understanding emr error control practices among gynecologic physicians
TRANSCRIPT
Understanding the EMR Error Control Practices
among Gynecologic Physicians
Ritu Kharea, Yuan Anb, Sandra Wolfa, Paul Nyirjesy,a Longjian Liua, Edgar Choua
aDrexel University College of Medicine
bDrexel University College of Information Science and Technology
Philadelphia, PA, USA
1
iConference 2013,
February 13 2013
Motivation: EMR Errors
EMR Errors
◦ incomplete, inaccurate, or inconsistent information entered in Electronic Medical Records (Brown & Patterson, 2001; Phillips & Gong, 2009)
Occur because
◦ unusable EMR interfaces situated within demanding clinical environment
◦ clinicians inadvertently make mistakes while documenting patient visits and diagnosis information
Are expensive …
◦ poor data quality
◦ unsafe quality of care
◦ physicians liable for medical malpractice (Classen, Pestotnik, Evans, Lloyd, & Burke, 1997; Fichman, Kohli, & Krishnan, 2011).
2
Controlling EMR Errors Develop computational error control algorithms (Redwood, Rajakumar, Hodson, & Coleman,
2011).
◦ alert physicians in real time, and minimize further medical errors
An “inside-out” approach
Step1: Understand the existing error control mechanisms
Step II: Design the algorithms according to the observed limitations, and
opportunities.
Existing EMRs offer limited error control functionality
◦ Clinicians resort to MANUAL techniques to review, detect, and
resolve the errors (Phillips & Gong, 2009).
We take the first step toward algorithm development, and
investigate the manual practices.
◦ assess abilities of physicians to detect EMR errors
◦ elucidate their strategies
◦ derive implications for algorithm design
3
The Error Simulation User Study
Outpatient Clinics
◦ clinicians document the patient
visit information into the EMRs in
an on-the-spot “narrative”
manner.
◦ documentation occurs under
extreme time constraints
◦ conducive to a variety of data
errors (George & Bernstein, 2009).
Gynecologic Field of Medicine
◦ Physicians responsible for
documenting information on
yeast infections, bacterial vaginitis,
menstrual cycle issues, pre-natal
and post-natal complaints, regular
gynecologic examination, etc.
Study Workflow
1. Fabricate gynecologic visit
scenarios and develop the
corresponding EMR patient visit
notes.
2. Purposefully introduce several
EMR errors into the notes.
3. Present experienced gynecologic
physicians with the flawed notes,
and ask them to
◦ Identify any data errors
◦ Reveal error detection and
resolution strategies
4
EMR Gynecologic Visit Notes
5
Reason for Visit
Ms ABC presents for a new patient visit
Chief Complaints
Yeast Infections
HPI
After having the PE and DVT had to go off birth control which she was on for endometriosis. Few months later, started to have yeast infections.
Initially, noticed external itching, soreness, redness. Her OB/GYN told she had yeast and put her on fluconazole to take as two dose treatment.
Sometimes it would work but not reliable. Treated herself monthly in this manner. Started to have more internal symptoms, as well as little
discharge. Started nystain-triamcinolone which helped the external itching. They then tried inserts x7 days , it helped again. Now finds that her
partner is getting sore at times. Finally went on fluconazole once a week, started it a few months ago. Has not been much better on this regimen.
Allergies
Benadryl: CAPS
Lovenox: SOLN
Coumadin: TABS
Latex Exam Gloves MISC
PMH:
Bipolar disorder
Current Meds
Diflucan 150 mg oral tablet: Qty0, R0, RPT
Terconazole 0.8% Vaginal Cream; Qty0; R0; RPT
Probiotic CAPS; ; Qty0; R0; RPT
Assessment
• Bacterial vaginosis
Tests
LABS ORDERED; yeast culture
Orders
Fluocinolone Acetonide 0.025% External Ointment; APPLY SPARINGLY TO AFFECTED AREA(S) TWICE DAILY; Qty1; R1; Rx.
Plan
Reviewed at length w/pt. Discussed vulvar hygiene sheet. Discussed pathophysiology of lichen simplex. Start fluocinolone ointment. Stop
fluconazole as she has no evidence of VVC today. RTC – 1 month.
Signature
Electronically Signed by XYZ, MM-DD-YY
Type: Amb New Pt Visit
Owner: XYZ, XYZ
Status: Final
ABC, ABC, 34 yr old F, DOB: DD-MM-YY
Providers find it far more efficient to enter impromptu
narrative notes (as opposed to structured data) during
patient visits (Doğan et al., 2010).
A typical gynecoloigic
visit note is organized
into 19 Sections
(1) Reason for Visit
(2) Chief Complaint
(3) History of Present
Illness (HPI)
(4) Allergies
(5) Current Medicines,
(6) Active Problems
(7) Past Medical History
(PMH)
(8) Past Surgical History
(PSH)
(9) Family History
(10) Personal or Social
History
(11)Gynecologic History
(12)Obstetric History
(13)Review of Systems
(14)Vital Signs
(15)Physical Examination
(PE)
(16)Assessment
(17)Tests
(18)Plan
(19)Orders
The 5 Kinds of EMR Errors
Inconsistent Information ◦ Contradictory information
across sections
Incorrect Information ◦ Based on note scenario
◦ Based on clinical guidelines
Incomplete Information ◦ Essential information
omitted
Missing Section ◦ An entire section omitted
Miscellaneous Errors ◦ Inappropriate placement
◦ Use of un-established acronyms
6
User Study Design
Participants
◦ 11 females, 9 males
◦ Gynecologic physicians with
Drexel College of Medicine Using Allscripts EMRs since 2008 or later
Data
◦ 7 fabricated visit notes
different hypothetical
gynecologic patients
Purposefully introduced errors Gold standard errors (total 97)
prepared by 2 clinical investigators
with 20 years of patient visit
documentation
Introduced more incomplete and
missing section errors as they are
more frequent in real world
7
5
52
8
Number of Participants
1 yearexperience
2 yearexperience
3 yearexperience
4 yearexperience
86
2554
4
Total Number of Introduced ErrorsInconsistent
Incorrect
Incomplete
MissingSection
Miscellaneous
User Study Design
Objectives
◦ assess participants’ ability to detect/resolve errors
◦ explicate intuitive strategies
◦ infer guidelines for algorithm design.
Conducted one-to-one session with each participant.
◦ Analysis Stage
present the paper prototypes of the patient notes
participant carefully studies the note, detect any data error(s), and document/annotate them on the same sheet of paper.
◦ De-briefing Stage
the participant answers follow-up questions regarding the detected errors what makes you conclude that certain data are erroneous?
what in your medical training allowed you to detect this error?
what measures would you take to resolve a certain error?
why do you think these errors occur?
8
Results: Note Analysis Stage (avg. 32.6 min)
Re-grouped Errors:
◦ Mod-liability: missing section, missing
information, misc. errors
◦ Hi-liability: incorrect and incomplete
information
Error Precision: 100%
◦ Each detected error could be mapped to a
gold standard item
Error Recall: Recall for hi-liability (0.49)
statistically higher(p<=0.05) than recall for
mod-liability (0.36)
Correlation (Pearson's) between
◦ Task performance(recall) and task
duration
◦ Task performance and years of
experience
◦ Not significant at 0.05 level
Best Recall Performance 70%
◦ Participant P5 had 4 years experience
and spent 53 minutes reviewing the
note
Lowest Recall 17%
◦ Participant P10 had 1 year experience
and spent only 19 minutes reviewing
the note
9
Time
spent
Experience
Hi-liability 0.29 0.43
Mod-liability 0.1 0.19
Results: Debriefing Stage (avg. 13.3 min)
Participants were very confident of their performances during the
note analysis stage, and were very vocal about their experiences.
How do you gain the ability to detect errors? ◦ Field experience of writing EMR
notes in clinical settings 4 participants
◦ Training in medical school 6 participants
◦ Both (Experience + Academic training) 5 participants
Why do the errors occur? ◦ Because of poor
documentation practices 17 participants believe that
physicians Should write for others to be
able to read
Should ask more questions of the patient
Should write in a list format
◦ Because of system design 5 participants
Allscripts EMR propagates all problems through previous visits creating obsolete information
Clinicians tend to write free text note because structured interface is not friendly enough.
10
Results: Debriefing Stage What are the triggers for detecting mod-liability errors?
Detection of abnormal history events
◦ if history of abnormal pap smear in the gynecologic history section is observed, it
must always accompany more information such as the diagnosis date
◦ If history of hypertension in the family history section is specified then it should
also be specified which family member suffered hypertension.
General Information Recall
◦ If the patient is having an annual visit, then HIV screening must be
present in at least one of the sections.
◦ If the patient is over 60 years of age, then a health monitoring plan
should be created and specified in the Plan section
11
Results: Debriefing Stage What are the triggers for detecting hi-liability errors?
Observation of discrepant information between two sections
◦ The reason for visit should be consistent with the active problem list.
◦ The information on the same drugs should match across different
sections, e.g., in one of the study notes
strength of the drug “Fluconazole” is 200mg in the Plan section, and strength of the drug
“Diflucan” (market brand name for same drug) is 20mg in the Orders section
Only 5 participants detected the above error in our study
Detection of abnormal results
◦ Any abnormal body mass index should be alerted in the Plan and
Assessment sections.
Identification of broken information links
◦ Each abnormal result from Physical Examination, should be linked to a
corresponding diagnosis in the Assessment section. Each diagnosis item
should have a corresponding item in the Plan section.
12
General Implications for Error Control Algorithm Design
Despite the experience and expertise, error recall performance < 50% (3 participants had >55%, 5 participants had <30%)
It is imperative to replace existing manual strategies with effective computational algorithms
The participants delivered statistically better performance for hi-liability errors than mod-liability errors
Underlines the significance of learning from their expert abilities to minimize potential physician liability
The computed correlations between performance and experience/time were not significant
No clear conclusion regarding learning-based algorithms
The results on the provenance of abilities suggest that future algorithms should
incorporate domain knowledge from a wide range of sources
learn and infer from the contextual information in the EMR data
13
Guidelines to programmatically fire the Error Detection Triggers
14
Trigger Computational Technique Knowledge Sources
General Information Recall Basic if-then rules Extraction of key information such as age, visit type, etc.
Clinical guidelines • Centers for Disease Control and Prevention • American Congress of Obstetricians and Gynecologists
Detection of Abnormal Results
Basic if-then rules Extraction of examination results.
• Davis’s Laboratory and Diagnostic Tests • Agency for Healthcare Policy and Research • Archimedes 360 Medical Calculator
Detection of abnormal history event
Advanced if-then rules Extraction of abnormal events and their attributes Extraction of medications and their attributes.
Conceptual model for drugs, disease conditions, and habits. • UMLS RxNorm • DailyMed • MedlinePlus
Observation of discrepant information between two sections
Comparison of problems, and medications across sections Drug and Disease recognition
Controlled vocabulary for describing problems and drugs, linkages between drug ingredients, and brand names, drug-drug interactions. • DrugBank • UMLS • FDA National Drug Code Directory • Classification of Diseases, Functioning, and Disability • RxDrugs • Davis’s Drug Guide
Identification of broken links across multiple sections
Extraction of results, diagnosis, plan, order information Linking items from different sections, and discovering the missing links.
Drug indication, prescriptions, physical examination resources. • SIDER 2 • Health Assessment Through the Life Span • Outlines in Clinical Medicine • DailyMed • NDF-RT
Study Summary and Contributions As a pre-step to design EMR error control algorithms, we learn algorithm design lessons from
their abilities and behaviors of physicians on gynecologic visit notes.
Participants could detect only 49% of the inaccuracy and inconsistency
errors, and only 36% of the omission errors from the notes.
◦ Need for more effective and efficient error control solution.
An in-depth investigation of manual strategies helped develop guidelines for algorithm design.
1. 5 data triggers that naturally prompt participants to sense a potential error in gynecologic notes: detection of abnormal examination results, recall of generic clinical guidelines, detection of abnormal history events, observation of discrepant information, identification of broken
information links.
2. Participants can identify the triggers using NLP abilities, and an immense amount of intuitive domain knowledge accumulated through experience and medical school training.
◦ In addition to sophisticated NLP techniques, the algorithms should incorporate a wide range of federally established free resources for clinical guidelines, controlled vocabularies, drugs, diseases, drug indications, gynecologic best practices, etc.
3. We briefly provide the linkages among triggers, NLP techniques, and the relevant trustworthy knowledge sources.
15
Study Limitations
Demanding schedules of our participants
◦ Could devote limited time to the study
◦ Set of derived implications is by no means complete
Study had inherent bias
◦ Participants knew in advance that the note contain errors
◦ This is contrary to real-world practices
Frequency of errors introduced in each note (avg. 13) is
not based on empirical evidence (due to lack of related work)
◦ Some participants might have assumed the notes to contain
fewer errors and terminated their analysis earlier
16
Conclusions and Future Work
the first step to algorithm design by exploring an untapped knowledge resource, i.e., the physicians.
Explored their abilities to detect EMR data errors
Derived algorithm design implications from their intuitive knowledge and personal strategies.
In comparison to the manual expert strategies, the existing automated algorithms only scratch the surface of error control
We plan to design customized algorithms for gynecologic notes by building on the identified triggers
◦ To simulate the narrative information extraction existing NLP algorithms to extract drug, disease and specific clinical information from
texts (Doğan et al., 2010; Li et al., 2012; Névéol & Lu, 2010).
◦ To simulate the physicians’ knowledge in the head utilize, integrate, and organize several available trustworthy knowledge sources hosted by
the US Government.
17
In the USA, medical errors kill more people than highway accidents every year. (Kohn, Corrigan, & Donaldson, 1999).
“30 years from now the EMRs should make sense – otherwise defeats the purpose of EMRs”
– a study participant on documentation malpractices
THANK YOU
Acknowledgements - The 2 anonymous reviewers - The 20 User Study Participants - Drexel’s Jumpstart grant for health informatics - Drexel’ Institutional Review Board
18