CONFIDENTIAL & PROPRIETARY Expanding the Reach of Predictive Models: Using Clinical, HRA, and Consumer Data Dan Dunn, PhD, Senior VP of R&D, Ingenix The.

CONFIDENTIAL & PROPRIETARY

Expanding the Reach of Predictive Models: Using Clinical, HRA, and Consumer Data

Dan Dunn, PhD, Senior VP of R&D, Ingenix

The National Predictive Modeling Summit December 13, 2007 ● Washington, DC

© Ingenix, Inc. 2CONFIDENTIAL & PROPRIETARY

Agenda

Context for Innovation New Sources of Data and Changing the Focus of

Measurement – a Conceptual Model Using Alternative Data Sources in Risk Modeling


Context for Innovation

Information tools to support care and health management – current state: Primary focus is on disease populations or individuals of

moderate to higher risk Clinical information and concepts supported by administrative

medical and pharmacy claims, some use clinical data Outputs include measures of risk, some add gaps in care Many tools add reporting and some cohort modeling

capabilities Limited use of alternative sources of data


Context for Innovation

Increasing interest in focusing on healthier members in a population, or members of emerging risk Extend interventions to the lower end of the risk spectrum Improve wellness, healthy behaviors and lifestyle Improve attitudes on health Intervene “upstream” in a more pro-active way, e.g., pre-

diabetes, and “pre-pre”-diabetes Interest in creating a personal health record (PHR)

Integrates information from a number of data sources to provide a multi-dimensional profile of an individual’s health

Support interventions in a more complete way – from “end-to-end”


Changing focus for information solutions

0

Increasing demand for information solutions that support interventions for relatively healthier members or those of emerging risk

Members without medical or pharmacy claims

Members of emerging clinical risk

- pre-diabetic- onset of chronic

condition

Higher PM Risk- higher cost conditions- multiple co-morbidities- recent acute events

Moderate PM Risk- chronic conditions- some co-morbidities- recent history, stable

Lower PM Risk- smokers,

- sleep problems,- obese, inactive

“Sweet spot” for current state of predictive modeling (PM) is patients of moderate to higher risk – supporting more traditional disease and care management


Identification and StratificationIdentification and Stratification

Intervention and Management

Medical ClaimsRx Claims

DemographicsClinical Data

HRAsConsumer Data

Support “end-to-end” intervention solutions

SegmentationSegmentation

ActivationActivation

Risk PredictionClinical Profile

Health Behaviors

Match Patients to Programs

Support for Engagement and Intervention


Change of Focus and Requirements

Support analysis of healthier populations and emerging patients Leverage existing and new sources of data, including HRA/self

report and consumer information Integrate these different sources of data in innovative ways:

Improve on existing concepts, e.g., measures of future risk Support new domains of measurement, including behaviors, attitudes,

and social context Accommodate different data scenarios – consistent data

availability unlikely across and within populations Create a useful context for analysis

We are pulling together even a larger number of concepts and variables

Add value by developing a context – organize information for analysis, presentation, and operations – in a flexible way


New Information and Domains – Opportunities Address disease and lifestyle risk

Whole-person approach to health management – across the full continuum of health and risk

Complement and expand opportunities to address further domains of health that they may not be concentrating on

Expand models of clinical, risk and cost with the addition of new dimensions and sources of data Prediction based on a set of new concepts Bring behavior and attitudes to the equation Bring social and consumer variables to bear on risk

Tailor interventions based on a central repository of data that has key variables associated with outreach, intervention and outcome

Support a Personal Health Record – informed by multiple sources of data, describing key dimensions of health


Using New Sources of Data and Changing the Focus in Measurement:

A Conceptual Design


What model of health can be used to structure a more complete approach?

Wilson Cleary model (1995) of HRQOL is helpful because it represents a full picture of health

A model of health


Operational Model of Health: Concepts and Domains

A more complete approach requires methods and outputs to measure individuals along the different domains that describe health

Domains that support identification/stratification, segmentation, and activation

Intervention Groups – a context for integrating the five domains

Note – prediction and “risk” are only one component

Risk and Severity

Intervention Groups

Clinical

Health Behaviors

Social Context

Health Attitudes


Health Model Concepts and Domains

Information/domains to support identification and stratification: Clinical

– A clinical description of an individual, based on diagnostic and procedural concepts – from claims, clinical results and self report

– Examples – diabetes, pre-diabetes, CHF, depression, sleep disorder, obesity, propensity for a clinical condition

Risk or Severity– Predictive model risk, condition severity, self-report health status– Examples – relative risk, condition episode severity, health status

Behavior (Healthy behaviors)– HRA and claims-based measures of behavior, behaviors inferred

from consumer data– Examples – smoking, physical activity, compliance with chronic

and preventive quality rules (gaps in care), prescription adherence


Health Model Concepts and Domains

Further segmentation and activation can be supported by: Attitudes about Health

– Readiness to change, activation and perceived social support Social Context (Social Score)

– Ascribed and achieved status, plus consumer-oriented variables– Examples – Age, gender, race ethnicity, education, income, SES


What will Intervention Groups do? Provide a context to organize and focus information – in a way

that is consistent from both a clinical perspective and also from an operational perspective

Describe both clinical and wellness concepts – e.g., diabetes, smoking, sleep disorder

Have defined levels – that map to potential cohorts for intervention – e.g., level of acuity; categories of smoking status; level of physical activity

Have rules and algorithms that assign an individual to an Intervention Group – and further to a level

Incorporate methods to accommodate different data availability scenarios for each individual


Examples of Intervention Groups

Disease Management Wellness

Asthma/COPD Smoking/Tobacco

CAD Physical Activity

CHF Nutrition

Diabetes Safety

Back, Joint, Surgical Option ProblemsStress

Mental Health (Depression) Safety

Obesity Alcohol Abuse

Sleep Problems Sexual Risk Activity

Pain Syndromes


Diabetes Intervention Group Levels

Severe Diabetes Moderate Diabetes Mild Diabetes Pre-Diabetes “Pre-Pre-Diabetes” No Diabetes

Information used to identify and stratify Medical and Rx: diagnoses, drug therapies Predictive model risk HRA and consumer: self-report, obesity, behaviors consistent with

propensity for diabetes Map relevant clinical and family history to further define levels

Ask ourselves: If I run a diabetes management program, what would I want to understand about my members? Severity of diabetes, propensity Associated health behaviors, co-morbid conditions and attitudes What factors are associated with engaging members?


Sleep Problems Intervention Group Levels

Severe sleep problems Moderate sleep problems Mild sleep problems No sleep problems

Information used to identify and stratify Medical and pharmacy: diagnoses, drug therapies for treatment,

diagnostic tests Predictive model risk HRA: self-report, sleep problem questions, medication self report

Ask ourselves: If I run a program for sleep problems what would I want to understand about my members? Severity of sleep problem Other behaviors, conditions and attitudes associated What factors are associated with engaging members?


Outputs

Summary and detail results for an individual along each of the domains

Information, reports and views centered around the concept of an Intervention Group with links between a patient, a group, their level detailed information supporting:

– Intervention Group assignment– appropriate segmentation– activation for intervention– the intervention itself

Risk scores and other summary measures


New Data Sources and Domains: Challenges

Consistency in the availability of information across individuals – most will have claims some will have HRAs and/or consumer data clinical lab results may be available timeliness of the information

Opportunities for risk models – leveraging different types of information

Creating a flexible context for using this information – it translates in different ways depending on the appropriate focus for a patient and the domains


Using New Sources of Data in Risk Modeling


Measuring Health Risk – Overview

· Demographics· Medical Claims· Rx Claims· HRA· Lab Results· Consumer

Disease Prevalence, Co-

Morbidities, Complications

Condition-Based Risk Markers

Service-Based Risk Markers

Member Clinical Profiles

Weighting of Profile to Compute

Risk

Complete Member Risk Profile

Grouping of Inputs to support Disease Identification and

Disease Severity (e.g., Episodes of Care)

Grouping of Diseases and Conditions into

Clinically Homogeneous Risk Marker Categories

- High Acuity Events - Moderate/Lower

Risk Markers- Rx Markers

Array Markers for each Member to

Create a Clinical Risk Profile

Apply Weights Measuring

Contribution of each Marker to Overall

Risk

Combine Profile and Risk Results to

Complete Member Profile

Outputs

Markers of Risk

Translating Markers into Risk Measures

Data Inputs


New sources of data in risk modeling*(*in addition to administrative claims and enrollment)

HRA surveys What it adds

– Clinical indicators – e.g., self report of a condition not observed in claims

– Overall assessment of health status– Behaviors that indicate propensity for a higher risk clinical

condition Modeling approach

– New indications for disease risk markers– Propensity-based markers of risk – e.g., likelihood of diabetes– Behaviors, other – smoking, obesity– Estimate risk weights for new markers – use to adjust risk score

Challenges– Data availability and timeliness– Reconciling conflicting information



Clinical lab results What it adds

– Condition severity – e.g., organ function tests and cancer tumor/stage diagnostics

– Trends in levels Modeling approach

– Add lab-result based risk markers to a model– Estimate risk weights for new markers – use to adjust risk score

Challenges– Data availability– Timing– Benefits a relatively small percentage of population – although

impact can be significant for these patients


Lab Results and PredictionAdded risk indicated by lab result markedly outside of normal range

Lab performed in last 90 days. Comparison of predicted (Impact Pro without Lab Model) and actual PMPM and relationship of prediction error with lab results ranges (“Difference”). Only most extreme lab result findings included on slide.

-2000

200400600800

1,0001,2001,4001,6001,800

Pred

ictio

n D

iffer

ence

($

PM

PM)

Albumin ALP CRP Chol Ratio CA-125 HbA1c

Using lab results in risk modeling



Consumer data What it adds

– Social Context – income, education– Consumer habits – purchases, auto registration– Categories – groupings of individuals to

Modeling approach– Categories and derived variables– Test risk weights for new markers – use to adjust risk score?

Challenges– Data availability– Timing– TBD on general contribution to predictive accuracy on top of

claims – likely most helpful for lower risk


Summary

Information tools to support care and health management – current state: Primary focus: disease populations, moderate to higher risk Limited use of alternative sources of data Mostly support ID & stratification

Use of alternative data sources both provides new opportunities and requires a new conceptual idea about “predictive modeling” More complete view of the patient Supporting the full cycle of care and health management, including

segmentation, activation and the intervention itself Focus on healthier individuals and wellness programs is not best

supported by a risk “score” – but by a multi-domain description of that individual

Challenges – consistent availability of data and creation of a context that supports operational realities


Questions/Comments

CONFIDENTIAL & PROPRIETARY Expanding the Reach of Predictive Models: Using Clinical, HRA, and Consumer Data Dan Dunn, PhD, Senior VP of R&D, Ingenix The.

Documents