Healthcare NLP Models

Learn about the Language service Healthcare NLP models to extract entities from healthcare records such as electronic health records (EHR), progress notes, and clinical trial documents.

The healthcare models constitute a foundational layer for business use cases and other AI services. These business units in Oracle aim to leverage AI/ML building blocks offered by OCI Language Services to build applications and ML models for use cases such as Readmission Predictive Risk Models, disease-specific Risk Models, Clinical decision support systems, and so on, for which OCI Language Services must develop foundational healthcare NLP models such as Health entity extraction, Health entity linking to medical standards, Assertion Status Detection and Relation Prediction. These healthcare NLP models are built in the framework OCI Healthcare services, using deep learning techniques.

The Healthcare NLP model is used to process healthcare text records such as EHR to extract entities, determine assertion statuses, identify related entities, and link those entities with supported ontologies

Healthcare NLP Model Types

Healthcare NLP is a suite of four models:

Health Named Entity extraction or Health Named Entity Recognition (HNER)

The aim of the task is to find and classify named entities mentioned in unstructured text into categories such as person names, medical conditions, medications, dosages, symptoms, test results, treatments, and procedures, and so on.

Example: Bold key phrases denote spans which appear along with mapped entity types in parentheses.

Using entity types:

MEDICINE_NAME
QUALIFIER.MODIFIER
MEDICINE_STRENGTH
MEDICINE_FREQUENCY

"Tacrolimus (MEDICINE_NAME) taper (QUALIFIER.MODIFIER) halted (QUALIFIER.MODIFIER), now at 2.5mg (MEDICINE_STRENGTH) BID (MEDICINE_FREQUENCY)"

Health Relation Extraction/Health Relation Prediction (HRE)

The aim of the task is to identify possible semantic relations that can occur between the entities. For example, the relation between medicine and its dosage in the healthcare text.

Example: Bold key phrases denote spans which appear along with mapped entity types in parentheses.

Using entity types:

MEDICINE_DURATION
MEDICINE_NAME
REGIMEN_THERAPY
QUALIFIER.MODIFIER

"She has received 4 cycles (MEDICINE_DURATION) of Ruxience (MEDICINE_NAME) Plus CVP (REGIME_THERAPY) completed (QUALIFIER.MODIFIER) in [**DATE**]

Relationship extracted is:

DURATION_OF_MEDICINE (Ruxience, 4 cycles)
MODIFIER_OF_MEDICINE_NAME (Ruxience, completed)
MODIFIER_OF_REGIME_THERAPY (CVP, completed)

Health Assertion Detection (HASD)

The aim of Health Assertion detection is to identify assertion types for medical entity types (as they appear as spans) in the clinical text, namely certainty (whether the medical concept is positive, negated, possible, or hypothetical), temporality (whether the medical concept is for present, the past, or the future history), subject (whether the medical concept is described for the physician, patient, a family member or other) and so on.

Examples:


SI	Text	Span with Entity Type	Modality/Dimension	Value/Qualifier
1	Prescribing sick days due to diagnosis of influenza	influenza (DISORDER)	Certainty	Certain
2	His kidneys are deteriorating	kidneys (BODY_STRUCTURE)	Course	Worsening
3	He has acute pain in left leg	pain in left leg (SIGN_SYMPTOM)	Severity	Severe

Health Medical Entity Linking (HMEL)

The aim of the task is to associate or link mentions (spans) of recognized entities to their corresponding node in a knowledge base or an ontology. In practice, entity linking is helpful for automatic linking of Electronic Health Records (EHR) to medical entities, supporting downstream tasks such as diagnosing, decision making and the like.

Example:

"Indication: Acute hypoxia, Relapsed AML, GVHD, and renal failure with new hypoxia with clear chest X-ray"


Span for the entity type ‘DISORDER’	ICD 10 CM code (Ontology)
Acute hypoxia	J96.01
Relapsed AML	C92.02
GVHD	D89.813
Renal failure with new hypoxia	N17.1

Supported Ontologies

Rxnorm: See https://www.nlm.nih.gov/research/umls/rxnorm/index.html.
SNOMED CT US: See .https://www.nlm.nih.gov/healthit/snomedct/us_edition.html
ICD 10 CM: See https://www.cdc.gov/nchs/icd/icd-10-cm/index.html.
Tumult: See https://www.drugs.com/mtm/.

Pipeline Architecture of Four Services

These healthcare NLP models are built in the framework OCI Healthcare services and deployed on OCI healthcare NLP endpoint using a pipeline architecture.

The following example shows text as input to Health NLP endpoint and the output produced for different modules.

Input text: pain in armpit; advised Aceclofenac twice a day for 3 days.

Specific input and outputs on Health NLP pipeline

Note

When working with the Oracle NLP model, it's important to review the provided confidence scores for accuracy. These scores can help you decide the appropriate confidence threshold for your particular use case. However, to ensure compliance with regulations, it's always advisable to verify the accuracy of any detected Health entities through other means such as human review.

Use Cases

Healthcare NLP models have a wide range of use cases in healthcare, revolutionizing the industry by improving patient care, streamlining operations, and facilitating research.

Clinical Documentation Improvement: NLP can help providers by extracting relevant information from patient records to provide recommendations for treatment options.
Clinical Decision Support: NLP can help providers by extracting relevant information from patient records to provide recommendations for treatment options.
Medical Coding: NLP can help automate the coding of medical procedures and diagnoses by analyzing physician notes.
Telemedicine: Develop voice-activated assistants that can transcribe doctor-patient interactions, update electronic health records, and provide quick access to relevant patient data during appointments.

Supported Entity Types


	Entity Type	Description
1	HEADER	Chief Complaint → HEADER Detect the main section headers within the document. Marking the HEADER is highly dependent on the document structure. Use the correct context to mark document sections as HEADER.
2	SUB_HEADER	All child headers of the main header. This entity type might include sub-headers or sub-sub header.
3	BODY_STRUCTURE	The organ names, organ sites, body parts, or body regions.
4	MORPHOLOGIC_ABNORMALITY	The abnormal anatomical body structure.
5	CELL	The cell types.
6	FINDING.SIGN_SYMPTOM	The signs or symptoms of the medical condition. Signs: Objective findings that can be observed by a healthcare provider. Symptoms: Subjective experiences that are reported by the patient.
7	FINDING.OTHER	The findings that aren't sign or symptom, are considered as FINDING.OTHER. Observations: The active acquisition of subjective or objective information from a primary source. This includes general findings of observation of the patient. This entity type can capture aspects such as: Personal characteristics: For example, eye-color. Social histories: Examples include: Substance abuse or use. For example: Smoker Alcoholic Drug use Cognitive status or psychological evaluations. For example: Alert Oriented x3 Awake Oriented Calm Pleasant Core characteristics: Examples include: Pregnancy status Death assertion Physical observations: Examples include: Soft Non-tender Well-developed Normal medical conditions: Examples include: Normal eyes Normal bowel sounds Normal heart rate EOMI PERRLA Clear lungs
8	DISORDER	The diseases and disorders. Always and necessarily abnormal. Necessarily have an underlying pathological process. Have temporal persistence (might be under treatment, in remission, or inactive, even though they're still present). Might be present as a propensity for certain abnormal states to occur, even when treatment mitigates or resolves those abnormal states.
9	STAGING_SCALE	Chest pain rating, Breathlessness rating Symptom rating, ...
10	ASSESSMENT_SCALE	Pain Scale Visual analog pain scale Pain Descriptor Scale Karnofsky score Token test Dolo test Borg scale ...
11	TUMOR_STAGING	M+ tumor staging N+ tumor staging H+ tumor staging Level II tumor staging Lung stage L2 ...
12	MEDICATION_ORDER	The sentences or segments of the EHR document that contains medication order-related entities in it.
13	MEDICINE_NAME	The generic name of the drug.
14	MEDICINE_FREQUENCY	The frequency for medication. For example: `Two times a day, daily, q4h`
15	MEDICINE_DOSE	All words mentioning the medication dosage.
16	MEDICINE_DOSE.FORM	The only form of dose.
17	MEDICINE_ROUTE	The route of administration.
18	MEDICINE_DURATION	The duration of the medicine.
19	MEDICINE_STRENGTH	The strength of the medicine.
20	MEDICINE_DISPENSE	The total dispense units of medicine.
21	MEDICINE_PRN_ASNEEDED	The PRN prescription stands for 'pro re nata,' which means that the administration of medication isn't scheduled. Instead, the prescription is taken as needed.
22	MEDICINE_REFILL_AMOUNT	The number of times to refill a medication.
23	MEDICATION_CLASS	The collective names for groups of medications. Drugs can be classified in different ways according to: Mode of action, for example, visit type Indications Chemical structure
24	OBSERVABLE_ENTITY.VITALS	Vitals: Examples include: Blood pressure Body temperature Heart rate Respiratory rate Body measurements: Examples include: Height Weight Body Mass Index Head circumference Pulse oximetry
25	OBSERVABLE_ENTITY.OTHER	The observable entity is the name of something that can be observed and represents a question or assessment that produces an answer or result. Functions carried out by the body or organ. This excludes VITALS.
26	PROCEDURE.LAB_TEST	The laboratory tests are performed on a sample of blood, urine, or other substance from the body.
27	PROCEDURE.OTHER	The procedure is a one-time action performed on the patient to treat a medical condition or to provide patient care.
28	REGIME_THERAPY	The treatment is interventions performed over a period of time (days, weeks, months) to treat a disease or disorder.
29	MEASUREMENT	The measurements related to lab, procedure, treatment, vitals, Observalbe_entities, and so on. It includes Measurement value (Numerical) and unit.
30	ALLERGEN_AGENT	The drug and food allergies.
31	IMMUNIZATION	The vaccine names, including: Hepatitis A Vaccine, Covid Shot, Flu shot, MMR, Tetanus, polio, varicella, pneumococcal, small pox, Hepatitis B, Hip, mums, Rubella, IPV, Influenza A, Influenza B, Rabies, OPV, Hepatitis B B19.10, flu, meningococcal ACWY, Tdap, Influenza B +, Influenza A J10.1, Measles, DT, meningococcal ACWY, and so on.
32	OCCUPATION.MEDICAL_ROLE	The specific medical occupations/professions are considered under this category. Examples include: Doctor Nurse Pharmacist
33	OCCUPATION.OTHER	The other non medical occupations / professions
34	PERSON.FAMILY	The person that the information is maintained for. Examples include: Employee Person Patient Care Professional Relative of a patient
35	PERSON.OTHER	The other persons that might not be a family or relatives.
36	SUBSTANCE	The concepts that can be used for recording and modeling: Chemical constituents of medicinal and non-medicinal products including: Allergies Adverse reactions Poisoning Physicians and nursing orders laboratory reports and results Sub-hierarchies of SUBSTANCE. Examples include: Body substance (substance) Chemical (substance)
37	EVENT	The situation around the individual at a specific time, which is relevant to their healthcare. Occurrences impacting health or health care, not including procedures or interventions.
38	PHYSICAL_OBJECT.MEDICAL_DEVICE	The physical devices relevant to health care, or to injuries/accidents.
39	RECORD_ARTIFACT.DOCUMENT_TYPE	The item/document/note component of the request. The clinical documents, or parts. Record artifacts don't have to be complete reports or records. They can be parts of a larger record artifact.
40	RECORD_ARTIFACT.OTHER	The subsections of the documents.
41	SPECIALTY	The related to departments.
42	ENVIRONMENT.CARE	The environment or location where patients are given care. Examples include: Emergency room Physicians office Cardio unit Hospice Hospital Location of person, pharmacy, any specialty wards, any generic location.
43	INDEPENDENT_HISTORIAN	The individual (for example, parent, guardian, surrogate, spouse, witness) who provides a history in addition to a history provided by the patient who is unable to provide a complete or reliable history (for example, because of developmental stage, dementia, or psychosis) or because a confirmatory history is judged to be necessary. When there's conflict or poor communication between several historians and more than one historian is needed, the independent historian requirement is met. The independent history doesn't need be obtained in person but does need to be obtained directly from the historian providing the independent information.
44	SITUATION	The phrases that must be recorded in the patient record but change the default context. Concepts that include context information, such as a subtype of the situation to which it applies with an attribute associating it with the relevant clinical finding or procedure Might be used to represent conditions/procedures that already occurred, haven't yet occurred, or refer to someone else (not patients)
45	ORGANISM	The organisms of significance to human and animal medicine used in modeling cause of disease.
46	SPECIMEN	The entities that are obtained (usually from patients) for examination or analysis.
47	QUALIFIER.MODIFIER	The qualifiers are the words or phrases that add details to the term. We annotate only words related to the following potential categories as qualifiers. Severity: The severity level is a measure of the intensity. Chronicity: A measure of persistence; a state of continuing to exist. Course of a Medical Condition: A fixed or ordered series of actions or events. Other Generic modifiers: Normal is a modifier in the Normal Lungs. Result: A qualitative entity (non-numeric) representing the result of a lab test, treatment, procedure, vitals, or observable entities. Includes values such as: Negative Positive Normal Low Elevated Directions: Includes values such as: Left Right Top Bottom Lateral Peripheral