Evariant has partnered with, and are using DataRobot for multivariate predictive analytics because it is a flexible, robust, and extremely efficient tool for maximizing our modeling efforts, as well as an example of leveraging high-end data science and tools in the healthcare industry.

1. MODELING IN THE
HEALTHCARE INDUSTRY:
A COLLABORATIVE APPROACH
William B. Disch, Ph.D.
Director, Analytics
Evariant
O P E N
D A T A
S C I E N C E
C O N F E R E N C E
BOSTON 2015
@opendatasci

2. Abstract
Evariant has partnered with, and are using
DataRobot for multivariate predictive analytics
because it is a flexible, robust, and extremely
efficient tool for maximizing our modeling
efforts, as well as an example of leveraging high-
end data science and tools in the healthcare
industry.
… DataRobot helps in automating many routine processes like finding important variables, variable
transformation, variable selection, model building and scoring. As a result, we data analysts/scientists
have more time for analytical thinking…

3. Overview of Evariant/Propensity Models
Evariant is on a mission to move healthcare providers to the
cloud with the data and analytics required to confidently
identify and execute on the most important strategic growth,
patient engagement, and physician alignment initiatives.
 The primary goal of Evariant’s predictive modeling is to
identify and target patients and non-patients who are likely
candidates for health services
 Patients and non-patients in healthcare markets have
differential levels of response propensity for different
disease-states and health screening programs
 Our predictive modeling is evolving with the healthcare
industry to not only capture traditional volume based targeted
marketing, but to also incorporate the rapid move to value
based marketing initiatives

4. Optimal modeling can incorporate volume and value based
targeted marketing initiatives.
The Healthcare industry in Transition
Why incorporate both volume and value based modeling and analytics? ?
Curve 1: Volume Based/Static
Example: Mammography Screenings
• Current state
• Standard for most healthcare marketers
• All about volume
• Little incentive for real integration
Curve 2: Value Based/Dynamic
Example: Traffic in Women’s Health Center
• Current + Future State
• Few healthcare marketers taking advantage
• Shared savings program
• Bundled/global payments
• Value-based reimbursement
• Rewards integration, quality, outcomes, and
efficiency

5. Types of Models
Patient Model - Which patients are likely to respond to a disease-specific marketing campaign (cross-sell, upsell,
retention)
Non-Patient Model - Which non-patients in the market are most likely to respond to a disease-specific marketing
campaigns (acquisition, re-acquisition)
 Certain individual patients and non-patients in a healthcare market have a higher likelihood of benefitting from
different health screening and treatment programs
 Multivariate statistical analyses (predictive scores) can optimize the precision in which these patients and non-
patients are identified and targeted for marketing purposes
 If your recipe for targeted marketing include traditional volume based approaches, limitations include only relying on
preselect criteria against “prospect lists” that include sociodemographic, lifestyle, response, transactional or other
elements
 Propensity models assign propensity scores to patients and non-patients that represent their likelihood to respond to
a given campaign, based on a broader set of predictive elements
We have a core set of approximately 130+ disease and health screening models available for your patient
population and consumers in your market

6. Evolving Solutions
Multivariate predictive modeling is employed along the Needs continuum, and incorporates
both volume and value based initiatives.

7. Sample: Model Debriefing Agenda
There are three components to the model
debriefing:
1. Overview of Modeling Processes
2. Overview of Tableau Visual Output
3. Overview of Dynamic List Builder (DLB)
 Q & A/Next Steps

8. Model Inputs/Parameters
Multivariate Comprehensive Datasets Include:
 Patient demographics
 Patient visit data history
 Appended Consumer Data
– Personal Information
o Lifestyle
o Sociodemographic/socioeconomic
o Health behavior
o Reported prescription data
– Household Information
o Ailments
o Family size/children
o Income/lifestyle variables (mortgage, dwelling size, location)
 Derived and proprietary variables such as behavior profile and comorbidity index

9. Cardiology Patient Model
Summary Statistics and Scoring Validation

10. Gender – Cardiology
71
158
0 20 40 60 80 100 120 140 160 180
female
male
Market Value Index
The market value index discriminates target group predictive characteristics from the prospect universe. Index values greater than 120 or less than
80 are indicators of statistically important over or under-value. A value of 100 indicated no over or under-value. For example, an overvalue index of
150 means that the prospects are 1.5x more likely than chance to statistically resemble the target group.

11. Marital Status – Cardiology
84
121
107
77
172
0 50 100 150 200
Other
Divorced
Married
Single
Widowed
Market Value Index
The market value index discriminates target group predictive characteristics from the prospect universe. Index values greater than 120 or less than
80 are indicators of statistically important over or under-value. A value of 100 indicated no over or under-value. For example, an overvalue index of
150 means that the prospects are 1.5x more likely than chance to statistically resemble the target group.

12. Age – Cardiology
32
41
59
87
109
146
192
262
0 50 100 150 200 250 300
youngest to 24 years
25-34 years
35-44 years
45-54 years
55-64 years
65-74 years
75-84 years
85 years and older
Market Value Index
The market value index discriminates target group predictive characteristics from the prospect universe. Index values greater than 120 or less than
80 are indicators of statistically important over or under-value. A value of 100 indicated no over or under-value. For example, an overvalue index of
150 means that the prospects are 1.5x more likely than chance to statistically resemble the target group.

13. Occupation – Cardiology
139
65
93
75
93
69
121
80
105
96
156
94
0 20 40 60 80 100 120 140 160 180
Blue Collar
Blue Collar Infer
Farm Related
Farm Related Infer
Other
Other Infer
Professional/Technical
Professional/Technical Infer
Retired
Retired Infer
Sales/Service
Sales/Service Infer
Market Value Index
The market value index discriminates target group predictive characteristics from the prospect universe. Index values greater than 120 or less than
80 are indicators of statistically important over or under-value. A value of 100 indicated no over or under-value. For example, an overvalue index of
150 means that the prospects are 1.5x more likely than chance to statistically resemble the target group.

14. Top Mosaics – Cardiology
121
122
128
128
129
131
132
133
138
147
159
166
176
187
193
193
0 50 100 150 200 250
Silver Sophisticates
Wired forsuccess
Fragile families
Birkenstocks and beemers
Small town shallow pockets
Aging in place
Homemade happiness
Footloose and family free
Gospel and Grits
Golf carts and gourmets
Reaping Rewards
Rural escape
True grit americans
Town elders
Senior Discounts
Settled and sensible
Market Value Index
The market value index discriminates target group predictive characteristics from the prospect universe. Index values greater than 120 or less than
80 are indicators of statistically important over or under-value. A value of 100 indicated no over or under-value. For example, an overvalue index of
150 means that the prospects are 1.5x more likely than chance to statistically resemble the target group.

15. Sample - Predictive Drivers – Cardiology Patient Model
 For the general cardiology patient model, the top three statistical drivers are comorbidity, age (older), and Mosaic groups (mostly
including older folks)
 Family history of cardiology related procedures, as well as ethnicity (higher risk for African-Americans and Latinos) are also strong
predictors
 Even though females make up a greater portion of the non-cardio population, males have a higher likelihood of being cardio patients
 Sales/service, professional/technical, and blue collar are the three occupational categories most predictive of having cardio services
Rank Variable Name Variable Description Direction
1 COMORBIDITY_MATCH_CLS * FLAG Comorbidity - Cardio
2 EX_EXACTAGE_CLS * FLAG Age Older individuals
3 MOSAICHOUSEHOLD_CLS * FLAG Mosaic More Mosaics that include older folks
Segment J Autumn Years
Segment Q Golden Year Guardians
Segment N Pastoral Pride
Segment C Booming With Confidence
Segment L Blue Sky Boomers
4 FMLY_HSTRY Family History Increased Family Hx
5 VST_ETHNICITYRACE * FLAG Ethnicity Increased for African American, Latino
6 EX_AWARNS_PRFL_CLS_ALL * FLAG Awareness of Health
HEALTHINSTITUTIONCONTRIBUTOR Higher donating behavior
MAILRESPONDER More multiple responders
FEMALEORIENTEDMAGAZINE More female-oriented magazines
BEHAVIORBANKINTERESTINREADING More general reading behavior
7 EX_OCCUPATIONMODEL_CLS * FLAG Occupation
Higher for Sales/Service,
Professional/Technical, Blue Collar
8 EX_GENDER_CLS * FLAG Gender
Males overpenetrated compared with
females

16. Sample - Predictive Drivers - Cardiology – Consumer Model
 For the general cardiology consumer model, the top three statistical drivers are age (older), Mosaic groups (mostly including older folks), and
socioeconomic variables
 Gender (higher for males) and general ailments (appended health flags, most related to cardio procedures) are also strong predictors
 Note that in the consumer model, without patient data, both the ailment conditions as well as the ailment medications are significant predictors
 Proactive health behaviors are negative predictors of cardio prospects
Rank Variable Name Variable Description Direction
1 EX_EXACTAGE_CLS Age Older individuals
2 MOSAICHOUSEHOLD_CLS Mosaic More Mosaics that include older folks
Segment J Autumn Years
Segment Q Golden Year Guardians
Segment N Pastoral Pride
Segment C Booming With Confidence
Segment L Blue Sky Boomers
3 EX_WEALTH_PRFL_CLS_ALL Economic Index
Median Home Value Lower and higher home values more predictive
Travel Travel behavior is a positive predictors
New Market Auto In the market for a new auto positive predictor
4 EX_GENDER_CLS Gender Males overpenetrated compared with females
5 EX_AILMENT_PRFL_CLS_ALL Dx Condition
Top 5 general appended ailments most predictive of cardio
patient status
Osteoarthritis
High Cholesterol
Heart Disease
High Blood Pressure
Sinuses/sinusitis
6 EX_BEHV_PRFL_CLS_ALL Proactive Health Behavior
Gardening, fitness, and outdoors interests are negative
predictors
GARDENINGFARMINGBUYER
INTERESTINFITNESS
INTERESTINTHEOUTDOORS
7 EX_AWARNS_PRFL_CLS_ALL Awareness of Health
HEALTHINSTITUTIONCONTRIBUTOR Higher donating behavior
MAILRESPONDER More multiple responders
FEMALEORIENTEDMAGAZINE More female-oriented magazines
INTERESTINREADING More general reading behavior
8 EX_MED_PRFL_CLS_ALL Medication Profile Increase in top medications related to cardio
9 EX_BUSINESSOWNER_CLS Business Owner Increase in risk for business owners

17. Sample: Model Performance and Testing
Sample of Relationship between Lift and “Best Patient Prospects” for Targeted Marketing Campaigns
 Once a final predictive model is created, a multivariate predictive score is produced. Each unique record in a given file is scored, then the
scores are broken into deciles.
 Decile 1 includes the “Best Patient Prospects” and should be targeted first. Prospects in Decile 1 have the highest probability of looking like
those in the Event Group having the behavior of interest (e.g., Cardiology Screening).
 Looking at the “Lift” Column in the Lift Calculation table, scored patient prospects in Decile 1 are 2.7x more likely (have greater than
chance probability) to look like an existing member of the Target Group (cardio targets).

18. Best Practices
Model Maintenance
• Models are updated regularly – new patients/non-patients added to database, run through model
and assigned a score/decile
• Models should be refreshed when there is a significant change in population parameters:
• Large number of people moved in/out
• Organization acquired/sold service location
Modeling Best Practices
• Evariant will review the need to refresh models
• Evariant will assist in synching marketing and modeling
calendars
• Models can be merged to maximize
campaign impact
• Consider testing + advanced reporting
• Built-in test-controls can be leveraged to assess
the efficacy of propensity models, including
refining when necessary

19. Using a Model for Targeted Marketing Campaign:
Breast Cancer Screening
Note: All patient and consumer IDs you have access to come from your own
facilities and markets.

20. MODELING IN THE
HEALTHCARE INDUSTRY:
A COLLABORATIVE APPROACH
William B. Disch, Ph.D.
Director, Analytics
Evariant
Thank you!
Q and A
O P E N
D A T A
S C I E N C E
C O N F E R E N C E
BOSTON 2015
@opendatasci

21. Q&A
Appendix