RE: control charts in health care

[jmsiegel@sbcglobal.net]

I agree with Mike Woolbert and Paul Hollingworth's recent comments
regarding the issue of how to control chart the length of an ambulance
ride. I spent a couple of years in the mid-90s doing control charts in
health care, and I found what may be a similar situation to the
ambulance one.

I found that in the situations I was looking at, people often seemed to
be using the same category to describe, and attempting to chart on the
same chart, very different phenomena whose variation they may have had
little ability to control. We charted things like home health visits of
various types per patient by diagnosis. Some diagnoses generally
resulted in a handful of home care visits but for some patienst there
were dozens, or even hundreds or thousands. We took a further look and
found that because diagnoses were made by hospital care physicians, they
tended to serve hospital care objectives, and hence the same diagnosis
category at hospital admittance could result in a wide variety of home
health situations and responses. We also found that many factors in a
patient's home situation influenced care requirements. For example home
care often involved healing from a surgery wound after surgery and
preventing or treating infections. The care involved often had little to
do with the condition which originally required the surgery and more to
do with things like complications in the surgery itself, the sanitation
in the patient's home, and underlying patient disabilities. 

The purpose of a control chart is to help distinguish between common and
special causes within a single process. In service industries like
health care, however, one is often looking at phenomena that come from a
mixture of multiple underlying processes. When this occurs, there is no
one common cause to find, and a control chart stops being a useful tool.
And it may even conceal problems because a signal related to a problem
occurring within a process can be drowned out by the noise coming from
between-process variation. 

It's helpful in these situations to identify that they're occurring, and
to come up with a way to separate the processes out by classifying
outcomes into categories and to identify (always approximately) what is
the result of what process. If one can come up with a set of
characteristics or factors that can  help one predict what category
something will end up in when it walks in the door, this can be an
incredibly helpful tool.

It might be helpful to look at some plots and histograms of the
ambulance times. Depending on the underlying distribution of the times,
there are several possible indicators of multiple systems. They include
more than one distinct peak with clear, substantial separation between
them (a little jaggedness at the edges, a small foothill peak, etc.
would be mere ordinary variation), or an "extreme value" or "heavy
tailed" situation where extremely high values are relatively rare but
still substantially more common than standard approaches might predict.
Dr. Deming mentions the Gumbel distribution in TNE; I found something
like it seemed to show up a number of times in the health care data I
looked at, and its (approximate) presence seemed to be a useful
indicator that we were looking at more than one distinct system. If it
might be possible to provide data on the ambulance times that could be
plotted over time and histogrammed, perhaps this could be discussed
further.

If this is the situation, the first step would to make a rough
classification e.g. into "short" and long trips. Once you are in a
position to classify into types, you may be in a position to identify
risk factors which tend to be present in long but not short cases (or
vice versa) or even cause the extended length. You might want to start
by looking anecdotally at individual ambulance situations, perhaps
picking cases in each category at random. It might be helpful to start
by looking at the extremes and leaving the middle cases out for the
moment. Perhaps you may find a pattern. You may not be systematically
collecting the data that's really relevant to distinguishing the
situations. You may want to interview people involved in these cases if
your records don't seem to give you much help.  But you may be lucky and
you may have relevant data on hand. 

Ideally you you will have useful theories that, perhaps after an
investment in collecting relevant data, identify risk factors for a
longer ambulance ride. Perhaps you will be in a position to do something
about some of these factors. But perhaps some of the factors are
external to you and even though you can record the data and use it to
predict, there may nonetheless be little ability to influence or avoid
them. You may not be able to reduce your response time at all. 

Even so, if you know better how to predict who is at risk for a long
wait time or ride and have knowledge that helps you understand why, you
may still be better able to use your predictions to meet other important
goals like managing availability and improving health care outcomes, and
perhaps even survival, in the patients who are in for a long wait and/or
a long haul. These goals may ultimately be more important. 

Jonathan Siegel 
 Jonathan Siegel
1204 Banbury Rd
Kalamazoo, MI, 48188
(269) 381-0829
jmsiegel@sbcglobal.net