1. Applying Natural Language Processing in the Clinical Setting Peter Haug, MD Homer Warner Center for Informatics Research Intermountain Healthcare, Salt Lake City, Utah

2. Goals: –Affect Care Delivery Extract Clinical Data from Medical Documents Use Extracted Data to Alter Care –Improve Documentation Identify Necessary Data Identify Eligible Patients –Support Clinical Research Data Extraction Phenotype Recognition Research Alerting –Improve Administrative Data Improve Data for Business Planning Improve Data for Billing Examples: –Support for Diagnostic Systems Screening for Disease Assess Risk Triggering Orders Activating Clinical Protocols –Encode Admit Diagnoses –Identify Patients for Trauma Registry –Complete Problem List –Identify Patients for Research Recruitment Applied NLP Research 2

3. Process Begins with a Document Chest Xray Report History: Cough and fever. Previous history of right-sided pneumonia. Exam: PA and Lateral Chest Film. Observations: Prior films showed confluent opacification of the RLL. This finding remains in today’s exam. These opacities, seen in multiple previous films, have spread to the right and left upper lobes. Interpretation: Extension of previously diagnosed pneumonia to right and left upper lobes. Find Document StructureFind Document Structure Extract Meta- DataExtract Meta- Data Plan Further ParsingPlan Further Parsing Find Sentence StructureFind Sentence Structure Determine Meaning (Concepts)Determine Meaning (Concepts) Store/Process ConceptsStore/Process Concepts Document Parsing Process

4. Find Document Structure Find Section Structure Find Sentence Structure Determine Meaning (Semantics) –Map to Concepts –Build Data Structures A Sequence of Processing

5. Document Parsing Process Find Document StructureFind Document Structure Extract Meta- DataExtract Meta- Data Plan Further ParsingPlan Further Parsing Find Sentence StructureFind Sentence Structure Determine Meaning (Concepts)Determine Meaning (Concepts) Store/Process ConceptsStore/Process Concepts Prior films showed confluent opacification of the RLL. This finding remains in today’s exam. These opacities, seen in multiple previous films, have spread to the right and left upper lobes. These opacities, seen in multiple previous films, have spread to the right and left upper lobes.

6. Find StructuresFind Structures Extract Meta- DataExtract Meta- Data Plan Further ParsingPlan Further Parsing Find Sentence StructureFind Sentence Structure Determine Meaning (Concepts)Determine Meaning (Concepts) Store/Process ConceptsStore/Process Concepts Sen> These opacities, seen in multiple previous films, have spread to the right and left upper lobes. 1234 1235 Coded Findings: Localized Infiltrate-RUL Localized Infiltrate-LUL Document Parsing Process

7. Categorize Words Identify Phrasal Boundaries Find Relationships Among Phrases Test for Semantic Congruence Extract Concepts These opacities, seen in multiple views, have spread to the right and left upper lobes. Sentence Parsing Process these=>POS: Adj. Semantic Rep: N/A opacities=>POS: Noun Num; Pleural Semantic Rep: RadFind.finding.opacities ID: 1 seen=>POS: verb Num; N/A Semantic Rep: N/A ………………………... right=>POS: adj. POS: Noun Num; Singular Semantic Rep: RadFind.side.right ID: 2 left=>POS: adj. POS: Noun Num; Singular Semantic Rep: RadFind.side.left ID: 3 upper=>POS: adj. Num; N/A Semantic Rep: RadFind.sup_inf.upper ID: 4 lobes=>POS: noun Num; pleural Semantic Rep: RadFind.anat_loc.lobes ID: 5 Etc. these opacities to the left upper lobes

8. Categorize WordsCategorize Words Identify Phrasal BoundariesIdentify Phrasal Boundaries Find Relationships Among PhrasesFind Relationships Among Phrases Test for Semantic CongruenceTest for Semantic Congruence Extract ConceptsExtract Concepts These opacities, seen in multiple views, have spread to the right and left upper lobes. Document Parsing Process From Words to Coded Findings: Localized Infiltrate-RUL Localized Infiltrate-LUL these opacities to the left upper lobes

9. Output of a Semantic Parse Instantiated Event: 1001 *Overall Concept : *localized infiltrate (0.998669) 1002 *State Concept : *present (0.780993) 1003 Presence Term : null (0.779583) 1004 *Topic Concept : *poorly-marginated opacity (infiltrate) (1.0) 1005 Topic Term : opacity~n (1.0) 1006 Topic Modifier Term: hazy~adj. (1.0) 1008 Topographic Location Term : null (0.588844) 1009 *Severity Concept : *null (0.969009) 1010 Severity Term : null (0.962739) 1011 *Link Concept : involving (0.686011) 1012 Topic Location Link Term : in (1.0) 1013 *Anatomic Concept : *right upper lobe (1.0) 1014 Anatomic Location Mod : null (0.9375) 1015 Anatomic Location : lobe~n (1.0) 1016 Anatomic Location Mod1 : right (1.0) 1017 Anatomic Location Mod2 : upper (1.0) 1018 Anatomic Location Mod3 : null (1.0) 1019 Anatomic Location Mod4 : null (1.0) 1020 Anatomic Location Mod5 : null (1.0) Instantiated Event: 1001 *Overall Concept : *localized infiltrate (0.998669) 1002 *State Concept : *present (0.780993) 1003 Presence Term : null (0.779583) 1004 *Topic Concept : *poorly-marginated opacity (infiltrate) (1.0) 1005 Topic Term : opacity~n (1.0) 1006 Topic Modifier Term: hazy~adj. (1.0) 1008 Topographic Location Term : null (0.588844) 1009 *Severity Concept : *null (0.969009) 1010 Severity Term : null (0.962739) 1011 *Link Concept : involving (0.686011) 1012 Topic Location Link Term : in (1.0) 1013 *Anatomic Concept : *right upper lobe (1.0) 1014 Anatomic Location Mod : null (0.9375) 1015 Anatomic Location : lobe~n (1.0) 1016 Anatomic Location Mod1 : right (1.0) 1017 Anatomic Location Mod2 : upper (1.0) 1018 Anatomic Location Mod3 : null (1.0) 1019 Anatomic Location Mod4 : null (1.0) 1020 Anatomic Location Mod5 : null (1.0) PARSE: A hazy opacity is seen in the right upper lobe.

10. Probabilistic Semantics

11. Goal: –Identify Pneumonia Patients in the ED Rapidly –Assess Risk –Suggest Intervention Approach: –Use Probabilistic System to Identify Patients –Suggest Enrollment in Pneumonia Protocol –Provide Therapeutic Suggestions Requires Data Extracted from the X-ray Report Example: NLP in Pneumonia (a computer-based intervention)

12. Care Delivery Framework

13. Example: Community-Acquired Pneumonia

14. Care Delivery Framework Example: Community-Acquired Pneumonia Does the patient have pneumonia?

15. Care Delivery Framework Example: Community-Acquired Pneumonia Does the patient have pneumonia? Should we used the guideline?

16. Care Delivery Framework Example: Community-Acquired Pneumonia Does the patient have pneumonia? Should we used the guideline? Apply Pneumonia Care Protocol.

17. Web Services Infrastructure A Bayesian Network Supported by a Production Rules System (DROOLS) Using an NLP System –Sentence Isolation –Random Forests-Based Semantics Implimented Using:

18. Patient Tracking Board

22. Natural Language Processing Does Play A Role In Patient Care Useful Applications Will Blend NLP-Derived Data With Structured Data From The EHR Radiology Reports Are A Data-rich Target For NLP Conclusion

23. 23 Questions???