Agent-Based Modeling in Health Care · 2015. 10. 16. · Paolo Gaudiano, CTO, Icosystem Corporation Agent-Based Modeling in Health Care. 2 Joint Colloquium of the IACA, PBSS and IAAHS

JointJoint Colloquium of the IACA, PBSS and IAAHS SectionsColloquium of the IACA, PBSS and IAAHS Sections of the International Actuarial Associationof the International Actuarial Association

Westin Copley Place Hotel, Boston, U.S.A. –

4-7 May 2008

Paolo Gaudiano, CTO, Icosystem Corporation

Agent-Based Modeling in Health Care

2Joint Colloquium of the IACA, PBSS and IAAHS Sections

Westin Copley Place Hotel, Boston, U.S.A. – 4-7 May 2008

OverviewOverview

• Predictive modeling

• Agent-based modeling

• The Bean Machine

• Case Studies

• Closing remarks



Predictive Modeling in HealthcarePredictive Modeling in Healthcare



What is Predictive Modeling?

• What is a model?

• How are models used to make predictions?

Defining Defining ““Predictive ModelingPredictive Modeling””

Predictive modeling is the process of creating or selecting a model to predict the likelihood of an event or outcome.



What is a model?What is a model?



Is this a model?Is this a model?



It takes one to know oneIt takes one to know one



What is a What is a mathematical mathematical model?model?

A mathematical model is an abstract model that uses

mathematical language to describe the behavior of a system.

From: Wikipedia, the free encyclopedia



Types of mathematical modelsTypes of mathematical models

Mathematical models typically focus on a specific functional form or process, e.g.:

• Linear models

• Generalized linear models

• Hidden Markov models



Using mathematical models for predictionUsing mathematical models for prediction

Steps for mathematical modeling

1. Determine what model type to use

2. Determine what data to use

3. Adjust model parameters to fit the available data

4. Extrapolate to estimate future outcomes



An exampleAn example

DataModel

Prediction

Accurate fit to historical data does not guarantee predictive accuracy!

Accurate fit to historical data does not guarantee predictive accuracy!



Some (personal) issues with mathematical modelsSome (personal) issues with mathematical models

• The model itself becomes the focus: is it linear? is it exponential? is it logit?

• Unclear connection between statistical accuracy and “true model accuracy”, e.g., r2 =0.5

• The more complex the model, the more data are required to calibrate it

• Models are unable to predict the future in the presence of significant discontinuities (correlation vs. causation)

• Many statistical models were designed prior to the advent of the computer



IcosystemIcosystem’’s approach to predictions approach to prediction



A betterA better approachapproach

Develop world-class, custom computer models of complex systems

• Extract domain information about problem

• Identify “correct” elements for simulation

• Blend scientific rigor, experience and intuition

• Take advantage of the power of today’s computers to capture complexity instead of simplifying!



What is a What is a complex systemcomplex system??

A system consisting of many elements is complex if its overall

behavior emerges from the behavior of the individual elements

and their interactions.



Sample systems with emergent behaviorsSample systems with emergent behaviors

a traffic jam, a financial market, a crowd of people, a

sports event, a termite nest, a flock of birds,

…

a traffic jam, a financial market, a crowd of people, a

sports event, a termite nest, a flock of birds,

…



Emergent behavior: the Icosystem gameEmergent behavior: the Icosystem game



Computer simulation of the Icosystem gameComputer simulation of the Icosystem game

A is the aggressor, B the defender, try to keep B between you and A

A is the aggressor, you are the defender, keep yourself between A and B

See demo at http://www.icosystem.com/game.htm

http://www.icosystem.com/game.htm



Why is this interesting?Why is this interesting?

The Icosystem Game: a deceptively simple complex system

• Knowing the behavior of every individual is not enough to predict system behavior

• Slight changes in rules or interactions can lead to dramatic changes in system behavior

• The traditional reductionist (i.e., divide-and- conquer) does not work!



Controlling emergent behaviorControlling emergent behavior

• How can we control emergence?

• How do we define individual behaviors and interactions to produce desired emergent patterns?

“Here is where we think the

problem is...



AgentAgent--based modelingbased modeling

• Shift viewpoint from system (centralized) to individual elements (de- centralized)

• Each agent follows local rules

• Behavior depends on interactions with other agents

• Overall system behavior emerges from local interactions



AgentAgent--based modeling analogybased modeling analogy

• Simulate individual behaviors

• Capture key elements of agents

• Simulate interactions between agents

• Let the simulation unfold over time

• Look for patterns and trends



The Bean Machine:The Bean Machine:

Statistics vs. AgentStatistics vs. Agent--Based ModelingBased Modeling



The Bean MachineThe Bean Machine

• Developed by Sir Francis Galton to illustrate the Law of Error and the Normal Distribution

• The machine consists of a vertical board with interleaved rows of pegs

• Marbles are dropped from the top, and bounce randomly left and right as they hit the pegs

• Marbles collect into bins at the bottom

• The height of the marbles in the bins approximates a normal curve



The Icosystem Bean MachineThe Icosystem Bean Machine

• Pegs can be biased to the left or to the right

• The emergent distribution resulting from peg biases is observed at the bottom

• Two primary uses:

1. Given a certain set of biases, what will be the resulting distribution?

2. Given an observed distribution, what sort of biases might have given rise to it?



Can you Can you explainexplain the resulting distributions?the resulting distributions?

1 2 3 4

Compare to statistical approach: select a function, adjust parameters to fit distribution.

Compare to statistical approach: select a function, adjust parameters to fit distribution.



How does this help predictive modeling?How does this help predictive modeling?

• The bottom-up approach to modeling explains how behaviors lead to observed outcomes

• The resulting distribution emerges from behaviors, interactions and context

• The model can be used “in reverse” to try to determine behaviors that can explain the observed outcomes

• The model makes it possible to ask natural questions, e.g.: “how much benefit do we get from intervening earlier?” or “how much do I need to influence this stage to get an x% improvement?”



Sample Case StudiesSample Case Studies

PROPRIETARY29

Understanding supermarket shopper behaviorUnderstanding supermarket shopper behavior

Parking lot

Produce

Deli

RegistersEntrance

Fro

zen

foo

d

CS

D

entrance

targetClient: PepsiCo - a Fortune 500 consumer goods company

Challenge: Understand behavior of shoppers moving through a supermarket •How do they navigate?•What do they purchase?•Where to place products?

Complexities:•Correlate cart tracking data with consumer behavior •Predict behavior for novel supermarket configurations

Client: PepsiCo - a Fortune 500 consumer goods company

Challenge: Understand behavior of shoppers moving through a supermarket•How do they navigate?•What do they purchase?•Where to place products?

Complexities:•Correlate cart tracking data with consumer behavior•Predict behavior for novel supermarket configurations

PROPRIETARY30

Predicting Health Insurance enrollmentPredicting Health Insurance enrollment

Client: Humana, Inc. (a Fortune 500 Health Insurance Co.) Challenge:Increase prediction accuracy for health plan enrollment •What factors drive plan selection? •Impact of employer contribution?Complexities:•Many plan options across industries, locations •No historical data for new plans•Slow sales process

Icosystem’s predictive engine now rolled out to sales force, see

press release on Humana & Icosystem web sites

Client: Humana, Inc. (a Fortune 500 Health Insurance Co.)Challenge:Increase prediction accuracy for health plan enrollment•What factors drive plan selection?•Impact of employer contribution?Complexities:•Many plan options across industries, locations•No historical data for new plans•Slow sales process

Icosystem’s predictive engine now rolled out to sales force, see

press release on Humana & Icosystem web sites

RiskRisk

CostsCosts BenefitsBenefits

RiskRisk


RiskRisk


RiskRisk


RiskRisk


RiskRisk


RiskRisk


RiskRisk


RiskRisk


RiskRisk


Select PlanSelect Plan

OfferingsOfferings

PROPRIETARY31

The Impact of social nets on product adoptionThe Impact of social nets on product adoption

Client: Fortune 500 drug company puzzled by variability of drug adoption across clinics.

Challenge: Why do some clinics prescribe this emergency room drug much more readily than others, even with similar marketing efforts?

Outcome: Identified different clinic types based on social interactions due to scheduling and patient volume. Clinic social networks correlate well with required marketing effort.

Client: Fortune 500 drug company puzzled by variability of drug adoption across clinics.

Challenge: Why do some clinics prescribe this emergency room drug much more readily than others, even with similar marketing efforts?

Outcome: Identified different clinic types based on social interactions due to scheduling and patient volume. Clinic social networks correlate well with required marketing effort.

Fully connected:

Rapid spread of new ideas.

Sparsely connected:

Repeated efforts required for

spread.

Mostly disconnected:

Each party needs individual attention.

PROPRIETARY32

Medicare Product Launch and DesignMedicare Product Launch and Design

Client: Humana, Senior Products Division

Challenge: Plan marketing strategy for new Medicare products in 2006

Outcome:

• Helped planning for new product launch - Humana now #2 in Medicare space

• Extended the same tool to assist with product design

• Used tool to identify operational issues

Client: Humana, Senior Products Division

Challenge: Plan marketing strategy for new Medicare products in 2006

Outcome:

• Helped planning for new product launch - Humana now #2 in Medicare space

• Extended the same tool to assist with product design

• Used tool to identify operational issues

PROPRIETARY33

Treatment ComplianceTreatment Compliance

Client: Fortune 100 Pharma with an antiviral drug, confused as to why so few patients seek treatment.

Challenge: What makes patients enter a complex treatment regime and comply with it over a long period of time?

Outcome: Discovered unexpected pockets of opportunity in diagnostic procedures and challenged client assumptions about leverage points.

Client: Fortune 100 Pharma with an antiviral drug, confused as to why so few patients seek treatment.

Challenge: What makes patients enter a complex treatment regime and comply with it over a long period of time?

Outcome: Discovered unexpected pockets of opportunity in diagnostic procedures and challenged client assumptions about leverage points.

Prioritization Matrix from Customer ReportPrioritization Matrix from Customer Report

Ability to WinAbility to Win

Size

of O

ppor

tuni

tySi

ze o

f Opp

ortu

nity

22

HighHigh

MedMed

LowLow

11

3*3*

7/87/8

1010

1111

99

1212

1313

11 22 33 44 55

6a6a

--

1414

6b6b

11

223*3* 6a6a6b6b

7/87/8

991111 1212

13131414

1010

Icosystem Results MatrixIcosystem Results Matrix10001000

100100

11.0001.0001 .001.001 .01.01

11

22

6a6a6b6b

7/87/899

111112121414

1010

1010

SensitivitySensitivityPo

tent

ial (

% in

crea

se o

f pat

ient

s)Po

tent

ial (

% in

crea

se o

f pat

ient

s)

11, 12, 1311, 12, 13

3*3*

PROPRIETARY34

Drug development portfolio processDrug development portfolio process

Client: Fortune 100 Pharma

Challenge: Redesign drug development process to reduce complexity, costs and time while improving success.

Approach: Detailed simulation of decision-making and processes and visualization to identify sources of inefficiencies and design/test a more flexible networked, portfolio-centric clinical development organization.

Outcome: New approach tested on small portfolio, recently (‘05) expanded to broader portfolio

Client: Fortune 100 Pharma

Challenge: Redesign drug development process to reduce complexity, costs and time while improving success.

Approach: Detailed simulation of decision-making and processes and visualization to identify sources of inefficiencies and design/test a more flexible networked, portfolio-centric clinical development organization.

Outcome: New approach tested on small portfolio, recently (‘05) expanded to broader portfolio

Asset Asset Group EstimateNPLC

Approval

Implement Work

Interpretable Data

Uninterpretable Data

Yes

NoReplan

DMS Approval Yes

NoReplan

TOXADME Approval Yes

NoReplan

SAFE Approval Yes

NoReplan

TASC/BPC Approval Yes

NoReplan

PCAT II Approval Yes

NoReplan

DDOC Approval Yes

NoReplan

Complex Governance

Inefficient Resourcing

Difficult Planning

Networked Organization

ProblemSet

Definition

ScenarioGeneration Module

Definition

PortfolioScheduling

ModuleImplement

NewData

Analysis

ResourceContract



Detailed CaseDetailed Case StudiesStudies



AgentAgent--based modeling for healthcarebased modeling for healthcare

Two examples of healthcare applications:

• How physician scheduling in hospitals impacts prescribing behavior

• The impact of social interactions on the perception of healthcare products for seniors



Sample application #1 Sample application #1

How physician scheduling in hospitals How physician scheduling in hospitals impacts prescribing behaviorimpacts prescribing behavior



Problem statementProblem statement

• Large Pharma company observed different prescription rate for ER drug at different hospitals.

• Sales/promotional activities were uniform.

• What drives different adoption rates at different hospitals?



A simplistic view of the prescribing dynamicA simplistic view of the prescribing dynamic

message

frequency

Sales rep

opinion

Physician

compliance

Patient



The reality of drugThe reality of drug--based healthcare is complexbased healthcare is complex

message

frequency

Sales rep

Other patients Care

givers

message

frequency

Sales rep

opinion

Physician

compliance

Patient

substitution

Pharmacy

Physician thought leaders

Physician colleagues

Other medical

colleagues

Competing sales reps

formulary

Health insurer

Treatment support teams

Web detailing



Top prescribing clinics, can be

classified by social network type,

through analysis of observable data.

Once classified, the network can be

modeled to predict the response to

proposed marketing efforts.

Social networks to classify clinic typesSocial networks to classify clinic types



These types of clinics respond differently to marketing activities!

Fully connected: Rapid spread of

new ideas.

Sparsely connected: Repeated efforts

required for spread.

Mostly disconnected: Each party needs

individual attention.

Three general types of clinics observedThree general types of clinics observed



Drivers of social network structure

Shift structures for staff.

Patient volume.

Observability of patient benefit.

Numbers of attending physicians

Socializing opportunities.

Physical layout of building.

VERY STRONG ++++

STRONG +++

MODERATE ++

MODERATE ++

WEAK +

STRONG +++

Social network Social network ““driversdrivers”” can be identifiedcan be identified



Sample application #2 Sample application #2

The impact of social interactions on The impact of social interactions on the perception of healthcare products the perception of healthcare products

for seniorsfor seniors



Predictive modeling for MMA Part DPredictive modeling for MMA Part D

Problem statement:

• The implementation of MMA Part D will cause a profound discontinuity in the Medicare landscape.

• The lack of precedent leads to uncertainty.

• Market research can help -- but how?• What existing data can be used to gain insights?

• What data should be collected after implementation?

• How should data be used: Marketing campaign? Product design? Competitive strategy?



In reality: many events affect senior opinionIn reality: many events affect senior opinion

Go to seminar Go to

seminar

Contact service & guidance

professionals

Contact service & guidance

professionals

Visit doctorVisit doctor

Evaluate Plan Benefits

Evaluate Plan Benefits Interact

with peers Interact

with peers

Fill prescription

Fill prescription

View information

View information

Go to hospitalGo to hospital

Talk with family members

Talk with family members



The model: highThe model: high--level summarylevel summary

HMO

FFSPPO

Plans

Selection

Service & Guidance Professionals

SocialInteractions

Marketing

Providers



1. SimSenior: key factors1. SimSenior: key factors

Personal factors:

• Age

• Income

• Marital status

• Social network

• Physician

Healthcare factors:

• Medical condition

• Current coverage

• Medical expenses

• Prescription expenses

Subjective factors:

• Brand perception

• Sensitivity to plan benefits

• Inertia



Marketing2. The influence of information2. The influence of information

• Brand ads affect brand perception

• Product ads affect the senior’s sensitivity to specific plan benefits

• Seminars affect• Brand perception• Perceived importance of the plan benefits highlighted in

seminar

• Seminars also foster social interactions

• Competitor and government information exerts additional influence on senior choice process



3. The influence of social interactions3. The influence of social interactions

• When SimSeniors exchange information, they influence each other’s brand and plan perception.

Low brand perception

High brand perception

Raised brand perception

Social Interaction



4. The influence of customer experience4. The influence of customer experience

Ability to Provide Good Customer Experience

0 1Distributions

0

0.005

0.01

0.015

0.02

0.025

0.03

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-0.1 0.10

Bad

Med

Good

Probabilistically sample Experience = change in Brand Perception

Service & Guidance Professionals



Simulation tool: screenshotsSimulation tool: screenshots

Main screen



Result Visualization: Baseline exampleResult Visualization: Baseline example



When drug cost benefits are important When drug cost benefits are important ……

… HMOs are most popular (when available).

Reason: HMOs provide best drug coverage at a very low premiumReason: HMOs provide best drug coverage at a very low premium



When freedom of choice is important When freedom of choice is important ……

… HMOs capture only 25% of market share.

x

xx

x

xx



SummarySummary



Emergent behaviorsEmergent behaviors

Incomplete understanding of interactions can lead to surprising results...

... but predictive modeling can help!



AgentAgent--based simulations for healthcarebased simulations for healthcare

• Not all models are created equal

• Agent-based simulations offer a powerful tool when:• Significant, complex interactions• Insufficient data• Highly dynamic environment• Discontinuities

• Important take-home messages:• Understand the limitations of your model• Find (or build) the right model for problems

Joint Colloquium of the IACA, PBSS and IAAHS Sections�of the International Actuarial Association�Westin Copley Place Hotel, Boston, U.S.A. – 4-7 May 2008OverviewPredictive Modeling in HealthcareDefining “Predictive Modeling”What is a model?Is this a model?It takes one to know oneWhat is a mathematical model?Types of mathematical modelsUsing mathematical models for predictionAn exampleSome (personal) issues with mathematical modelsIcosystem’s approach to predictionA better approachWhat is a complex system?Sample systems with emergent behaviorsEmergent behavior: the Icosystem gameComputer simulation of the Icosystem gameWhy is this interesting?Controlling emergent behaviorAgent-based modelingAgent-based modeling analogyThe Bean Machine:�Statistics vs. Agent-Based ModelingThe Bean MachineThe Icosystem Bean MachineCan you explain the resulting distributions?How does this help predictive modeling?Sample Case StudiesUnderstanding supermarket shopper behaviorPredicting Health Insurance enrollmentThe Impact of social nets on product adoptionMedicare Product Launch and DesignTreatment ComplianceDrug development portfolio processDetailed Case StudiesAgent-based modeling for healthcareSample application #1 Problem statementA simplistic view of the prescribing dynamicThe reality of drug-based healthcare is complexSocial networks to classify clinic typesThree general types of clinics observedSocial network “drivers” can be identifiedSample application #2 Predictive modeling for MMA Part DIn reality: many events affect senior opinionThe model: high-level summary1. SimSenior: key factors2. The influence of information3. The influence of social interactions4. The influence of customer experienceSimulation tool: screenshotsResult Visualization: Baseline exampleWhen drug cost benefits are important …When freedom of choice is important …SummaryEmergent behaviorsAgent-based simulations for healthcare

Documents

Agent-Based Modeling in Health Care · 2015. 10. 16. · Paolo Gaudiano, CTO, Icosystem Corporation Agent-Based Modeling in Health Care. 2 Joint Colloquium of the IACA, PBSS and IAAHS