Re-purposing The Electronic Medical Record For Public Health Sylvain DeLisle MD, MBA VA Maryland Health Care System and University of Maryland Parade to Promote Sale of War Bonds, Philadelphia, September 28, 1918 Overall Objective • To find out if a comprehensive EMR can contribute to the early detection of an infectious disease epidemic Case Detector EMR Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Case Detector CPRS Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS CPRS: Provider Interface CPRS: Free-text data entry CPRS: Structured data entry CPRS: Structured data entry CPRS is really VISTA Data Extraction: “MDE” SQL: Data Transformation Sequences SQL: Primary Warehouse Case Detector CPRS Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Case Detector SQL Database VISTA /CPRS Data Extractor Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Case Detector SQL Database VISTA /CPRS Data Extractor Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Focus on ILI • We focused on influenza-like illness (ILI) as a syndrome that may indicate an event of public health significance – Anthrax – Plague – SARS – Influenza ILI Case Definition • Positive influenza culture or antigen OR • Any two of the following (<= 7 days duration) – Cough – Fever or chills or night sweats – Pleuritic chest pain – Myalgia – Sore throat – Headache AND • Illness not attributable to a non-infectious etiology Gold Standard Detector: Manual Case Review • VA Maryland Health Care System (VAMHCS) and the Salt Lake City VA (SLCVAMC) • Study period: 10/01/03 to 3/31/04 • 15,377 (of 253,818) random sample, ER and selected outpatient clinics • All ILI cases and a 10% subsample of the records were rereviewed by a MD, discordant pairs were adjucated by a panel of three MDs • Found 280 cases ICD-9-based ILI Detectors • Compared “respiratory” ICD-9 groupings from – BioSense (CDC) – Essence (DoD) – Optimized (VA) Structured Parameters-based Case Detectors • Vitals: Temp >38, RR > 22, HR > 100 • Orders/dispense for Rx: expectorants, antibiotics, antitussives, decongestants, anti-emetics, antidiarrheals • Order/results for tests: CBC, Diff, Strep. screen, Sputum cultures, Gram stain, Respiratory serologies, Influenza cultures/antigens, Chest/sinus XRays or CT scans ILI Case Detector Retained parameters • “Cold remedies” • Fever >= 38ºC “Cold Remedies” CN101 opioid analgesics like ”codeine" OR CN900 CNS medications, other acetaminophen/diphenhidramine, OR MS102 non-salicylate NSAIS (does not include antirheumatics) , OR NT100 decongestants, nasal, OR NT200 anti-inflammatories, nasal, OR NT400 antihistamine, nasal, OR NT900 nasal and throat, topical, use “other throat lozenge" only, OR RE200 decongestants, systemic, OR RE301 opioid-containing antitussives/expectorants, OR RE302 non-opioid-containing antitussives/expectorant, OR RE501 antihist/decongest, OR RE502 antihist/decongest/antitussive , OR RE503 antihist/decongest/expectorant , OR RE507 antihist/antitussive, OR RE508 antihist/antitussive/expectorant , OR RE513 decongest/antitussive/expectorant , OR RE516 decongest/expectorant , OR RE599 cold remedies, OR AH102 antihistamines, ethanolamine, OR AH104 antihistamines, alkylamine TEXT-based ILI Case Detectors • Examine the text of all clinical encounter notes on the day of an index visit • Used modified NegEx algorithm (Wendy Chapman) TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc Negation? TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc Negation? TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc Negation? TXT ILI Case Detectors NegEx Mumbo jumbo fever nonsense trivia blabla etc Negation? Which One Should We Use? Case Detector SQL Database VISTA /CPRS Data Extractor Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Time Series: VAMHCS, Jan 1999 – Dec 2004 80 Number of ILI Cases 70 60 50 40 30 20 10 0 1999 2000 2001 2002 2003 2004 ILI Cases VAMHCS 2002-2003 70 Number of ILI Cases 60 50 40 30 20 10 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Case Detector SQL Database VISTA /CPRS Data Extractor Outbreak Generator Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS Outbreak Generator for Baltimore 1. Age-structured deterministic epidemic model generates the number of new infections in each age group, each day, for each zip code 2. Stochastic metapopulation spatial model determines how the outbreak will extend in space-time 3. Stochastic clinical features algorithm determines which of these infections will be recognized cases at the VA, and the severity, clinical profile and outcome for each recognized case ILI Cases VAMHCS 2002-2003 70 Number of ILI Cases 60 50 40 30 20 10 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun ILI Cases VAMHCS 2002-2003 70 Number of ILI Cases 60 50 40 30 20 10 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun ILI Cases VAMHCS 2002-2003 70 Number of ILI Cases 60 50 40 30 20 10 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Case Detector SQL Database VISTA /CPRS Data Extractor Outbreak Generator Outbreak Detector N Case Valid? Y N Outbreak Valid? Y Stop N Outbreak Y Sign.? Escalate PHS ILI Cases VAMHCS 2002-2003 70 Number of ILI Cases 60 50 40 30 20 10 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun BIOSENSE (Current) Fixed Threshold (p-value) Optimized Threshold Improved ICD-9 Codesets Adding Structured Parameters Adding Text Analysis Optimizing for Positive Predictive Value Surveillance for Febrile_ILI Conclusions (1) • EMR data can significantly enhance automated Case detection of ILI compared to the use of ICD9 codes alone • Whole-system simulation and is required to evaluate and calibrate the performance of alternative single-case detectors Conclusions (2) • For influenza surveillance, case-detection algorithms should aim for high positive predictive value, and target ILI cases who are febrile Bob Sawyer VISTA /CPRS Data Extractor Jill Anthony Shawn Loftus Brett South Shobha Phansalkar Brett South Matt Samore Outbreak Generator Gary Smith Holly Gaff Ericka Kalp Case Detector SQL Database Outbreak Detector Hongzhang Zheng Case Zhilian Ma Fang TianN Valid? Paul Sun Y “Vibrio” N Outbreak Valid? Sylvain DeLisle Trish Perl Y Stop N Outbreak Y Sign.? Steve Altman, Raju Vatsaval Escalate PHS