Predictive Modeling and the Economics of Medical Management

1
Predictive Modeling and the Economics of Medical
Management

• Ian Duncan, FSA, MAAA

Lotter Actuarial Partners 15 East 26th Street New
York, NY 10010 www.lotteract.com
Lotter Actuarial Partners
2
Healthcare Trend has accelerated
If you think Healthcare is expensive now, wait
until the Government provides it for free
(P.J. ORourke)
Source Towers Perrin Health Care Cost Survey.
3
Many believe that a small number of high
utilizers drive cost increases
How do we identify the future high utilizers?
How do we help themmanage their care?
How will we measurethe cost reduction?
4
The basis of Predictive Modeling Member Costs
dont behave the same way all the time
Average
5
Predictive Modeling is NOT about Cost
Prediction..

• ..it IS about resource allocation.

• Where/how should you allocate resources?
• Who is intervenable or impactable?
• What can you expect for outcomes?
• How can you manage the key drivers of the
  economic model for better outcomes?

6
Key drivers of the economic model

• Prevalence within the population (numbers)
• Population Risk Ranking
• Data quality
• Reach/engage ability
• Cost/benefit of interventions
• Timeliness
• Resource productivity
• Random variability in outcomes

7
The yield curve Ranking population by risk
characteristics and likely outcomes
8
Some observations on Care Management

• Low prevalence of target conditions in
  population.
• Difficult to identify (incomplete data privacy).
• Difficult to reach and engage (member data call
  blocking lethargy and resistance).
• When member wants to engage, its too late!

9
Some observations on Care Management

• Interventions are expensive.
• Professional staff
• Cost of outreach relative to lift
• Cost of increased compliance.
• We dont yet have a good idea of what works.

10
An Interactive Economic Model

• Allows you to
• Plan the scope, timing and staffing of a program.
• Project Program Financial Results.
• Test the sensitivity of the results to different
  variable values.
• Track progress of a program during the
  implementation.

11
Dont be alarmed.just an example!
From Ian Duncan and Arthur Robb Health Risk
Management in Jain and Shapiro, eds.
Intelligent and Other Computational Techniques
in Insurance. World Scientific, 2003.
12
Understanding the underlying economicslets you
optimize your investment
Gross cumulative yield and program cost (000)
From Duncan, I.G. Understanding the Economics of
Disease Management. Part 4 of the Study
Actuarial Issues of Care Management
Interventions. Schaumburg, IL. Society of
Actuaries, 2004. Available at www.soa.org.
13
Understanding the underlying economicslets you
optimize your investment
Distribution of Return on Investment
14
Types of Predictive Modeling Tools

• Risk Groupers
• Statistical Models
• Rules-Based Models

15
Different Models may be appropriate in different
circumstances

• What matters sensitivity and specificity
• Risk Groupers Actuarial/pricing models Need
  specificity
• Statistical Models Case/Disease Management
  Need sensitivity
• Rules-Based Models Population Management
  Provider Interventions Both?

16
Selected Risk Groupers
See Cumming et al A Comparative Analysis of
Claims-Based Methods of Health Risk Assessment
for Commercial Populations Schaumburg, IL.
Society of Actuaries, 2002.
17
Comparison of Different Risk Groupers
See comments later about R2 model with
re-calibrated weights.
18
Statistical Models

• Is often done in-house, or by DM vendors.
• Leverage one of the risk-groupers, or use
  in-house models.
• Advantage you can constantly refine and develop
  models to support your programs. Additional
  (self-reported) data can be incorporated
  dynamically

19
Rules-based Models

• Based on clinical literature applied to claims
  data.
• Needs a vendor.
• Advantage one type of model that is
  provider-focused.

20
Evaluating Model Performance

• R2 is the popular test statistic, but flawed.
• Importance of Sensitivity and Specificity for
  your project.
• ROC Curves are a good test.
• Dont forget Impactibility and Plausibility.
• Finally, did you save money?

21
Its no use identifying patients unless you have a
program/process to intervene with them

• Examples end-stage renal disease rare diseases
  population programs, etc.
• But if you have a good program, Predictive
  Modeling can help find patients, target them, and
  allocate resources efficiently.

22
What do you want Predictive Modeling to do?

• EXAMPLES
• Assessing health risk for lifestyle/behavior
  modification
• Outreach for early impact on preventable diseases
• Finding members for case management/disease
  management
• Stratification of patients for disease management
• Find potential gaps and errors in care
• Cost savings through quality improvement
• Provider profiling/reimbursement
• Underwriting/actuarial uses

23
Example Hospital Admission Targeting

• The problem
• Long-stay, expensive admissions
• 15 of admissions account for 40 of acute days.
• Where should the health plan focus its Case
  Management Resources?

In any health plan, a significant share of
hospital resources is consumed by a small
sub-population.
24
Model Development

• Create episode severity variable.
• Develop model.
• Test model for significance/accuracy.
• Implement
• Significant variables in model
• Patient age.
• Previous acute admit in the past 90 days.
• Hospital claims in the past year (acute or
  otherwise).
• Disease Management Program Risk Rank
• Day of the week of admission
• AND Wild Card Variable does the patient have a
  pet at home.

25
Testing Model for significance/accuracy
Using the Risk Ranking Algorithm, newly admitted
patients can be classified into three categories

• High Risk (red) 14 of admits
• Medium Risk (yellow) 42 of admits
• Low Risk (green) ) 44 of admits

48 of High-Risk (red) patients have episodes of
Severity 4 or 5
50 of Low-Risk (green) patients have episodes of
Severity 1 or 2
26
Implementation daily routine
Step 1
The intake coordinator receives a daily admit
notification.
For each patient the intake coordinator completes
a brief admission survey in DSManager.
Step 2
DSM
High Risk
DSManager combines the admission survey with
historical patient information and assigns the
patient a risk ranking in real time on nurse C.M.
task-list.
Step 3
Medium Risk
DSM
Low Risk
The intervening nurse contacts admitting hospital
on behalf of high and medium risk patients.
Tasks are prioritized according to Risk ranking.
Step 4
High Risk
Medium Risk
DSM
27
DMAA has developed a prototype Predictive
Modeling Questionnaire for RFPs
28
References/Further Reading

• Ash AS et al. How well do models work?
  Predicting health care costs Proc Section on
  Stat Epidemiol in American Statistical
  Association, 1988
• Celebi D. "The power of predictive modeling"
  August 2003, Healthcare Informatics
• Cousins M et al. "Introduction to predictive
  modeling for disease management risk
  stratification" DM Volume 5, Number 3, 2002
• Cumming RB et al. "A comparative analysis of
  claims-based methods of health risk assessment
  for commercial populations" SOA May 24, 2002 at
  www.soa.org
• Duncan I et al. A prediction model for targeting
  low-cost, high-risk members of managed
  organizations Am J. Mgd. Care, May 2003
• Duncan, I. Understanding the Economics of Care
  Management. Part 4 of the study Actuarial
  issues of Care Management Programs. Published
  by the Society of Actuaries. Available at
  www.soa.org.
• Duncan I, Robb, AS, Population Risk Management
  in Jain Shapiro, eds. Intelligent and other
  Computational Techniques in Insurance, World
  Scientific, 2003.

29
Reference/Reading

• 8. Iezzoni LI. The risks of risk adjustment
  JAMA 19972781600-07
• 9. IHCIS "Predictive Risk Modeling in the
  Healthcare Industry The needs-awareness gap
  among healthcare payers
• 10. Mari Edlin "New data sources multiply the
  power of predictive modeling The evolution of
  predictive modeling continues with integrated
  data sources and weighted variables" February
  2003, Managed Healthcare Executive.
• Zhao Y et al. Measuring population health risks
  using inpatient diagnoses and outpatient pharmacy
  data Health Serv Res 200136180-192
• "Predictive Modeling in Health Care Going beyond
  prediction to prevention and care" Vol. 3, Issue
  5, May 2003, Disease Management and Quality
  Improvement Report
• 13. DMAAs Dictionary of Disease Management
  Terminology (2004)