Requirements for Useful Data Common data models Standardized - PDF document

Afternoon Breakout Session: Using Big Data to Evaluate Clinician-sensitive Outcomes Connie White Delaney, PhD RN, FAAN, FACMI – Objective: To identify, in this more “nuts and bolts” session, key strategies and tactics that investigators should consider in using big data to evaluate clinician-sensitive outcomes. Requirements for Useful Data Common data models Standardized coding of data Standardized queries http://www.pcornet.org/resource-center/pcornet-common-data-model/ 1

Vision – Inclusion of Nursing & Other Interprofessional Data Clinical Data NMDS Other Data Management Sets Data NMMDS Continuum of Care 4 Example Flowsheet Flowsheet Data Challenges Volume of data There are multiple measures for the same concepts – Different people building screens – Software upgrades – Discipline/ practice specific needs No information models exist – Data driven information modeling required 2

Information Model Development Process Identify Map Clinical Identify Flowsheets Present Validate Data Model Concepts to Topic Concepts UMN – Academic Health Center CDR Flowsheets constitute 34% of all data • 14,564 measure types • 2,972 groups • 562 templates • 1.2 billion observations • 2,000 measures cover 95% of observations Sample Data Source -Clinical Data Models T • Flowsheet Data from 562 10/20/2010 - 12/27/2013 • 66,660 patients Groups • 199,665 encounters 2,696 Flowsheet Measures 14,550 Data Points 153,049,704 3

Development Process Details Identify clinical topic important to researchers/ operations Develop a list of concepts from research questions, clinical guidelines and literature Search for concepts in templates/groups/measures – Search associated groups for additional concepts Add matched concepts to running list Categorize into assessment and interventions Organize into hierarchy Combine similar concepts that have similar value sets Validated by a second researcher Flowsheet Information Models Pain Neuromusculoskeletal System Falls/ Safety Respiratory system Peripheral Neurovascular Vital Signs, Height & Weight (VTE) Genitourinary System/ Aggression and Interpersonal CAUTI Violence Pressure Ulcers Psychiatric Mental Status Exam Cardiovascular System Substance Abuse Gastrointestinal System Suicide and Self Harm Example Information Model 4

What is i2b2? Informatics for Integrating Biology and the Bedside (i2b2) Framework for research cohort discovery Create Flowsheet Ontology in i2b2 14 information models - approximately 81 million new rows i2b2 OBSERVATION_FACT table I2b2 – every row has to be unique Informatics Issues Encountered Redundancy – flowsheet and value sets – 7 blood pressure and 10 heart rate measures – Mapped multiple flowsheet measures to same concept Variations in value sets – Created a unique list of all for same concept Measures with similar names represented different concept – i.e. search “display name” – Urine Output – R IP URINE FOLEY – URINE OUTPUT – URINE OUTPUT.MODIFIED ALDRETE – R NEPROSTOMY URINE OUTPUT 0-unable to void and – URINE OUTPUT (ML) uncomfortable 1-unable to void but comfortable 2-has voided, adequate urine output per device, or not applicable 5

Technical Issues Encountered Free text response – Included name of measure, no data included in i2b2 Multi-response items – Created a separate row OBSERVATION_FACT table Choice list - comment or “other” option – Created a row for each type of comment Numeric response measures - units of measure not clearly identifiable – Modified name to include unit of measure Mapping issues – Changed names to exclude “* | / \ “ < > ? %” – Constructed synthetic value item id’s Names must be unique within first 32 characters – Changed from fully specified names to multiple levels Certified WOC Nurses – Incontinence & Wounds Outcome Variables Description Pressure Ulcers Total number of pressure ulcers (M0450 a-e) Stasis Ulcers Total number stasis ulcers (M0470/ M0474) Surgical Wounds Total number of surgical wound (M0484/ M0486) Urinary Incontinence Presence/management of urinary incontinence or need for a catheter (M0520) Urinary Tract Infection Treated for UTI in past 14 days (M0510) Bowel Incontinence Frequency of bowel incontinence (M0540) Improved/ Not Worse (Stabilize) Outcomes Not Worse Score Bowel Incontinence Frequency Improved (Stabilize) Very rarely /never has BI or has ostomy 0 for bowel elimination 1 Less than once weekly 2 One to three times weekly 3 Four to six times weekly 4 On a daily basis 5 More often than once daily 6

Individual Patient Outcomes Using Data Visualization to Detect Client Risk Patterns Monsen, K. A. et al., 2014 Each image (sunburst) was created in d3 from public health nursing assessment data for a single patient. Data were generated by use of the Omaha System signs and symptoms and Problem Rating Scale for Outcomes Key: • Colors = problems • Shading = risk • Rings = Knowledge, Behavior, and Status • Tabs = signs/symptoms Documentation patterns suggest a comprehensive, holistic nursing assessment. Kim et al. found that the presence of mental health signs and symptom tends to be associated with more diagnostic problems and worse patient condition Kim, E., Monsen, K. A., Pieczkiewicz, D. S. (2013). Visualization of Omaha System data enables data-driven analysis of outcomes. American Medical Informatics Association Annual Meeting, Washington D. C. Funded by a gift from Jeanne A. and Henry E. Brandt. Using Data Visualization to Detect Nursing Intervention Patterns Monsen, K. A. et al., 2014 Each image (streamgraph) was created in d3 from longitudinal public health nursing intervention data for a single patient. Data were generated by use of the Omaha System in clinical documentation Key: • Colors = problems • Shading = actions (categories) • Height = frequency • Point on x-axis = one month From 403 images, 29 distinct patterns were identified and validated by clinical experts Documentation patterns suggest both a unique nurse style and consistent patient- specific intervention tailoring Monsen, K.A., Hattori, K., Kim, E., Pieczkiewicz, D. (In review). Using visualization methods to discover nurse-specific patterns in nursing intervention data. Streamgraph development funded by a gift from Jeanne A. and Henry E. Brandt. 7

Do PHNs Tailor Interventions? Public Health Nurses Signature Styles? Challenges of Secondary Analysis of Big Data Database/Data Dictionaries Data Extraction – Feature selection Data Cleansing – Missing values, outliers, errors, redundancies, transformation… Analysis – Exploratory • Statistics, data mining – Predictive • Machine learning • Testing Algorithms Cleansing Model Evaluation – Extraction Testing on new data Analysis 8

Requirements for Useful Data Common data models Standardized coding of data Standardized queries Vision for Data in a Clinical Data Warehouse Clinical Data NMDS Other Data Management Sets Data NMMDS Continuum of Care Celebrating our foundation for “Big Data/Data Science” Global standards eMeasures EHRs Magnet, etc Resources Workforce 9

Nursing Minimum Data set (NMDS) National Standard – SNOMEDCt, LOINC Werley, HH & Divine, E., & Zorn, C. (1988). Nursing Minimum Data Set Data Collection Manual. University of Wisconsin, Milwaukee, WI Huber D, Schumacher L, Delaney C. Nursing management minimum data set (NMMDS). J Nurse Adm . 1997;27(4):42-48. z.umn.edu/bigdata 30 10