Doctors "Versus" Predictive Modeling To Identify Patents At Risk: It's Really Docs AND Predictive Modeling.

Who can find patients for a DM program

One of the major advances in disease and population-based programs for chronic illness has been the discovery that not all persons with a chronic condition will benefit from care management.  Instead of blanketing everyone with education programs, the "secret sauce" is targeting just those patients who have the twin attributes of:

1) being at high future risk of a bad outcome (such as an avoidable hospitalizations or multiple emergency room visits) and

2) having modifiable risk, which is ultimately a function of having the "right" disease severity, receptivity to engagement in self care and social supports which, in turn, respond to outreach.

In their American Journal of Managed Care article, "Identification of Patients Likely to Benefit from Care Management Programs," Tobias Freund et al use the terms "cost intensive" and "care sensitive" to describe the twin attributes above. They set out to examine how they compared in two candidate populations for disease management programs: 1) one that was chosen by a "predictive modeling" program that uses mathematical algorithms to weigh demographic and insurance claims inputs to compute future risk and 2) one that was made up of patients chosen by their physicians.

If this sounds like a TV Jeopardy contest between IBM's "Watson" versus human brainiacs, it shouldn't.  In this study, there was no "gold standard" that could be used to gauge whether computers or docs were better at picking the patients at risk.  Instead, the authors merely compared the two populations.

This study was conducted in Germany, using one of the public and private funded "sickness funds'" claims and demographic data that provided insurance to a population of 6026 patients that were being cared for in 10 southwestern primary care practices.  The predictive model was "Case Smart Suite Germany," which selected the top 90th percentile of patients at risk using data from 2007 and 2008,  while the primary care physicians were simply asked to screen "all" of the sickness fund patients from 2009 and "select up to 30" patients that may benefit from a program aimed at helping them stay out of the hospital.

What the DMCB found interesting about this study was that the predictive model identified 301 patients, while the docs identified 203.  Only 32 patients were selected by both.  The PCP patients were generally younger (mean age of 66 years vs. 75 years), had lower rates of prior hospitalizations in 2007 and 2008 (.53 vs. 2.74), had a lower degree of co-morbid illness (a Charlson Index of 2.1 vs. 4.3) and generally lower rates of chronic conditions like diabetes, COPD and depression (mean number of conditions was 3.0 vs. 4.5).  Last but not least, the physicians were better at picking the "care sensitive" patients that had been previously entered into a previously existing disease management.

The main lesson here?  It's not predictive modeling "versus" physician identification, or trying to decide whether to target a care program based on what your IT Department is saying "or" what the docs are saying, it's both.  The physicians, thanks to their "proximity" to their patients are using yet-to-be-identified heuristics to spot patients that predictive modeling may miss.  

While this study is limited by being outside the U.S. (making it less generalizable) and restricted to a relatively small population, the DMCB thinks the implications are intriguing enough to possibly warrant a repeat study here in the States.  In addition, the disease and population care management community can use this approach and eventually compare how each selected population responds on a prospective and outcomes basis.