Medicare Hospital Readmissions: Bad. Our Ability To Understand or Do Much About Them: Worse

"I  think I need to go back
to the hospital....."

The Disease Management Care Blog remembers her very well. She had come to the United States after escaping on foot from post-war Hungary.  She liked to have sliced pears and cognac for lunch. As her health began to fail, her stubbornness began to exceed her self-sufficiency, leading to multiple hospitalizations. One time, despite making sure she had a post-discharge outpatient appointment, a fresh pile of prescriptions and home visiting nurse services, her illness flared and she had to come back to hospital.

She was not only my patient, but she represented one of Medicare's dreaded readmission statistics.

By now, DMCB readers know that CMS, buoyed by its value-based purchasing program, has targeted readmissions by reducing payment levels to hospitals that fall outside the expected norm. Threatened by the loss of income, it’s assumed that hospitals will respond by developing higher quality discharge planning and care programs that keep patients from having to come back.

An important part of reducing readmissions is to identify those patients that are at greatest risk. That would help on two levels:

1) if a hospital had more than its fair share of patients at risk, it could argue that an increased number of readmissions is the result of a sicker patient population and not quality of care. As a result, the hospital could be held "harmless;

2) hospitals would be able to focus extra care resources on those patients who are spotted early as likely to come back, thereby reducing the readmission rate.

In other words, patients like the lady from Hungary would not necessarily lead to a cut in hospital payment rates and, for example, she could be proactively given extra care, such as a doctor appointment within 48 hours, a week’s supply of free medications and twice a day home nurse visits.

Which is why this just published JAMA article "Risk Prediction Models for Hospital Readmission" by Devan Kansagara, Honora Englander, Amanda Salanitro David Kagen, Cecelia Theobald, Michele Freeman and Sunil Kripalani is important.  The authors set out to see what the evidence-based published scientific literature had to say about predicting readmissions.   They filtered thousands of references, reviewed 286 publications found 30 rigorous studies that described 26 models. 

To the DMCB's delight, the authors applied a “c statistic” to the 30 publications to assess a wide variety of retrospective and concurrent prediction methodologies using a host of data inputs such as age, gender and past diagnoses. According to this article, the c (or "concordance") statistic measures how well a test can predict the presence or absence of a "condition" which, in this case, was being readmitted to the hospital.  One way to think of this is the likelihood of correctly identifying a condition when there are two people, one with it and one without it.  If the likelihood is 50%, that's no better than random guessing.  If it's 100%, that's perfect.  By the way,if this sounds a lot like the area under the receiver operator curve, you're right.

And what did the all-seeing "c statistic" say?  All of the published models had disappointingly similar levels of performance that ranged between the extremes of .52 to .83 with most in the .50 to .7 range. What's more, only one study examined the most important question of all: is it possible to find patients with preventable readmissions?

What does the DMCB think?

1) This may be another area where national health policy has gotten out in front of the scientific evidence.  If we can't reliably assess or predict readmissions with sufficient accuracy, there is a distinct likelihood that statistical variation, not quality of care, will lead to some hospitals being victimized by CMS with lower payment rates. What's more, if hospitals can't tell which patients are likely to come back, how are they supposed to target their expensive care management programs at those who are most likely to benefit?

2) There are undoubtedly some proprietary predictive models that haven't been reported in the literature that claim to have higher levels of accuracy.  Yet, without the scrutiny of successful peer-reviewed publication, it'd be difficult to believe that they're really any better than the mainstream published range of .5 to .7.  The next time the DMCB runs into one of these outfits, it's going to ask about the "c statistic" and if they haven't published their results, why not.

3) Last but not least, while the hospital payment rates are being held hostage by CMS, it's the doctors that are making the call on readmissions based on the best interest of their patient.  The c statistic suggests that that will be the most important determinant in the readmission rate.