Research Abstract summarizing the impact of using One Care Street tm in a Fortune
100 company
Prospective Targeting of Potential High-Utilizers for Health Coaching:
Results of a Pilot Study of Pre-Medicare Retirees
Abstract
Objectives: (1) To determine the accuracy of a predictive model, using self-reported
health survey data, to prospectively identify high near-term utilizers of health
care. (2) To examine the effect of health coaching of these high utilizing individuals
on their subsequent health care claims at 6 and 12 months.
Design: Pre/post prospective cohort study using propensity scores to adjust for
biases inherent in high risk patients’ agreeing to participate in One Care StreetTM
(OCS) and its health coaching intervention.
Setting: In spring 2002, a Fortune 100 employer with a large national managed care
organization funded The Haelan Group (Haelan) to perform a pilot study assessing
the potential for improved health and lowered cost through the use of OCS. OCS is
a population health management system, that utilizes proprietary predictive modeling
(based on demographic data and personal health perceptions and behaviors) and highly
individualized, proactive health coaching for individuals at high risk for generating
higher costs. The pilot study included 5,127 pre-Medicare retirees, then insured
through the employer with traditional administrative and care coordination services
provided by the health plan. The study population became eligible to take the OCS
survey in 7 deployment groups from June through December of 2002 with subsequent
enrollment of high-risk individuals into the health-coaching program as they were
identified. A division of the health plan provided the health coaches (all registered
nurses) for the pilot program. The coaches utilized the OCS coaching models and
software system. The final cohorts of continuously-enrolled individuals for whom
both OCS and claims data were available included 1,692 who had cost data available
for the 12 months before and 6 months after OCS screening, and 1,556 who had full
claims data for 12 months before and after screening.
Measures: All eligible subjects were asked to complete the OCS survey, a 45-item
health perception assessment either online or on paper as each deployment group
was notified between June and December of 2002. Claims data were subsequently obtained
from the health plan and pharmacy-benefit manager for the entire 24 month study
period. The final study cohort was determined by including all those with both OCS
and continuous claims data during the 12 months prior to each person’s participation
in OCS and the 12 months after (determined as each person’s “Index Date”). All claims
dollars were indexed to the last month of each subject’s 24-month study period to
control for inflation. Therefore, regardless of the person’s Index Date, data could
be aggregated across all participants into the 12-month pre- and post-periods. Outliers,
defined as those individuals with greater than $50,000 in either the pre- (N=12)
or post-periods (N=17) were left in for descriptive analyses but were excluded from
the analysis of the intervention’s effect.
Analysis: To determine the accuracy of the OCS predictive model in identifying high-care
users, the predicted status of each individual (high or low-risk for being a high
utilizer as determined by OCS risk score) was compared to their observed status
(high = 20% highest costs; low = 80% lowest costs) at 6 months post-screening among
the 1,692 individuals with 6-months post-screening claims. Additional analyses (which
will be reported separately) compared the additional predictive power gained by
combining OCS’ variables and the health plan’s claims-based predictive model. All
tentative models then underwent a business analysis to determine the model which
best optimizes highest accuracy at the lowest identification cost going forward.
To determine the intervention effect, we first established an equivalent control
group of subjects determined by OCS to be high-risk but were not coached. To accomplish
this, we developed a propensity score of those variables that discriminated between
subjects coached and not coached. The subsequent intervention effect was then calculated
by determining the difference in costs between the pre-12 months and the post-12
months total costs between the treatment and control groups, adjusting for treatment
bias via the propensity score. Further, we determined the effect of the number of
coaching sessions on treatment effect by comparing subjects who received 1-3 coaching
sessions with those who received 4 or more sessions.
Results: Predictive accuracy:
Though the original OCS predictive model (at a 0.5
risk score threshold with 77% sensitivity) added differential value to the health
plan’s predictive model, the false positive rate was unacceptably high at 58%. A
re-derived OCS model for this particular age group (i.e., re-weighting the variables
for this subject population) yielded a sensitivity of 57% and specificity of 79%,
thus lowering the false positive rate to more acceptable levels and still identifying
more than half of the subjects who were in the top 20% group of post-screening costs.
Intervention effect: Analyzing the 1,556 who had the full 12 months of pre- and
post-claims data, 612 (39%) were in the OCS low-risk group and 944 (61%) were in
the OCS high-risk group. Among high-risk subjects, 556 (59%) were coached. The most
common reasons for not coaching subjects were the inability to contact them, subject
refusal, subject’s issues were resolved, and subject already felt he/she had adequate
resources for dealing with his/her health care issues. Of the 556 who were coached,
64 (12%) had 4+ coaching sessions while 492 (88%) had 1-3 sessions. To better control
for the regression-to-the-mean bias inherent in a pre/post design, an equivalent
control group (adjusted for treatment bias using a propensity score) was established
to determine the effect of the coaching intervention on pre- to post-12 month claims
cost. The propensity score (derived using all OCS, demographic, and claims data
as potential entering variables) had an acceptable c-statistic (ROC curve area)
of 0.74. Using this control group, the coached high-risk group experienced an average
$371 greater decrease in total costs comparing pre- to the post-12 month periods
than the non-coached high-risk group. Those who received 1-3 sessions averaged a
cost decease of $300 while those who received 4+ sessions had an average cost decrease
of $2,251.
Conclusions: Self-reported factors improve the ability to predict an adult’s probability
of becoming a near-term high care utilizer. Once identified, the cost of care for
these high-care users can be positively impacted through an individualized coaching
intervention. Going forward, effects can be enhanced by improving response rate,
using the re-derived OCS model, engaging a higher percentage of high-risk into 4+
coaching sessions.
Respectfully submitted,
William M. Tierney, M.D.
Chancellor’s Professor of Medicine and Director,
Division of General Internal Medicine and Geriatrics
Indiana University School of Medicine
Senior Research Scientist, Regenstrief Institute, Incorporated
|