Healthcare NLP v4.2.2 Release Notes

 

4.2.2

Highlights

  • Fine-tuning Relation Extraction models with your data
  • Added Romanian support in deidentification annotator for data obfuscation
  • New SDOH (Social Determinants of Health) ner model
  • Improved oncology models and 4 pretrained pipelines
  • New chunk mapper models to map entities (phrases) to their corresponding ICD-10-CM codes as well as clinical abbreviations to their definitions
  • New ICD-10-PCS sentence entity resolver model and ICD-10-CM resolver pipeline
  • New utility & helper modules documentation page
  • New and updated notebooks
  • 22 new clinical models and pipelines added & updated in total

Fine-Tuning Relation Extraction Models With Your Data

Instead of starting from scratch when training a new Relation Extraction model, you can train a new model by adding your new data to the pretrained model.

There are two new params in RelationExtractionApproach which allows you to initialize your model with the data from the pretrained model:

  • setPretrainedModelPath: This parameter allows you to point the training process to an existing model.
  • setОverrideExistingLabels: This parameter overrides the existing labels in the original model that are assigned the same output nodes in the new model. Default is True, when it is set to False the RelationExtractionApproach uses the existing labels and if it finds new ones it tries to assign them to unused output nodes.

Example:

reApproach_finetune = RelationExtractionApproach()\
    .setInputCols(["embeddings", "pos_tags", "train_ner_chunks", "dependencies"])\
    .setOutputCol("relations")\
    .setLabelColumn("rel")\
    ...
    .setFromEntity("begin1i", "end1i", "label1")\
    .setToEntity("begin2i", "end2i", "label2")\
    .setPretrainedModelPath("existing_RE_MODEL_path")\
    .setOverrideExistingLabels(False)

You can check Resume RelationExtractionApproach Training Notebook for more examples.

Added Romanian Support in Deidentification Annotator For Data Obfuscation

Deidentification annotator is now able to obfuscate entities (coming from a deid NER model) with fake data in Romanian language.

Example:

deid_obfuscated_faker = DeIdentification()\
    .setInputCols(["sentence", "token", "ner_chunk"]) \
    .setOutputCol("obfuscated") \
    .setMode("obfuscate")\
    .setLanguage('ro')\
    .setObfuscateDate(True)\
    .setObfuscateRefSource('faker')

text = """Nume si Prenume : BUREAN MARIA, Varsta: 77 ,Spitalul Pentru Ochi de Deal, Drumul Oprea Nr. 972 Vaslui"""

Result:

Sentence Masked with entity Masked with Chars Masked with Fixed Chars Obfuscated
Nume si Prenume : BUREAN MARIA, Varsta: 77 ,Spitalul Pentru Ochi de Deal, Drumul Oprea Nr. 972 Vaslui Nume si Prenume : <\PATIENT>, Varsta: <\AGE> ,<\HOSPITAL>, <\STREET> <\CITY> Nume si Prenume : ****, Varsta: ** ,********, ****** ** Nume si Prenume : **, Varsta: ** , **, ** ** Nume si Prenume : Claudia Crumble, Varsta: 18 ,LOS ANGELES AMBULATORY CARE CENTER, 706 north parrish avenue Piscataway

New SDOH (Social Determinants of Health) NER Model

  • Social Determinants of Health(SDOH) are the socioeconomic factors under which people live, learn, work, worship, and play that determine their health outcomes.The World Health Organization also provides a definition of social determinants of health. Social determinants of health as the conditions in which people are born, grow, live, work and age. These circumstances are shaped by the distribution of money, power, and resources at global, national, and local levels. Social determinants of health (SDOH) have a major impact on people’s health, well-being, and quality of life.
  • SDOH include lots of factors, also contribute to wide health disparities and inequities. In this project We have tried to define well these factors. The goal of this project is to train models for natural language processing focused on extracting terminology related to social determinants of health from various kinds of biomedical documents. This first model is Named Entity Recognition (NER) task.
  • The project is still ongoing and will mature over time and the number of sdoh factors (entities) will also be enriched. It will include other tasks as well.

Example:

ner_model = MedicalNerModel.pretrained("sdoh_slim_wip", "en", "clinical/models")\
        .setInputCols(["sentence", "token", "embeddings"])\
        .setOutputCol("ner")

text = """ Mother states that he does smoke, there is a family hx of alcohol on both maternal and paternal sides of the family, maternal grandfather who died of alcohol related complications and paternal grandmother with severe alcoholism. Pts own drinking began at age 16, living in LA, had a DUI at age 17 after totaling a new car that his mother bought for him, he was married. """

Result:

+-------------+-------------------+
|        token|          ner_label|
+-------------+-------------------+
|       Mother|    B-Family_Member|
|           he|           B-Gender|
|        smoke|          B-Smoking|
|      alcohol|          B-Alcohol|
|     maternal|    B-Family_Member|
|     paternal|    B-Family_Member|
|     maternal|    B-Family_Member|
|  grandfather|    B-Family_Member|
|      alcohol|          B-Alcohol|
|     paternal|    B-Family_Member|
|  grandmother|    B-Family_Member|
|       severe|          B-Alcohol|
|   alcoholism|          I-Alcohol|
|     drinking|          B-Alcohol|
|          age|              B-Age|
|           16|              I-Age|
|           LA|B-Geographic_Entity|
|          age|              B-Age|
|           17|              I-Age|
|          his|           B-Gender|
|       mother|    B-Family_Member|
|          him|           B-Gender|
|           he|           B-Gender|
|      married|   B-Marital_Status|
+-------------+-------------------+

Improved Oncology NER Models And 4 New Pretrained Pipelines

We are releasing the improved version of Oncological NER models (_wip) and 4 new pretrained oncological pipelines which are able to detect assertion status and relations between the extracted oncological entities.

NER model name (MedicalNerModel) description predicted entities
ner_oncology_anatomy_general Extracting anatomical entities. Anatomical_Site, Direction
ner_oncology_anatomy_granular Extracting anatomical entities using granular labels. Direction, Site_Lymph_Node, Site_Breast, Site_Other_Body_Part, Site_Bone, Site_Liver, Site_Lung, Site_Brain
ner_oncology_biomarker Extracting biomarkers and their results. Biomarker, Biomarker_Result
ner_oncology_demographics Extracting demographic information, including smoking status. Age, Gender, Smoking_Status, Race_Ethnicity
ner_oncology_diagnosis Extracting entities related to cancer diagnosis, including the presence of metastasis. Grade, Staging, Tumor_Size, Adenopathy, Pathology_Result, Histological_Type, Metastasis, Cancer_Score, Cancer_Dx, Invasion, Tumor_Finding, Performance_Status
ner_oncology Extracting more than 40 oncology-related entities. Histological_Type, Direction, Staging, Cancer_Score, Imaging_Test, Cycle_Number, Tumor_Finding, Site_Lymph_Node, Invasion, Response_To_Treatment, Smoking_Status, Tumor_Size, Cycle_Count, Adenopathy, Age, Biomarker_Result, Unspecific_Therapy, Site_Breast, Chemotherapy, Targeted_Therapy, Radiotherapy, Performance_Status, Pathology_Test, Site_Other_Body_Part, Cancer_Surgery, Line_Of_Therapy, Pathology_Result, Hormonal_Therapy, Site_Bone, Biomarker, Immunotherapy, Cycle_Day, Frequency, Route, Duration, Death_Entity, Metastasis, Site_Liver, Cancer_Dx, Grade, Date, Site_Lung, Site_Brain, Relative_Date, Race_Ethnicity, Gender, Oncogene, Dosage, Radiation_Dose
ner_oncology_posology This model extracts oncology specific posology information and cancer therapies. Cycle_Number, Cycle_Count, Radiotherapy, Cancer_Surgery, Cycle_Day, Frequency, Route, Cancer_Therapy, Duration, Dosage, Radiation_Dose
ner_oncology_unspecific_posology Extracting any mention of cancer therapies and posology information using general labels Cancer_Therapy, Posology_Information
ner_oncology_response_to_treatment_wip Extracting entities related to the patient’s response to cancer treatment. Response_To_Treatment, Size_Trend, Line_Of_Therapy
ner_oncology_therapy Extracting entities related to cancer therapies, including posology entities and response to treatment, using granular labels. Response_To_Treatment, Line_Of_Therapy, Cancer_Surgery, Radiotherapy, Immunotherapy, Targeted_Therapy, Hormonal_Therapy, Chemotherapy, Unspecific_Therapy, Route, Duration, Cycle_Count, Dosage, Frequency, Cycle_Number, Cycle_Day, Radiation_Dose
ner_oncology_test Extracting mentions of oncology-related tests. Oncogene, Biomarker, Biomarker_Result, Imaging_Test, Pathology_Test
ner_oncology_tnm Extracting mentions related to TNM staging. Lymph_Node, Staging, Lymph_Node_Modifier, Tumor_Description, Tumor, Metastasis, Cancer_Dx
Oncological Pipeline (PretrainedPipeline) Description
oncology_general_pipeline Includes Named-Entity Recognition, Assertion Status and Relation Extraction models to extract information from oncology texts. This pipeline extracts diagnoses, treatments, tests, anatomical references and demographic entities.
oncology_biomarker_pipeline Includes Named-Entity Recognition, Assertion Status and Relation Extraction models to extract information from oncology texts. This pipeline focuses on entities related to biomarkers
oncology_diagnosis_pipeline Includes Named-Entity Recognition, Assertion Status, Relation Extraction and Entity Resolution models to extract information from oncology texts. This pipeline focuses on entities related to oncological diagnosis.
oncology_therapy_pipeline Includes Named-Entity Recognition and Assertion Status models to extract information from oncology texts. This pipeline focuses on entities related to therapies.

Example:

from sparknlp.pretrained import PretrainedPipeline

pipeline = PretrainedPipeline("oncology_general_pipeline", "en", "clinical/models")

text = "The patient underwent a left mastectomy for a left breast cancer two months ago. The tumor is positive for ER and PR."

Result:

**** ner_oncology_wip results ****
| chunk          | ner_label        |
|:---------------|:-----------------|
| left           | Direction        |
| mastectomy     | Cancer_Surgery   |
| left           | Direction        |
| breast cancer  | Cancer_Dx        |
| two months ago | Relative_Date    |
| tumor          | Tumor_Finding    |
| positive       | Biomarker_Result |
| ER             | Biomarker        |
| PR             | Biomarker        |

**** assertion_oncology_wip results  ****
| chunk         | ner_label      | assertion   |
|:--------------|:---------------|:------------|
| mastectomy    | Cancer_Surgery | Past        |
| breast cancer | Cancer_Dx      | Present     |
| tumor         | Tumor_Finding  | Present     |
| ER            | Biomarker      | Present     |
| PR            | Biomarker      | Present     |

**** re_oncology_wip results ****
| chunk1        | entity1          | chunk2         | entity2       | relation      |
|:--------------|:-----------------|:---------------|:--------------|:--------------|
| mastectomy    | Cancer_Surgery   | two months ago | Relative_Date | is_related_to |
| breast cancer | Cancer_Dx        | two months ago | Relative_Date | is_related_to |
| tumor         | Tumor_Finding    | ER             | Biomarker     | O             |
| tumor         | Tumor_Finding    | PR             | Biomarker     | O             |
| positive      | Biomarker_Result | ER             | Biomarker     | is_related_to |
| positive      | Biomarker_Result | PR             | Biomarker     | is_related_to |

New Chunk Mapper Models to Map Entities (phrases) to Their Corresponding ICD-10-CM Codes As Well As Clinical Abbreviations to Their Definitions

We have 2 new chunk mapper models:

  • abbreviation_mapper_augmented is an augmented version of the existing abbreviation_mapper model. It maps abbreviations and acronyms of medical regulatory activities to their definitions.

  • icd10cm_mapper maps entities to corresponding ICD-10-CM codes.

Example:

chunkerMapper = ChunkMapperModel\
    .pretrained("icd10cm_mapper", "en", "clinical/models")\
    .setInputCols(["ner_chunk"])\
    .setOutputCol("mappings")\
    .setRels(["icd10cm_code"])

text = """A 35-year-old male with a history of primary leiomyosarcoma of neck, gestational diabetes mellitus diagnosed eight years prior to presentation and presented with a one-week history of polydipsia, poor appetite, and vomiting."""

Result:

+------------------------------+-------+------------+
|ner_chunk                     |entity |icd10cm_code|
+------------------------------+-------+------------+
|primary leiomyosarcoma of neck|PROBLEM|C49.0       |
|gestational diabetes mellitus |PROBLEM|O24.919     |
|polydipsia                    |PROBLEM|R63.1       |
|poor appetite                 |PROBLEM|R63.0       |
|vomiting                      |PROBLEM|R11.10      |
+------------------------------+-------+------------+

New ICD-10-PCS Sentence Entity Resolver Model and ICD-10-CM Resolver Pipeline

We are releasing new ICD-10-PCS resolver model and ICD-10-CM resolver pipeline:

  • sbiobertresolve_icd10pcs_augmented model maps extracted medical entities to ICD-10-PCS codes using sbiobert_base_cased_mli sentence bert embeddings. It trained on the augmented version of the dataset which is used in previous ICD-10-PCS resolver model.

Example:

icd10pcs_resolver = SentenceEntityResolverModel\
  .pretrained("sbiobertresolve_icd10pcs_augmented","en", "clinical/models") \
  .setInputCols(["ner_chunk", "sbert_embeddings"]) \
  .setOutputCol("resolution")\
  .setDistanceFunction("EUCLIDEAN")

text = "Given the severity of her abdominal examination and her persistence of her symptoms, it is detected that need for laparoscopic appendectomy and possible open appendectomy as well as pyeloplasty. We recommend performing a mediastinoscopy"

Result:

+-------------------------+---------+-------------+------------------------------------+--------------------+
|                ner_chunk|   entity|icd10pcs_code|                         resolutions|           all_codes|
+-------------------------+---------+-------------+------------------------------------+--------------------+
|    abdominal examination|     Test|      2W63XZZ|[traction of abdominal wall [trac...|[2W63XZZ, BW40ZZZ...|
|laparoscopic appendectomy|Procedure|      0DTJ8ZZ|[resection of appendix, endo [res...|[0DTJ8ZZ, 0DT84ZZ...|
|        open appendectomy|Procedure|      0DBJ0ZZ|[excision of appendix, open appro...|[0DBJ0ZZ, 0DTJ0ZZ...|
|              pyeloplasty|Procedure|      0TS84ZZ|[reposition bilateral ureters, pe...|[0TS84ZZ, 0TS74ZZ...|
|          mediastinoscopy|Procedure|      BB1CZZZ|[fluoroscopy of mediastinum [fluo...|[BB1CZZZ, 0WJC4ZZ...|
+-------------------------+---------+-------------+------------------------------------+--------------------+
  • icd10cm_resolver_pipeline pretrained pipeline maps entities with their corresponding ICD-10-CM codes. You’ll just feed your text and it will return the corresponding ICD-10-CM codes.

Example:

from sparknlp.pretrained import PretrainedPipeline

resolver_pipeline = PretrainedPipeline("icd10cm_resolver_pipeline", "en", "clinical/models")

text = "A 28-year-old female with a history of gestational diabetes mellitus diagnosed eight years and anisakiasis. Also, it was reported that fetal and neonatal hemorrhage"

Result:

+-----------------------------+---------+------------+
|chunk                        |ner_chunk|icd10cm_code|
+-----------------------------+---------+------------+
|gestational diabetes mellitus|PROBLEM  |O24.919     |
|anisakiasis                  |PROBLEM  |B81.0       |
|fetal and neonatal hemorrhage|PROBLEM  |P545        |
+-----------------------------+---------+------------+

New Utility & Helper Modules Documentation Page

We have a new utility & helper modules documentation page that you can find the documentations of Spark NLP for Healthcare modules with examples.

New and Updated Notebooks

22 New Clinical Models and Pipelines Added & Updated in Total

  • abbreviation_mapper_augmented
  • icd10cm_mapper
  • sbiobertresolve_icd10pcs_augmented
  • icd10cm_resolver_pipeline
  • oncology_biomarker_pipeline
  • oncology_diagnosis_pipeline
  • oncology_therapy_pipeline
  • oncology_general_pipeline
  • ner_oncology_anatomy_general
  • ner_oncology_anatomy_granular
  • ner_oncology_biomarker
  • ner_oncology_demographics
  • ner_oncology_diagnosis
  • ner_oncology
  • ner_oncology_posology
  • ner_oncology_response_to_treatment
  • ner_oncology_test
  • ner_oncology_therapy
  • ner_oncology_tnm
  • ner_oncology_unspecific_posology
  • sdoh_slim_wip
  • t5_base_pubmedqa

For all Spark NLP for healthcare models, please check: Models Hub Page

Versions

Last updated