Copyright 2015
Use of Novel Predictive Models to Improve Hospital Readmission Program
1
Presenters
• Jason Burke, MA
• Senior Advisor & Faculty at UNC Health Care and School of Medicine
• Michael Cousins, PhD, MS
• President and Chief Analytics Officer, Forecast Health
2Copyright (2015)
Agenda and Objectives
Agenda
1) Overview of readmission program
2) Readmission model development process
3) Readmission model results
Learning Objectives
1) Describe the process of integrating EHR and socioeconomic, behavioral, and lifestyle factors behind the hospital’s firewall
2) List the variables that were found to be meaningful
3) Explain the predictive modeling methodology and the similarities and differences with claims-based models
3Copyright (2015)
Please ask questions throughout!
Who is UNC Hospitals?
History
Academic Medical Center in Chapel Hill with outpatient services across North Carolina
• 853 staffed beds (853 licensed)• >7,800 co-workers• >1,100 attending physicians• 780 residents
• >77,000 ED visits• >30,000 surgeries• 270,000 inpatient days• FY15 Net Rev = $1.5B
5
5Copyright (2015)
Statistical PerformanceLACE benchmark
Length of stay
Acuity
Charlson comorbidity
Emergency department
Readmission Program Overview
1) Risk modeling initiated with participation in CMS “Community-based Care Transitions Program” (CCTP)
2) Initial model developed for Medicare patients, then expanded to all adult patients
6Copyright (2015)
(3 admits OR 3 chronic Dx) AND
10+ discharge Rx?
2 admits OR 2 chronic Dx?
Yes
Yes
No
No
High Risk
Med Risk
Low Risk
AdmittedPatient
Characteristics of a New Approach
7Copyright (2015)
TraditionalApproach
New Approach
Which patients? Sicker patients Riskier patients
Which risk factors? ? Components of risk
What is changeable? ? Patient experience
What actually works? Overall program Specific program elements
• How do we rationalize fixed resources?
• How do we iteratively improve?
• What is often associated with bad outcomes?
• What is likely to happen in the future?
Powering a Different Approach
8Copyright (2015)
Predictive Analytics
EMR +
3rd Party Data
Clinical + Financial
Perspectives
Closed Loop Learning
Agenda
1) Overview of readmission program
2) Readmission model development process
3) Readmission model results
9Copyright (2015)
Readmission Model: Analytic Plan
10Copyright (2015)
(2) Index Hospitalization
30-day
90-day
Time
(1) Pre-Index Hospitalization (3) Readmit(s)
Design
• 3 time periods
Patients
• >18 yo and >1 Hospitalization
• 63k patients across 4 years with 4,500 readmits (15%)
Operational Process
11Copyright (2015)
PII Only+
ConsumerData
PII OnlyGeo
HouseholdPerson
PHI
Readmission Model: Data
12Copyright (2015)
• Consumer data at 3 levels
1) Geographic
• Census tract, ZIP, ZIP+4, Block group
2) Household• Street address
3) Person• Person
SocioeconomicImputed incomeRural/urbanMedian age
Ethnic distributionEducational attainmentFood desert, etc.
Retail purchasesDisposable incomeCaregiver availabilityEmployment status
Proximity to clinical and social Number of children
Marital statusDependents Automobile ownershipURL/website categoriesCredit risk proxyEducational attainmentGambling enthusiast
Health and fitness lifestyle interestsTraveling and arts interestsPet ownershipEthnicityPlus clothing size
Data Illustration
13Copyright (2015)
• ZIP codes may mask financial stress
0
5
10
15
20
25
30
0
50,000
100,000
150,000
200,000
250,000
300,000
20 25 35 45 55 80 100
N Pct of Total (%)
Co
nsu
mers
(N
)P
erc
en
t of T
ota
l (%)
<$20k $25k $35k $45k $55k $60-$99k >$100k
Household Income ($)
Selected Variables From Predictive Model
14Forecast Health Confidential
Hig
he
rR
isk
of
Re
adm
issi
on
Low
er
Ris
k o
f R
ead
mis
sio
n
• More outpatient encounters in pre-index hospitalization period
• 6 selected diagnoses including endocrine, nutritional and metabolic diseases; pneumonia, complications of procedures
• Diagnosis of hypertension complicating pregnancy childbirth
• More provider encounters and education (high school or higher)
• Higher blood pressure (Hypertension stage 2)
• More unique inpatient providers in pre-index hospitalization period
• Higher pain intensity reported at prior outpatient visit
• More unique inpatient providers in pre-index hospitalization period and affordability
• Diagnosis of complications of medical care and education (high school or higher)
Each variable is multiplied by a weighting factor (higher weights in larger font)
Agenda
1) Overview of readmission program
2) Readmission model development process
3) Readmission model results
15Copyright (2015)
16Copyright (2015)
False Positive
False Negative
Specificity
Sensitivity
Clinical and Financial Performance
Statistical Performance
Optimization Scenarios
Evaluated 3 ways:
1) Statistical Performance
2) Clinical Performance
3) Financial Performance
Predictive Model Performance
2,515more patients per year
17Copyright (2015)
Statistical PerformanceBetter than our Version 1
39%More correct identifications per year (sensitivity)
6%Fewer incorrect identifications per year (specificity)
0.91Overall predictive power (c-statistic)
18Copyright (2015)
Clinical Performance
-60.0%
-50.0%
-40.0%
-30.0%
-20.0%
-10.0%
0.0%
-
0.03
0.06
0.09
0.12
0.15
Pct
Re
du
ctio
n
Ne
w R
ead
mit
Rat
ePct Intervention Impact
V1 New Readmit Rate FH New Readmit Rate Pct Difference
Better ability of Version 2 to reduce readmissions
Readmissions as a Function of Program Impact
Category Readmission
Version 1 10%
Version 2 8%
At 50% Intervention Impact
V1 New Admit Rate V2 New Admit Rate Pct Difference
Readmissions 2 points or 20% lower
Statistical Performance:Claims vs EHR/consumer-based models
19Confidential | Copyright 2015
0
0.04
0.08
0.12
0.16
0
0.2
0.4
0.6
0.8
Claims EHR & Census
C-Stat IDI
• C-statistic = 18% higher• Integrated Discrimination Improvement = 721% higher
Readmission Accuracy Comparison• EHR and consumer-based better• Reasons may include:
• Lack of clinical detail (ex: vitals)• Timeliness
Recommendation• Use EHR and person-level consumer data:
• For better predictions• Use claims data:
• For in/out of system utilization• Combine if possible
Claims EHR & Consumer
Readmission Prediction
C-S
tati
stic
IDI
+18%
+721%
Conclusions
1) Readmission program overview
• Moving from traditional to the “new approach” based on the 4 pillars
2) Readmission model performance
• Version 2 outperforms LACE and our Version 1 model
• “Simpler isn’t always better – sometimes better is better”
• Expected to lead to substantial readmission improvements and improve our economics
20Copyright (2015)
Many Thanks!!!!
21Copyright (2015)
Contact Info:
22Copyright (2015)
Financial PerformanceFavorable Impact
Savings across:
1) Operating Expense: Availability to double-up or redeploy staff
2) Gain-share: New revenue from commercial gain-share contracts
3) CMS Penalties: Reduced due to lower readmission rate
4) Value-based Reimbursement: Higher margins due to reduced readmissions
23Copyright (2015)
Financial PerformanceFavorable Impact
Savings Under Value-Based Reimbursement1) 2,100 beds 2,515 new patients
a) 2,515 new patients x $11,000 avg readmission cost = up to $27m gross savingsb) $27m x 25% readmission program impact = $6.75mc) $6.75m - $2.5m new staffing and analytics = $4.25m net savings
2) 210 beds 252 new patientsa) 252 new patients x $11,000 avg readmission cost = $2.7m gross savingsb) $2.7m x 25% readmission program impact = $690kc) $2.7m - $250k new staffing and analytics = $440k net savings