Henry Ford Hospital Inpatient Predictive Model Mike Meitzner Principal Management Engineer Henry Ford Health System Detroit, Michigan
Outline HFHS background CMURC relationship Model Goals Data Cleansing Roadmap / Forecasting Steps Association / Sequencing Regression Model Development Results Next Steps
Henry Ford Health System Henry Ford Hospital (Detroit) 700 beds in operation, 40,000 annual admissions Community hospitals Bi-County Wyandotte Kingswood Henry Ford Medical Group ~700 physicians in 40 specialties Over 20 outpatient facilities Community Care Services Health Alliance Plan Corporate Offices (Accounting, AP, Audit)
HFHS/CMURC relationship Investigated a novel approach to the forecasting problem. Project Team: HFHS Management Services CMURC staff CMU faculty (College of Health Professions and the Statistics department)
Model Goals Predict inpatient volumes 3-123 months into the future based on HFHS current activity levels and historical trends. Use outpatient and emergency department (ED) activity data to predict future inpatient levels in the short term. Understand the typical service paths which patients take through the system Identify specific outpatient clinic services or diagnostic codes that translate into larger inpatient levels The duration of the project was scheduled to be 6 months
Data cleansing and preparation Combining data from different tables presented significant hurdles These hurdles could be addressed because the source tables shared a unique patient identifier that was consistently applied across the System. Complexity in the definitions of visits, admits, length-of of-stay (LOS), etc. were the cause of most of the project time.
Data cleansing and preparation Three major datasets Patient Encounters (Encounters) Stores billing information for every patient encounter, outpatient or inpatient, including the site and date of service. Inpatient Medical Record (Medical Records) stores inpatient data for HFH, and includes Length of Stay (LOS) and primary DRG Inpatient (PEMS). Includes unit level LOS and holds admit and discharge timestamps for each unit that the patient spent time on.
Patient Classifications Patients with encounters in HFHS into the three groups: Case A: Patients who have been admitted to HFH with prior encounters in the non-hfh environment (clinics and urgent care facilities); Case B: Patients who have not been admitted to HFH, although they have had encounters in the non-hfh environment; and Case C: Patients who have been admitted to HFH without any prior HFHS encounters.
Road Map Patient Encounters Classify records: a=i wi. E or O b= no I c=i w/o E or O a b c Medical Records Map weekly admits to day-of-week PEMS List of unusual events Map DR_SPEC to LOS & NUR_UNIT Seq Rules Recent E/R & Outpt Non-HFHS referrals HFHS referrals Admits by week by DR_SPEC Forecast integration DEMAND beds per NUR_UNIT Major Deliverable To Capacity & resource planning
Forecasting Steps HFHS referrals Last 6-mon E/R & O visits Visit to Admit rules Fcst Admits by doctor specialty, week using time-delays Non-HFHS referrals & low volume HFHS referrals Profile of future weeks to be forecasted Regression parameters Fcst Admits of non-hfhs by week Time-delay factors by rule Non_HFHS DR_SPEC factors Fcst Admits by DR_SPEC by week Fcst non-hfh by DR_SPEC by week Extend non-hfhs admits to DR_SPEC Apply day-of-week factors Apply length of stay factors Apply DR_SPEC to Nurse Unit factors Fcst by day by Nurse Unit
Referral Sources Henry Ford Medical Center - West Bloomfield Henry Ford Medical Center - Fairlane Henry Ford Hospital - etc. - Henry Ford Medical Center - Sterling Heights Referrals from non-henry Ford facilities
Association / Sequencing Each patient visit is associated with a location and a category Patient visits are sequenced, with particular attention placed on those resulting in inpatient visits.
Association / Sequencing SAS Enterprise Miner 4.3 was used as a tool for the Sequence analysis. Patients from case A (at least one I and at least one O/E ) were selected to relate admits (I) to preceding O and E visits, based on the service date.
Association / Sequencing Records identifying visits were used as input for Association Node. The number of generated rules depends on the support levels requested. If we used a support that was too low, the rules generated do not produce statistically significant time delay distributions.
Association / Sequencing For each rule we calculate the time lag (in weeks) between outpatient/emergency visits and admit to the hospital. The timelag distributions were fitted to parametric distributions such as Lognormal, Gamma and Weibull. The two parameter Weibull distribution offered the best fit. For each rule, the two parameter values, shape and scale, were calculated.
Regression Forecasts Stepwise regression was applied to the following datasets: Residuals from rule-based forecasting (A Residuals) Admits from outside the HFHS (Case C) Admits with no DR_SPCY_CD assigned Total forecast of admits for 6-months 6 to one year in future
Regression Forecasts Stepwise regression allowed only the variables with specified significance to enter and stay in the model. The entry parameter significance was 25% and 15% significance was needed to keep the variable. The values obtained from regression were used as parameters for the prediction. The regression was based on weekly total admits for each of the datasets.
Regression Forecasts Time variables were included to allow for changes in trends over time. Binary variables were included in the stepwise regression to offer further adjustments in the prediction. These variables were: Strict Holiday, Lax Holiday, Season (school), and Season2 (winter).
Forecast Timelines Jan2001 Generate forecasts E/R, O & Admits Generate forecasts Compare O/E visits & admits 2005 + Fcst to Actual Jan2000 Sept2003 Mar2004 Aug2004
Day of Week Distribution The percentage of weekly admits for each day of the week is very constant across non-holiday weeks. A set of daily factors for each quarter of the year was applied. For weeks containing a holiday, a column to the table was added to store the percentage of the holiday-week admits for holiday itself.
LOS Distribution Length of stay had to be incorporated into the model to calculate hospital census. An average length of stay (ALOS) was unsuitable for this prediction because the distribution of a length of stay is skewed. The complexity of patient types is not captured by an average. The distribution table used the length of stay (days), Dr. specialty code, and total number of patients with that code as the basis for bed requirements.
Nurse Unit Assignment Used historical admission patterns by specialty to determine unit placement The PEMS dataset contains information for each patient on the LOS spent in each nurse station. Matching PEMS with Medical Records was done in order to obtain more detailed information for each patient DR_SPCY_CD. A final mapping table was created Normalized list based on the number of hours spent in different Nurse Stations by specialty.
Daily Predicted Census 700 600 500 400 300 200 100 0 5/1/05 6/1/05 7/1/05 8/1/05 9/1/05 10/1/05 11/1/05 12/1/05 1/1/06 2/1/06 3/1/06 4/1/06
Sample Week Predicted vs. (Actual) Sunday Monday Tuesday Wednesday Thursday Friday Saturday Unit 12-Jun-05 13-Jun-05 14-Jun-05 15-Jun-05 16-Jun-05 17-Jun-05 18-Jun-05 ICU's C5M 25 (28) 26 (29) 27 (29) 28 (28) 28 (28) 28 (29) 27 (31) C6N 10 (11) 10 (10) 11 (10) 11 (9) 11 (10) 11 (10) 11 (9) Cardiology H5 25 (26) 27 (30) 29 (30) 29 (27) 29 (23) 30 (26) 27 (23) I5 25 (30) 27 (30) 28 (30) 29 (28) 29 (24) 29 (23) 27 (22) Medical GPU's B1 26 (30) 27 (30) 29 (28) 29 (27) 30 (28) 30 (28) 28 (29) B2 25 (29) 27 (27) 28 (28) 28 (26) 29 (26) 29 (27) 27 (23) B6 20 (19) 21 (21) 22 (22) 23 (20) 23 (19) 23 (18) 22 (16) F6 11 (11) 12 (13) 12 (16) 13 (15) 13 (13) 13 (11) 12 (10) Surgical GPU's B4 21 (16) 22 (19) 24 (24) 24 (28) 24 (28) 25 (21) 23 (14) F4 24 (28) 25 (27) 26 (26) 27 (28) 27 (27) 27 (25) 26 (23) Total predicted 511 542 571 579 587 592 553 Total census 551 586 603 596 589 575 547
Forecast vs. Actual IMG Forecast Versus Actual 240 220 200 180 160 140 120 100 5/2/2004 5/16/2004 5/30/2004 6/13/2004 6/27/2004 7/11/2004 7/25/2004 8/8/2004 8/22/2004 9/5/2004 9/19/2004 10/3/2004 10/17/2004 10/31/2004 11/14/2004 11/28/2004 12/12/2004 12/26/2004 1/9/2005 1/23/2005 2/6/2005 2/20/2005 3/6/2005 3/20/2005 4/3/2005 4/17/2005 IMG_A IMG_P
Model Applications Accurate forecast for budgeting purposes Determine bed capacity requirements Scenario testing and what if analysis. Analyze the flow of patients across the continuum. Outpatient Yields Origin Analysis
Next Steps Refine the model on a specialty by specialty basis. Use the model to aid in developing the 2007 budget. Use the model in strategic decision making.
Contributing CMURC staff Dr. Michael H. Kennedy FACHE, Associate Professor, College of Health Professions, Central Michigan University James Mentele Senior Research Fellow CMURC Allison Mentele Research Assistant CMURC Dr. Carl Lee Professor, Department of Mathematics, Central Michigan University Lyubov Fishman Research Associate CMURC Sri Sundaresan Research Assistant CMURC Eric Willoughby Research Assistant CMURC