Documentation Proliferation Effect in Electronic Medical Records Adele Towers, MD and Mark Morsch, MS DISCLAIMER: The views and opinions expressed in this presentation are those of the author and do not necessarily represent official policy or position of HIMSS.
Conflict of Interest Disclosure Adele Towers, MD, MPH Salary: N/A Royalty: N/A Receipt of Intellectual Property Rights/Patent Holder: N/A Consulting Fees (e.g., advisory boards): N/A Fees for Non-CME Services Received Directly from a Commercial Interest or their Agents (e.g., speakers bureau): N/A Contracted Research: N/A Ownership Interest (stocks, stock options or other ownership interest excluding diversified mutual funds): N/A Other: UPMC has a financial interest in the Optum Clinical Documentation Improvement Module 2 2013 HIMSS
Conflict of Interest Disclosure Mark Morsch, MS Salary: OptumInsight Royalty: N/A Receipt of Intellectual Property Rights/Patent Holder: N/A Consulting Fees (e.g., advisory boards): N/A Fees for Non-CME Services Received Directly from a Commercial Interest or their Agents (e.g., speakers bureau): N/A Contracted Research: N/A Ownership Interest (stocks, stock options or other ownership interest excluding diversified mutual funds): United Health Group Other: N/A 3
Learning Objectives Define the challenge of documentation proliferation in electronic medical records (EMR) Describe how Natural Language Processing (NLP) technology parses and analyzes the medical record and recognizes components of ICD-9 and ICD-10 codes Explain how natural language processing can help organizations find EMR documentation deficiencies before patient discharge 4
About University of Pittsburgh Medical Center UPMC is one of the leading nonprofit health systems in the United States, headquartered in Pittsburgh, Pennsylvania. UPMC s unique strategy of combining clinical and research excellence with business-like discipline translates into high-quality patient care. UPMC is Pennsylvania s largest employer, with more than 55,000 employees. UPMC Quick Facts Hospitals 20 Average Daily Census PUH: 626 SHY: 392 Inpatient Discharges Per Year PUH: 34,267 SHY: 24,980 Surgeries Per Year PUH: 23,540 SHY: 20,126 ED Visits Per Year PUH: 57,804 SHY: 39,686 5
EMR Environment at UPMC Cerner, HPF, MARS Cerner PowerNotes 100% electronic at one facility 50% electronic at other 2 facilities CAC since Sept 6, 2008 Medipac billing system 6
Documentation Gaps in the EMR Cut & paste phenomenon new information often buried When doctors type they don t include much information Symptoms not diagnosis are documented Doctors can t find correct diagnosis from pick-list Need to communicate with physician in their workflow 7
Financial Impact UPMC captures $12 million per year from retrospective review of medical records 2011 external documentation audit of UPMC s records showed that the system was losing $17.8 million per year despite best effort of current retrospective process Audit confirmed that since the system had moved from paper to electronic records, the case mix index (CMI) had decreased Typically means hospitals aren t getting paid as much due to lower documented severity of illness. 8
Clinical Documentation Improvement (CDI) Seeks to improve the quality of provider documentation to more accurately reflect services rendered. Important consideration in the transition to ICD-10. Address potential gap between the content of clinical documentation and the required specificity for ICD- 10 coding. Concurrent CDI is a proactive approach, identifying and correcting potential documentation deficiencies during the patient s stay. 9
Case Finding is Often a Wasted Effort ACDIS CDI Staffing Survey*: CDI specialists conduct 8-12 new reviews per day. Each CDI specialist spends between 33 and 48 minutes per initial review. Average salary for CDI Specialist $60K/yr. ($28.84/hour) Source: Simply Hired Percent of Reviews Resulting in a Query Percent of Respondents 0 10% 7% 11 20% 22% 21 30% 36% 31 40% 15% 41 50% 7% 51 60% 6% 61 70% 2% 71 80% 2% 81 90% 1% 91 99% 0% 100% 0% Total 100% 87% of respondents <50% of reviews result in a query. 36% of respondents 2 out of 3 reviews unnecessary *Source: ACDIS CDI Work Group. White Paper: CDI staffing survey provides estimates on record reviews, productivity considerations. HCPro, Inc, 200 Hoods Lane, Marblehead, MA, 2010 10
Transforming CDI with NLP Natural language processing (NLP) is transforming HIM & coding with computer-assisted coding (CAC) solutions Benefits - Productivity, accuracy, efficiency, transparency, manageability CDI programs shares these same goals Harness the power of CAC to drive CDI However CAC is not the same as CDI Not limited to finding only code-able facts, but clinically significant events that are evidence of an information gap 11
Natural language processing and CAC Computer Science Medical Coding Linguistics 12
Natural language processing and CAC NLP for CAC Computer Science NLP Medical Coding Linguistics 13
Natural language processing and CDI NLP for CAC CDI Computer Science NLP More General Medical Knowledge Medical Coding Linguistics 14
Factors Aligning NLP with CDI 1. Accurate abstraction of medical evidence to automate case-finding 2. Clinical information model that supports consistent query decisions 3. Compositional approaches to NLP to recognize complex query scenarios 15
Case Finding Automation with NLP NLP can extract the clinical evidence that indicate gaps in documentation Like in CAC, recall and precision are important measures of accuracy Goal is high recall and high precision High recall ensures that a high proportion of relevant clinical events are captured Capture important facts that can escape manual processes High precision means CDI specialists don t waste time reviewing cases that don t have gaps Comparing CDI evidence to CAC results provides automated validation 16
Alpha Testing NLP Case Finding Precision 100 95 90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0 n=308 cases with 387 total markers 17
CDI Information Model Consistent results require a well-defined set of policies with training and audit programs Use evidence-based criteria and national definitions to create markers Ensure information is abstracted and interpreted following standard guidelines Standardized information model for CDI Sound basis to construct queries Reduce variability in interpretation of potential CDI scenarios Drive requirements for NLP abstraction and business rules to combine data elements 18
Three Tier Information Model Scenario Marker Scenario Marker Marker Source=CDI Marker Label = Condition or Procedure Marker Type = Type of Marker Strength = High, Medium, Low SNOMED Concept ID = simple or complex SNOMED representation Scenario a group of indicators that indicate the reason for the Marker Indicator Indicator Indicator Scenario Label Strength SNOMED Concept ID Indicator Indicator Label Indicator Type Finding or lab or vital or meds or supplies with full inherited output SNOMED Concept ID 19
Compositional Approach to NLP NLP for CDI cannot solely rely on narrative text Lab orders or results, radiology reports, medication orders, and vital signs are all important sources of CDI evidence that are often structured data CDI markers are formed by logical combinations of indicators Two advanced forms of linguistics are important Pragmatics to recognize how context contributes to meaning Discourse analysis to synthesize meaning from multiple sources 20
Pragmatics What is the context? - Low sodium value -Patient completed marathon today 21
Discourse Analysis What are the broader meanings? Current Symptoms Medical History Findings Diagnosis Treatment New or Existing Problem? Findings Relevant or Incidental? Diagnosis Complicated by Chronic Condition? Which Symptoms Related to Final Diagnosis? How is the Treatment Supported by Medical Evidence? 22
Two types of CDI opportunities NLP must be able to handle Example 1: Specificity Physician documents CHF improving. NLP Identifies CHF in History and Physical CHF in progress note Suggests Code for Unspecified CHF Approach to Query Engage Physician to Provide Specificity in CHF Diagnosis Acute vs. Chronic Diastolic vs. Systolic Acute on Chronic Example 2: Clinical Clarity Physician documents fluid retention and shortness of breath improving. NLP Identifies Pulmonary Vascular Congestion in CXR Ejection Fraction of <30% in Echo BNP of 700 IV Lasix in MAR Approach to Query Engage Physician to Clarify Clinical Facts Ascertain if there is a diagnosis that could be added to reflect the clinical picture and rationale for treatment of this patient Subsequent query for specificity in diagnosis if indicated Easy to Moderate High 23
Workflow Concurrent CDI Case Finding Continuous Business Rules Logic processing of the EMR data through NLP to both code and apply case Financial Class Physician Service Revenue Code Location finding rules to each How should it be routed? admission. Directly to physician? Peer Advisor If a case is marked for CDI, ensure it conforms to business rules for presentation to a user: CDI Specialist Specific User CDI Manager Coder Passive Query Building Query passively built with minimal (if any) additional editing and update required by CDIS Presentation to physician either interfaced to EMR, Inbox or via PQRT Portal. Query Response Returned to NLP 24
System Built Queries vs. Manually Built Dear Dr. What kind of CHF is being treated? 25
EMR Case Example CEREBRAL EDEMA 26
27 CDI Marker Mention of Cerebral Finding
28 First Radiology Finding
29 Second Radiology Finding
30 Swelling Noted in Operative Note
Code Selection Financial Impact Original Post NLP/Rules Engine DRG 27 25 CC/MCC NA 348.5 Reimbursement $12,912.58 $29, 798.65 Severity of Illness 1 2 31
Conclusions Challenges of EMR documentation Clinical Documentation Improvement programs can address documentation gaps Three key factors aligning NLP and CDI Case finding automation Clinical information model Compositional NLP Concurrent CDI workflow integrated with electronic physician query Encouraging early results from alpha testing 32
Thank You! Contact Information Adele Towers - TowersAL@upmc.edu Mark Morsch - mmorsch@alifemedical.com 33