2.6.5', 'latest_version': '2.6.1'}}, {'jsl_ner_wip_greedy_clinical': {'current_version': '2.6.5', 'latest_version': '2.6.5'}}, {'jsl_ner_wip_modifier_clinical': {'current_version': '2.6.4', 'latest_version': '2.6.4'}}, {'jsl_rd_ner_wip_greedy_clinical': {'current_version': '2.6.1','latest_version': '2.6.2'}}]

5. Updated Pretrained Models:

(requires fresh .pretraned())

• ner_deid_large
• ner_deid_enriched