HAI prevalence and positive predictive value

Eli’s excellent post about problems associated with use of ICD-9 codes to assess HAI burden can be extended to other surveillance approaches. That PPV erodes as disease prevalence drops (for HAIs, presumably through better prevention), is Diagnostics 101. So let’s take the example of laboratory-based tracking of C. difficile infection (CDI). The figure below is from a meta-analysis of CDI diagnostics published last year in Clinical Infectious Diseases. In most hospitals using PCR, positive test rates approach 15-20%. In these settings, the PPV is good (>80%). If a hospital or region were somehow able to push their CDI prevalence to below 5% in the tested population, however, PPV falls below 50% and gets worse as prevalence drops further. 

This is a good problem to have, generally speaking, but it does present problems for surveillance approaches as HAI rates drop (and also provides a reality check regarding “zero” targets, which are only achievable using subjective definitions that allow for human judgment/adjudication).

P.S. As a corollary, remember that when you extend CDI testing to patients with a low pre-test probability of disease (e.g. by repeat testing of those who initially test negative, by testing formed stool, or by performing "tests of cure"), you are both wasting resources and reducing PPV.  

We can't predict HAI with ICD-9 codes and it's only going to get worse

I'm getting ready for a chat with a reporter concerned with issues surrounding HAI surveillance. During my preparation, I thought again through the issues of code-based algorithms (e.g. ICD-9) and I've come to the conclusion that they are useless for assessing the burden of HAIs and HAI trends and it's only going to get worse.

One area we (and many others) have looked at is the utility of ICD-9 code-based algorithms (ie administrative codes) for detecting HAIs efficiently. A key metric frequently reported by researchers is the sensitivity of a specific code or code algorithm, which is great if the purpose of the algorithm is to improve the efficiency of detection by manual methods. Thus, if the sensitivity is high-enough, you could use the code-based algorithm to reduce the number of charts that require an IP's review. If you are using codes in this way, great!  I have no problems with that.

However, many are now using code-based algorithms to track trends in specific HAIs and measure the burden of disease. My general feeling on these is that they should be completely avoided for several reasons:

1) No matter how sensitive the algorithm is, all we care about here (since we are not validating with manual review) is the positive predictive value (i.e. the proportion of all code-positive patients that actually have the HAI of interest)

2) The PPV is very low for almost all HAI algorithms

3) If we are doing our job and lowering the incidence of HAI per admission in our hospitals the PPV by definition will only get worse (given a fixed sensitivity and specificity)

To show you why I have these concerns I have constructed two 2x2 tables evaluating an excellent hypothetical code-based algorithm for UTI with a sensitivity and specificity set at 95%.  In this first 2x2, I have evaluated the performance of the algorithm when the HAI has a 5% incidence per admission (i.e., 5% of the admissions had a UTI). You can see that such a great algorithm with a high-prevalance of disease, has a poor PPV of 50% - like flipping a coin.

Now, assume we have done an amazing job and cut our HAI rate down to 1%.  Given the same hypothetical algorithm, our PPV is now a horrible 16%. Thus, as we get better at preventing HAIs, we get worse at detecting them using code-based algorithms. Are you comfortable saying UTIs are increasing or decreasing or are associated with a certain level of excess costs, when only 16% of the UTIs in your estimation are actual (true positive) UTIs?  Me neither.