iHT2 Health IT Summit in Seattle 2012 – Case Study “EMR Usability & NLP”


Published on

Case Study “EMR Usability & NLP”

Thomas Payne, MD, FACP
Medical Director, IT Services
UW Medicine

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

iHT2 Health IT Summit in Seattle 2012 – Case Study “EMR Usability & NLP”

  1. 1. Case Study EMR Usability & NLP Thomas H. Payne, MD, FACMI Medical Director, IT Services UW Medicine Clinical Associate Professor Departments of Medicine, Health Services, and Biomedical Informatics & Medical Education University of Washington IHT2, Seattle, 8/22/12 tpayne@u.washington.eduThursday, August 23, 12
  2. 2. Topics today • The story of moving from paper to electronic notes • NLP and clinical decision support • 3 examples: NLP in UW Medicine EMRs • Summary and discussion 2 tpayne@u.washington.eduThursday, August 23, 12
  3. 3. 5 reasons doctors have trouble with EMRs • Time requirement to enter notes and write orders • What notes look like and contain • Changed interaction with patients • Expense to buy and operate • Makes it harder to be a doctor 3 tpayne@u.washington.eduThursday, August 23, 12
  4. 4. 4,500 4,000 —Deaths —Billions of passengers-km 3,500 3,000 2,500 2,000 1,500 1,000 500 1918 1928 1938 1948 1958 1968 1978 1988 1998 2008 Source: Aircraft Crashes Record Office, Geneva (http://www.baaa-acro.com/) The Geography of Transportation Systems. http://people.hofstra.edu/geotrans/eng/ch3en/conc3en/airfatalities.html. tpayne@u.washington.eduThursday, August 23, 12
  5. 5. tpayne@u.washington.edu tpayne@u.washington.eduThursday, August 23, 12
  6. 6. Tools for structured note entry tpayne@u.washington.eduThursday, August 23, 12
  7. 7. tpayne@u.washington.eduThursday, August 23, 12
  8. 8. Problems with structured note entry • Training requirement higher • In our experience, most physicians don’t like to write with them. • Most physicians don’t like to read them (except narrative portion). • Important detail may be lost. tpayne@u.washington.eduThursday, August 23, 12
  9. 9. What hath we wrought? tpayne@u.washington.eduThursday, August 23, 12
  10. 10. More attractive Less attractive Payne TH, Patel R, Beahan S, Zehner J. The physical attractiveness of electronic physician notes. 10 tpayne@u.washington.edu AMIA Annu Symp Proc., 2010:622-626.Thursday, August 23, 12
  11. 11. tpayne@u.washington.eduThursday, August 23, 12
  12. 12. Why is narrative text valuable? • Narrative contains history, details of history and exam, and most importantly the thinking of the clinician. (This is a rare overlap between needs of reimbursement, clinical care, teaching, and research.) • Each note contains kernels of truth. Templates, direct entry aids, copy/paste can hide them. • In UW Medicine there are electronic notes from ~1.4 million visits and 68,000 admissions each year → great potential to improve decision-making and to learn. 12 tpayne@u.washington.eduThursday, August 23, 12
  13. 13. Hardware you already own Broca’s area auditory auditory association Graphic credit: Prof.Genevieve B. Orr http://www.willamette.edu/~gorr/classes/cs449/brain.html tpayne@u.washington.eduThursday, August 23, 12
  14. 14. Sound to text “45 year old female with pulmonary sarcoidosis…” 14 tpayne@u.washington.eduThursday, August 23, 12
  15. 15. Speech technologies are mainstream 15 tpayne@u.washington.eduThursday, August 23, 12
  16. 16. Med reconciliation accurate? tpayne@u.washington.eduThursday, August 23, 12
  17. 17. Early signs of sepsis? Other recommendations within imaging reports? tpayne@u.washington.eduThursday, August 23, 12
  18. 18. Broader differential? Appropriate care? Code status in EMR accurate? In compliance with billing rules? tpayne@u.washington.eduThursday, August 23, 12
  19. 19. Sound to meaning “45 year old female Parent/Child (Relationship Type) with pulmonary Sarcoidosis (disorder) {31541009 , sarcoidosis…” SNOMED-CT } Natural language processing A subfield of artificial intelligence and computational linguistics. It studies the problems of automated generation and understanding of natural human languages 19 tpayne@u.washington.eduThursday, August 23, 12
  20. 20. Analysis of UHC Core Measures data within UW Medicine All 3 Current availability of data elements HF AMI OPPS SCIP measures Structured, codified, machine-readable format 3 (17%) 6 (15%) 1 (11%) 10 (15%) Free text from canonical single source 9 (50%) 11 (28%) 6 (67%) 26 (39%) Free text from canonical multiple sources 1 (6%) 1 (3%) 0 (0%) 2 (3%) 85% Handwritten documentation, human 5 (28%) 22 (55%) 2 (22%) 29 (43%) interference required Analysis of 67 metrics reported to University Healthcare Consortium for heart failure (HF), acute myocardial infarction (AMI), and outpatient surgical care improvement project (OPPS SCIP) metrics. Together, these measures include 67 distinct metrics that require reporting to UHC. Supporting data in source systems (ORCA EMR, echocardiology, Docusys OR system, Reg/ADT, and other systems) for reporting these 67 metrics were analyzed. Slide courtesy of Tony Black, Biomedical Informatics Core, Institute of Translational Health Sciences, University of Washington tpayne@u.washington.eduThursday, August 23, 12
  21. 21. QA Tools: Required Clinical Variables by Anticipated Accessibility Accessible n Percentage Hard to Access n Percentage Demographics 482 100 Disease-specific history 142 30 Diagnosis 346 72 Care site 133 28 Prescription 167 35 Physical exam 67 14 Past medical history 148 31 Refusal 60 13 Procedure date 117 24 Patient education 53 11 Lab date 81 17 Social history 52 11 Problem/chief complaint 57 12 Treatment 25 5 Vital sign/weight/height 46 10 Diagnostic test result 18 4 Allergy 42 9 Imaging result 17 4 Lab result 38 8 Contraindication 15 3 Medication history 28 6 Pathology 11 2 Diagnostic test date 22 5 Family history 11 2 Imaging date 22 5 ECG result 6 1 Medications, current 13 3 X-ray result 2 <1 Vaccination 9 2 X-ray date 6 1 EKG date 6 1 American Journal of Medical Quality, Vol. 24, No. 5, Sep/Oct 2009 tpayne@u.washington.eduThursday, August 23, 12
  22. 22. tpayne@u.washington.eduThursday, August 23, 12
  23. 23. ness in gathering the patient’s one virtue of com history, findings from the physi- tems is that they cal examination, and other data. corded informatio Can Electronic Clinical Documentation Help Prevent Diagnostic Errors? Because information from pa- formats. Designer Gordon D. Schiff, M.D., and David W. Bates, M.D. tients’ previous clinical encoun- leverage the “visu Tinvest nearly $50technology testsquality without adversely discouraging independent data he United States is about to many questions about it persist. ing physicians from the patient, tersbillion in For example, can it bebe more read- assessment, and health information and improve will leveraged to gathering and capabilities of EH country to a tipping availablethe quality of electronic notes vision a redesigned documenta- ily point with Will with electronic than the aggregation, 4 (HIT) in an attempt to push the affecting clinicians’ efficiency? perpetuating errors. But we en- puterized records, which are records, degraded by the wide- to elec- paper ex- or will it be shifting approaches to improving diagno- patient’s weight o respect to the adoption of com- be better than that of paper notes, tion function that anticipates new pected to improve the quality and spread use of templates and cop- sis, not one that relies on the pu- tronic systems could substantial-diagnosticians” of for instance) 1 reduce the costs of care. A fun- ied-and-pasted information? tative “master tion, damental question is how best A fundamental part of deliv- past eras. The diagnostic process to design electronicimprove clinicians’ is get- must be made reliable,emphasis or disp ly health rec- eringthe diagnosis right. Unfor- and electronic documentation will good medical care knowledge not heroic, ords (EHRs) to enhance clinicians’ ting workflow and the quality of care. tunately, diagnostic errors areproblem about thecommon, outnumbering medica- be key toand clinicians will to facilitate rap patient. The velopers this effort. Systemsneed as de- Although clinical documentation of having too much information The second way plays a central role in EHRs and tion and surgical errors as causes to reconceptualize documentation occupies a substantial proportion of outpatient malpractice claims workflow as part of the next gen- tion practices have now surpassingbut we believe ers will need to lead by adopting is largely been multiple benefits, that of having can foster thought 3 of physicians’ time, documenta- and settlements. EHRs promise eration of EHRs, and policymak- dictated by billing and legal re- that one key selling point is their a more rational approach than quirements. Yet the primary role potential for preventing, become in- in which billingserving as a p too little, and it will mini- the current one, is by of documentation should be to mizing, or mitigating diagnostic codes dictate evaluation and man- clearly describe creasingly difficultevidencereview and providers are forced and communicate errors. Admittedly, to to agement all nicians, together what is going on with the patient. support the existence of such a to focus on ticking boxes rather Electronic prescribing patient information than on thoughtfully document succinc themedication benefit is currently lacking, and ing their clinical thinking. appears that is document- to reduce the rate of our hypothesis runs counter to errors, but the other benefits of the sentiments and claims of There are numerous ways in electronic records are less clear.2 many physicians, who argue that which EHRs can diminish diag- We must ensure that electronic electronic documentation in its nostic errors (see table). The first clinical documentation works ef- current incarnation is time-con- lies in filtering, organizing, and fectively to improve care if more suming and can degrade diag- providing access to information. benefits are to be achieved. Yet nostic thinking — by distract- Making accurate diagnoses has N ENGL J MED 2010; 362:1066-9 tpayne@u.washington.eduThursday, August 23, 12
  24. 24. CT CHEST VENOUS PROTOCOL CT ABDOMEN / PELVIS INDICATION Motorcycle crash. Comparison: Trauma series same date and outside noncontrast CT chest abdomen and pelvis same date. PROCEDURE: Reconstruction thickness: 2.5mm. Interval: 2.5 mm Superior extent: Thoracic inlet. Inferior extent: Ischial tuberosities Intravenous contrast: Positive. Oral contrast: Not used. MPRs: Coronal, Chest, Abdomen and Pelvis, venous. FINDINGS: **** CHEST **** Aorta and great vessels: Normal for a venous phase study. Mediastinal hematoma: Absent Pericardial fluid: Absent Endotracheal tube: Absent Intercostal tubes: Absent Right lung: Mild dependent upper lobe consolidation likely represents atelectasis. Left lung: Partial collapse of the left upper lobe. Mild basal atelectasis is also present. IMPRESSION: Right pneumothorax: Absent Left pneumothorax: Moderate left pneumothorax, increased compared to the prior outside study. Multiple left rib fractures and left clavicular fracture. Right pleural fluid: Absent Moderate of thefluid: Absent left rib fracturesleft present:is partially collapsed. Left pleural pneumothorax. The Bones left thorax: Multiple are lung Right sacral fracture and 4left7 pelvic ring disruption and 12. There is a minimally displaced mid shaft fracture of the left clavicle. Please refer to dedicated CT pelvis for details. Segmental 2nd, anterolateral to minimally displaced, posterior 11 with associated small extraperitoneal hematoma. ThereNo right sidedof the Thoracic spine for details. right kidney, which may be secondary to a 4-mm renal calculus at the level of the distal ureter. Delayed abdominal chest fractures identified. is mild hydronephrosis of the Spine: See CT radiograph is recommended. **** ABDOMEN AND PELVIS **** Liver: Normal Small amount of fluid surrounding the gallbladder, and nonspecific finding of uncertain clinical significance. Spleen: Normal Comment: Normal ducts: Normal. A small amount of fluid is noted surrounding the gallbladder, of uncertain significance. Gallbladder and bile Pancreas: Small left hemothorax. Portal veins: Normal Abdominal aorta and branch vessels: Normal Left apical extrapleural is mild right hydronephrosis and ureteral dilatation. At the level of the pelvis (2/215), there is a radiodense structure, which could Kidneys and ureters there hematoma associated with rib fractures. The above described fluid next to4 the gallbladder may simply represent the gallbladder wall. No definite fluid around the gallbladder or liver is represent a renal calculus which measures mm. Adrenal glands: Normal identified. duodenum and small bowel: Normal. Stomach, BetterColon: Normal on and retroperitoneum: Normal right pelvic calcification described above, which is external to the ureter and represents a phlebolith. No visualized Mesentery, omentum CT cystogram, is the evidence of ureteral calculi. Appendix:Not visualized Right adnexalIVC:Normal Aorta and Lymph Nodes: Normal cyst. Recommend nonemergent pelvic ultrasound for further evaluation to exclude cystic ovarian neoplasm. Small hiatal Normal Bladder: hernia. Uterus and ovaries: There is a right hypodense and well demarcated adnexal mass, which measures 2.8 cm in diameter. High-densitypelvic ultrasound. Consider fluid is present posterior and adjacent to the inferior aspect of the descending colon. This is contiguous with the pelvic retroperitoneal hematoma,Right sacral fracture A small amountdisruption is again visualized. noted. refer to dedicated bony pelvis CT for details. extension ofextravasation. Bones: and given the absence of abnormalities ofPlease colon most likely represents There is left parasymphyseal and peri-acetabular hematoma. and pelvic ring of presacral hematoma also the No definite intraperitoneal fluid is identified. No obvious active the retroperitoneal hematoma, which is not increased inintraperitoneal fluid or prior study. No free size since the air present. Subcutaneous tissues and body wall: There is left gluteal subcutaneous hematoma. IMPRESSION: Multiple left rib fractures and left clavicular fracture. Moderate left pneumothorax. The left lung is partially collapsed. Right sacral fracture and left pelvic ring disruption with associated small extraperitoneal hematoma. Please refer to dedicated CT pelvis for details. There is mild hydronephrosis of the right kidney, which may be secondary to a 4-mm renal calculus at the level of the distal ureter. Delayed abdominal radiograph is recommended. Small amount of fluid surrounding the gallbladder, and nonspecific finding of uncertain clinical significance. Comment: Small left hemothorax. Left apical extrapleural hematoma associated with rib fractures. The above described fluid next to the gallbladder may simply represent the gallbladder wall. No definite fluid around the gallbladder or liver is identified. Better visualized on CT cystogram, is the right pelvic calcification described above, which is external to the ureter and represents a phlebolith. No evidence of ureteral calculi. Right adnexal cyst. Recommend nonemergent pelvic ultrasound for further evaluation to exclude cystic ovarian neoplasm. Small hiatal hernia. High-density fluid is present posterior and adjacent to the inferior aspect of the descending colon. This is contiguous with the pelvic retroperitoneal hematoma, and given the absence of abnormalities of the colon most likely represents extension of the retroperitoneal hematoma, which is not increased in size since the prior study. Discussed with admitting surgical service at 8:30 p.m. 24 tpayne@u.washington.eduThursday, August 23, 12
  25. 25. Show term detail I clicked on AML. This window shows that ‘AML’ was recognized as a monocytic leukemia, SNOMED code 100744006. tpayne@u.washington.eduThursday, August 23, 12
  26. 26. NLP can help clinical decision support • Summarization • Enhanced search • Extracting key encoded information from narrative, such as problem lists but potentially far more • Focus our attention on needs that might be overlooked • As narrative text grows, so does the need for NLP 26 tpayne@u.washington.eduThursday, August 23, 12
  27. 27. July 2009 tpayne@u.washington.edu ICN: 006764Thursday, August 23, 12
  28. 28. NLP tagging, then CMS E/M rules applied E/M assigned by Code estimated by nCode MD Phrase recognized and assigned SNOMED code Narrative text—dictation, directly entered, or combination—is run through engine to tag SNOMED phrases. CMS algorithm used to assign E/M code suported by the note. 11/14/09 tpayne@u.washington.edu 28Thursday, August 23, 12
  29. 29. NLP improves E/M Financial Management accuracy and RVUs CODE OF CONDUCT Your coding questions answered Natural language processing • After calibration, 94% concordance improves coding accuracy between human coders and system Process enables group to provide rapid feedback to physicians and identify problems • Feedback within 3 seconds on each M any physicians find rules for assigning accurate E&M codes within three seconds and remained accu- rate, according to an ongoing audit. note written challenging1 and ask for feedback. Since coding clinical notes is an essential compo- nent of billing for professional medical During our test, NLP increased coding accuracy, provided rapid, regular feedback to physicians on their performance, and services and is required for payer compli- identified problems that might other- • With improved compliance comes ance, we used natural language processing wise have been difficult to identify. For (NLP) to assign E&M codes to electronic example, if a note was complete except clinical notes and to provide coding feed- for the absence of the chief complaint, revenue back. NLP is a subfield of computer science the level supported by the note would be that studies problems of automated genera- lower and the reason for this would be in tion and understanding of natural human the feedback screen. We believe this tool languages. will increase compliance with billing rules, Over the last two years, the University assist in practitioner education and reduce of Washington (UW) Medicine tested NLP billing costs. By Thomas H. Payne, MD, (shown to assign E&M codes to electronic clinical It has been well-accepted by physicians here) medical director, Information notes for billing professional services. and compliance professionals, who are Technology Services, UW Medicine, Seattle, tpayne@u.washington. We compared accuracy with coding by accustomed to the nCode, an applica- edu; Aidan Garver-Hume, systems practitioners (mostly physicians) and tion used to apply NLP techniques to analyst and programmer, Information coding experts, and assessed practitioner professional fee billing. It also satisfied Technology Services, UW Medicine, Seattle, aiding@u.washington.edu; acceptance, impact on labor costs and physician requests for feedback on coding Sheila Kirkegaard, practice plan the educational mission of our physician performance. administrator, Fred Hutchinson Cancer practice plan. Research Center, Seattle, skirkega@ fhcrc.org; Jamie Sweeney, charge capture This work was conducted under the Billing workflow supervisor, UW Physicians, Seattle, auspices of UW Physicians, our group prac- jsweeney@uwp.washington.edu; tice plan in the outpatient and inpatient Prior to this project, physician and Michael Ash, MD, RPh, vice president of provider strategy, Cerner Corporation, Blood and Marrow Transplant Service of nonphysician providers (NPPs) who bill Kansas City, mash@cerner.com; K.K. Seattle Cancer Care Alliance (SCCA), a for services completed paper fee sheets Kailasam, director and distinguished cancer care hospital and clinic created by with codes for outpatient visits. Several engineer, Cerner Corporation, Kansas the UW, Seattle Children’s Hospital and Fred Hutchinson Cancer Research Center. tpayne@u.washington.edu applications were developed including an electronic fee sheet (eFeesheet) and an City, kkailasam@cerner.com; Candace L. Hall, solution manager, Cerner Corporation, Kansas City, chall@ The clinic serves patients considered for interactive screen that replaced the paper cerner.com; Mika N. Sinanan, MD, PhD, professor, Department of Surgery, hematopoietic cell transplantation and fee sheet. It appears after a physicianThursday, August 23, 12 those undergoing or recovering from the signs a note and permits rapid entry of an UW, president, UW Physicians, Seattle,
  30. 30. Summary • We need to better match EMRs with human strengths and workflow. • Much of the clinical note content clinicians create is in narrative. • That content can help us make better decisions, esp if aided by NLP. • NLP today fits into the workflow of EMRs to capture important content from narrative. 30 tpayne@u.washington.eduThursday, August 23, 12