Natural Language Processing for Public Health

Transcription

1 Natural Language Processing for Public Health Coding OLIS Microbiology Data Kevin Brown Scientist Public Health Ontario Assistant Professor University of Toronto Adjunct Scientist ICES

2 Ontario Lab Information System Currently Covers 2007 to Fall ,892 unique test result types 202 Ontario labs 1 billion test orders 2.2 billion test results

3

4

5 Text result field Comment:\.br\The presence of C. difficile DNA is not diagnostic of\.br\infection and should be interpreted in conjunction with\.br\clinical presentation.\.br\ \.br\ Duplicate specimens received within 7 days of a NEGATIVE\.br\test or within 14 days of a POSITIVE test will be rejected. C. difficile Cytotoxin (PCR) Positive by RT PCR\.br\ **Infection Control Additional Precautions - Contact**\.br\ \.br\ \.br\ Phoned to (416) \.br\ Date: Apr 3,2015 \.br\ At: 14:10\.br\ By: John Doe

6 COMBAT-AMR Comprehensive Ontario Microbiology LaBoratory Administrative data for AntiMicrobial Resistance The stated purpose was to: Identify incidence and prevalence of AMR Measure the attributable mortality of each AMR Combine into Clinical Antimicrobial Resistance Index

7 Objective Identify tests of interest Code test results Describe test results Validate

8 Identify ICES provided us with a table of LOINC codes Logical Observation Identifiers Names and Codes A description of a microbiology result field LOINC -- Type C = Culture Test S = Susceptibility Test OST = Organism Specific Test LOINC -- Priority 1 = definitely of interest. 6 = definitely not of interest

9 Identify LOINC loincfullyspecifiedname frequency 05-Dec AMIKACIN:SUSC:PT:ISOLATE:ORDQN:MIC T' wave amplitude.lead V3:Elpot:Pt:Heart:Qn:EKG Surgical operation note preoperative Dx:Imp:Pt:Patient:Nar Lymphocytes/100 leukocytes:nfr:pt:csf:qn:manual count Monocytes/100 leukocytes:nfr:pt:csf:qn:manual count Monocytes/100 leukocytes:nfr:pt:body fld:qn:manual count CANCER AG 125:ACNC:PT:SER/PLAS:QN Hemoglobin A:ACnc:Pt:Bld:Qn:Electrophoresis Bacteria identified:prid:pt:gen:nom:aerobic culture Bacteria identified:prid:pt:nose:nom:aerobic culture MICROSCOPIC OBSERVATION:PRID:PT:BONE MAR:NOM:WRIGHT GIEMSA STAIN MICROSCOPIC OBSERVATION:PRID:PT:WND:NOM:GRAM STAIN COMPLEMENT C3 AG:ACNC:PT:TISS:ORD:IMMUNE STAIN ANION GAP 3:SCNC:PT:SER/PLAS:QN IGA AG:ACNC:PT:TISS:ORD:IMMUNE STAIN IGG AG:ACNC:PT:TISS:ORD:IMMUNE STAIN IGM AG:ACNC:PT:TISS:ORD:IMMUNE STAIN LUTROPIN:ACNC:PT:SER/PLAS:QN Microscopic observation:prid:pt:cvx:nom:cyto stain MICROSCOPIC OBSERVATION:PRID:PT:XXX:NOM:CYTO STAIN

10 Identify LOINC loincfullyspecifiedname frequency priority testype BACTERIA IDENTIFIED:PRID:PT:XXX:NOM:AEROBIC CULTURE C BACTERIA IDENTIFIED:PRID:PT:XXX:NOM:CULTURE C BACTERIA IDENTIFIED:PRID:PT:ISOLATE:NOM:CULTURE C BACTERIA IDENTIFIED:PRID:PT:THRT:NOM:CULTURE C BACTERIA IDENTIFIED:PRID:PT:URINE:NOM:CULTURE C BACTERIA IDENTIFIED:PRID:PT:BLD:NOM:CULTURE C BACTERIA IDENTIFIED:PRID:PT:STOOL:NOM:CULTURE C Bacteria identified:prid:pt:bld:nom:aerobic culture C TRIMETHOPRIM+SULFAMETHOXAZOLE:SUSC:PT:ISOLATE:ORDQN S NITROFURANTOIN:SUSC:PT:ISOLATE:ORDQN S GENTAMICIN:SUSC:PT:ISOLATE:ORDQN S CIPROFLOXACIN:SUSC:PT:ISOLATE:ORDQN S AMPICILLIN:SUSC:PT:ISOLATE:ORDQN S CEPHALOTHIN:SUSC:PT:ISOLATE:ORDQN S CEFAZOLIN:SUSC:PT:ISOLATE:ORDQN S

11 Identify Culture Tests N = 40 Priority 1 LOINCs Susceptibility Tests N = 197 Priority 1 LOINCs

12 Describe Culture Tests 4,552,482 test result records in ,312 unique values 1 every 70 records are unique Susceptibility Tests 3,823,864 test result records in ,217 unique values 1 out of 1700 records are unique Complexity Culture tests are 25X more complex

13 Coding Culture Tests List of 70 organisms (e. coli, staph aureus, etc) Multiple organisms Not classified Susceptibility Tests Susceptible Intermediate Resistant Other (MIC values, etc) Not classified

14 Coding Rule-based NLP (using regular expressions) Traditional Machine Learning Deep learning

15 Regular Expressions Character groups Any character =. Class of characters = [] E.g. lowercase letters [a-z], numbers [0-9], etc. Repetition? (absent or once) * (absent or any number) + (at least once) Concatenation = implicit or () ab = (a)(b) Alternation (OR statement) = 15

16 Regular Expressions R: uses POSIX standard grep(regexp, textvector) returns position grepl(regexp, textvector) whether or not it exists grepl( e.coli, DESCRIPTION) SAS (within a data step): uses Perl standard PRXPARSE compiles the regular expression PRXMATCH finds the position of the pattern match PRXMATCH(PRXPARSE('/e.coli/'), textvariable) > 0 16

17 Describing C & S coding This is a first pass 1. C coding 2. S coding 3. S x antibiotic

18 Describe C Organism N % No growth 976, Escherichia coli 525, Neisseria gonorrhoeae 263, Enterococcus sp. 149, Staphyloccocus aureus 98, Klebsiella sp. 71, Enterococcus faecalis 63, Campylobacter sp. 52, Klebsiella pneumoniae 49, Shigella sp. 48, Yersinia sp. 47,199 1 * Note 2.54 million coded records ~ 55.7% coded

19 Describe S Result N % Susceptible 2,802, Resistant 707, Intermediate 187, Other 117, Not classified 9, Total 3,823,864

20 Describe S x Antibiotic LOINC Antibiotic S I R %R Total TMP-SMX Nitrofurantoin Ciprofloxacin Gentamicin Ampicillin Cephalothin Cefazolin Tobramycin Amoxiclav Ceftriaxone Cefixime

21 Cleaning OLIS microbiology data 2-pronged approach Quick and dirty approach Subset of data we suspect (apriori) will be clean Outpatient urine cultures Rule-based NLP Sustainable approach Comprehensive Deep learning Active learning

22 Conclusions Susceptibility tests are easy to code we can probably do with rule-based NLP Culture tests will be more difficult to code and will require more sophisticated NLP methods