Healthcare NLP v5.0.2 Release Notes

5.0.2

Highlights

We are delighted to announce a suite of remarkable enhancements and updates in our latest release of Spark NLP for Healthcare. This release comes with the first Text2SQL module and ONNX-optimized medical text summarization models as well as 20 new clinical pretrained models and pipelines. It is a testament to our commitment to continuously innovate and improve, furnishing you with a more sophisticated and powerful toolkit for healthcare natural language processing.

• Text2SQL module to translate text prompts into accurate SQL queries
• Support for ONNX integration of Seq2Seq models such as MedicalTextGenerator, MedicalSummarizer, and MedicalQuestionAnswering
• 2 new medical QA models and 1 new summarization model optimized with ONNX.
• Brand new clinical NER model for extracting clinical entities in the Portuguese language
• 3 novel assertion status (negativity scope) detection models tailored for entities extracted from Voice of Patient (VoP) notes
• Detecting assertion statuses of entities related to Social Determinants of Health (SDOH) identified within clinical notes
• Classifying transportation insecurity within the context of Social Determinants of Health (SDOH)
• Text Classifier models to infer Age Groups from health records, even in the absence of explicit age indications or mentions.
• Mapping ICD-10-CM codes to Medicare Severity-Diagnosis Related Group (MS-DRG)
• Five new sentence entity resolver (terminology mapping) pretrained pipelines, designed to streamline solutions with a single line of code
• Various core improvements; bug fixes, enhanced overall robustness and reliability of Spark NLP for Healthcare
  • Improvement of the deidentification faker list (city, street, hospital, profession) for various language
• Updated notebooks and demonstrations for making Spark NLP for Healthcare easier to navigate and understand
• The addition and update of numerous new clinical models and pipelines continue to reinforce our offering in the healthcare domain

We believe that these enhancements will elevate your experience with Spark NLP for Healthcare, enabling more efficient, accurate, and streamlined analysis of healthcare-related natural language data.

Text2SQL module to translate text prompts into accurate SQL queries

We are excited to introduce our latest innovation, the Text2SQL annotator. This powerful tool revolutionizes the way you interact with databases by effortlessly translating natural language text prompts into accurate and effective SQL queries. With the integration of a state-of-the-art LLM, this annotator opens new possibilities for enhanced data retrieval and manipulation, streamlining your workflow and boosting efficiency.

Also we have a new text2sql_mimicsql model that is specifically finetuned on MIMIC-III dataset schema for enhancing the precision of SQL queries derived from medical natural language queries on MIMIC dataset. Please check the model card for more details.

Example:

text2sql = Text2SQL.pretrained("text2sql_mimicsql", "en", "clinical/models")\
    .setInputCols(["document"])\
    .setOutputCol("sql")

sample_text = "Calulate the total number of patients who had icd9 code 5771"

Results:

Please check: Text to SQL Generation Notebook for more information

Support for ONNX integration of Medical Seq2Seq models

ONNX integration empowers our Seq2Seq models to perform their tasks more efficiently and effectively. By leveraging the optimized capabilities of these LLMs through ONNX, the processing speed and overall performance of these models are substantially improved.

2 New Medical QA Models and 1 New Summarization Model Optimized with ONNX.

We’re excited to introduce the ONNX-powered versions of summarizer_clinical_laymen, clinical_notes_qa_base and clinical_notes_qa_large models, representing a significant advancement in efficiency and versatility.

Example:

med_qa = MedicalQuestionAnswering()\
    .pretrained("clinical_notes_qa_base_onnx", "en", "clinical/models")\
    .setInputCols(["document_question", "document_context"])\
    .setCustomPrompt("Context: {context} \n Question: {question} \n Answer: ")\
    .setOutputCol("answer")\

note_text = "Patient with a past medical history of hypertension for 15 years.\n(Medical Transcription Sample Report)\nHISTORY OF PRESENT ILLNESS:\nThe patient is a 74-year-old white woman who has a past medical history of hypertension for 15 years, history of CVA with no residual hemiparesis and uterine cancer with pulmonary metastases, who presented for evaluation of recent worsening of the hypertension. According to the patient, she had stable blood pressure for the past 12-15 years on 10 mg of lisinopril."

question = "What is the primary issue reported by patient?"

Results:

Brand New Clinical NER Model For Extracting Clinical Entities In The Portuguese Language

Portuguese clinical NER models provide valuable tools for processing and analyzing Portuguese clinical texts. They assist in automating the extraction of important clinical information, facilitating research, medical documentation, and other applications within the Portuguese healthcare domain.

For more details, please check the model card.

Example:

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

sample_text = """A paciente apresenta sensibilidade dentária ao consumir alimentos quentes e frios.
Realizou-se um exame clínico e radiográfico para avaliar possíveis cáries ou problemas na raiz do dente."""

Results:

3 Novel Assertion Status (Negativity Scope) Detection Models Tailored for Entities Extracted from Voice of Patient (VoP) Notes

We are excited to announce 3 new assertion status detection models that can classify assertions for the detected entities in VoP notes with the Hypothetical_Or_Absent, Present_Or_Past, and SomeoneElse labels.

Example:

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

sample_text = "I was feeling anxiety honestly. Can it bring on tremors? It was right after my friend was diagnosed with diabetes."

Results:

Detecting Assertion Statuses (Negativity Scope) of Entities Related to Social Determinants of Health (SDOH) Identified within Clinical Notes

We are introducing a new assertion_sdoh_wip assertion status detection model that can classify assertions for the detected SDOH entities in text into six distinct labels: Absent, Present, Someone_Else, Past, Hypothetical, and Possible.

For more details, please check the model card.

Example:

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

sample_text = "Smith works as a cleaning assistant and does not have access to health insurance or paid sick leave. But she has generally housing problems. She lives in a apartment now.  She has long history of EtOH abuse, beginning in her teens. She is aware she needs to attend Rehab Programs."

Results:

Classifying Transportation Insecurity within the Context of Social Determinants of Health (SDOH)

Introducing two new transportation insecurity classifier models for SDOH that offer precise label assignments and confidence scores. With a strong ability to thoroughly analyze text, these models categorize content into No_Transportation_Insecurity_Or_Unknown and Transportation_Insecurity, providing valuable insights into transportation-related insecurity.

Example:

generic_classifier = GenericClassifierModel.pretrained("genericclassifier_sdoh_transportation_insecurity_sbiobert_cased_mli", 'en', 'clinical/models')\
    .setInputCols(["features"])\
    .setOutputCol("prediction")

sample_text_list = ["Patient B is a 40-year-old female who was diagnosed with breast cancer. She has received a treatment plan that includes surgery, chemotherapy, and radiation therapy. She is alone and can not drive a car or can not use public bus. ", "Lisa, a 28-year-old woman, was diagnosed with generalized anxiety disorder (GAD), a mental disorder characterized by excessive worry and persistent anxiety."]

Results:

Text Classifier Models to Infer Age Groups from Health Records, Even in the Absence of Explicit Age Indications or Mentions.

Now we have three new Age Group Text Classifier models that are designed to analyze the age group of individuals mentioned in health documents, whether or not the age is explicitly mentioned in the training data. These models are trained using in-house annotated health-related text, and categorized into three classes:

• Adult: Refers to individuals who are fully grown or developed, usually 18 years or older.
• Child: A young human who is not yet an adult. Typically refers to someone below the legal age of adulthood, which is often below 18 years old.
• Unknown: Represents situations where determining the age group from the given text is not possible.

Example:

few_shot_classifier = FewShotClassifierModel.pretrained("few_shot_classifier_age_group_sbiobert_cased_mli", "en", "clinical/models")\
    .setInputCols(["sentence_embeddings"])\
    .setOutputCol("prediction")

sample_text_list = [["A patient presented with complaints of chest pain and shortness of breath. The medical history revealed the patient had a smoking habit for over 30 years, and was diagnosed with hypertension two years ago. After a detailed physical examination, the doctor found a noticeable wheeze on lung auscultation and prescribed a spirometry test, which showed irreversible airway obstruction. The patient was diagnosed with Chronic obstructive pulmonary disease (COPD) caused by smoking."],
 ["Hi, wondering if anyone has had a similar situation. My 1 year old daughter has the following; loose stools/ pale stools, elevated liver enzymes, low iron.  5 months and still no answers from drs. "],
 ["Hi have chronic gastritis from 4 month(confirmed by endoscopy).I do not have acid reflux.Only dull ache above abdomen and left side of chest.I am on reberprozole and librax.My question is whether chronic gastritis is curable or is it a lifetime condition?I am loosing hope because this dull ache is not going away.Please please reply"]]

Results:

Mapping ICD-10-CM Codes to Medicare Severity-Diagnosis Related Group (MS-DRG)

We have a new icd10cm_ms_drg_mapper chunk mapper model that maps the ICD-10-CM codes with their corresponding Medicare Severity-Diagnosis Related Group (MS-DRG).

Example:

chunkMapper = DocMapperModel.pretrained("icd10cm_ms_drg_mapper", "en", "clinical/models")\
      .setInputCols(["icd_chunk"])\
      .setOutputCol("mappings")\
      .setRels(["ms-drg"])

sample_codes = ["L08.1", "U07.1", "C10.0", "J351"]

Results:

For more details, please check the model card.

Five New Sentence Entity Resolver (Terminology Mapping) Pretrained Pipelines, Designed to Streamline Solutions with a Single Line of Code

We have five new sentence entity resolver pipelines that are meticulously designed to enhance your solutions by efficiently identifying entities and their resolutions within the clinical note. You can easily integrate this advanced capability using just a single line of code.

from sparknlp.pretrained import PretrainedPipeline

abbr_pipeline = PretrainedPipeline("abbreviation_pipeline", "en", "clinical/models")

result = abbr_pipeline.fullAnnotate("""Gravid with estimated fetal weight of 6-6/12 pounds.
LABORATORY DATA: Laboratory tests include a CBC which is normal. 
VDRL: Nonreactive""")

Results:

Various Core Improvements: Bug Fixes, Enhanced Overall Robustness, And Reliability Of Spark NLP For Healthcare

• Improvement of the deidentification faker list (city, street, hospital, profession) for various language

Updated Notebooks And Demonstrations For making Spark NLP For Healthcare Easier To Navigate And Understand

• New Extracting Public Health Related_Insights From Social Media Texts Using Healthcare NLP Notebook for automated health information extraction and co-occurrence analysis with JohnSnowLabs models
• Text to SQL Generation Notebook for automatically converting natural language questions into corresponding SQL queries
• Updated NORMALIZED SECTION HEADER MAPPER Demo with ner_section_header_diagnosis model
• Updated ENTITY RESOLVER LOINC Demo with sbiobertresolve_loinc_numeric and sbiobertresolve_loinc_augmented models

We Have Added And Updated A Substantial Number Of New Clinical Models And Pipelines, Further Solidifying Our Offering In The Healthcare Domain.

• summarizer_clinical_laymen_onnx
• clinical_notes_qa_large_onnx
• clinical_notes_qa_base_onnx
• ner_clinical -> pt
• text2sql_mimicsql
• assertion_sdoh_wip
• genericclassifier_sdoh_transportation_insecurity_e5_large
• genericclassifier_sdoh_transportation_insecurity_sbiobert_cased_mli
• bert_sequence_classifier_age_group
• genericclassifier_age_group_sbiobert_cased_mli
• icd10cm_ms_drg_mapper
• abbreviation_pipeline
• rxnorm_resolver_pipeline
• icd10cm_multi_mapper_pipeline
• rxnorm_multi_mapper_pipeline
• snomed_multi_mapper_pipeline
• assertion_vop_clinical
• assertion_vop_clinical_medium
• assertion_vop_clinical_large
• few_shot_classifier_age_group_sbiobert_cased_mli

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

Versions