Admitting ICU diagnosis is a poor predictor of infection

According to a recent study:

Methods

We studied a cohort of critically ill patients admitted with clinically suspected sepsis to two tertiary ICUs in the Netherlands between January 2011 and December 2013. The likelihood of infection was categorized as none, possible, probable or definite by post-hoc assessment. We used multivariable competing risks survival analyses to determine the association of the plausibility of infection with mortality.

Results

Among 2579 patients treated for sepsis, 13% had a post-hoc infection likelihood of “none”, and an additional 30% of only “possible”. These percentages were largely similar for different suspected sites of infection. In crude analyses, the likelihood of infection was associated with increased length of stay and complications. In multivariable analysis, patients with an unlikely infection had a higher mortality rate compared to patients with a definite infection (subdistribution hazard ratio 1.23; 95% confidence interval 1.03-1.49).

Conclusions

This study is the first prospective analysis to show that the clinical diagnosis of sepsis upon ICU admission corresponds poorly with the presence of infection on post-hoc assessment. A higher likelihood of infection does not adversely influence outcome in this population.

There are some important lessons from this study. Though we often have to aggressively give antibiotics “just in case” in severely ill patients, who often have undifferentiated presentations, many of these will turn out not to be infected. This speaks for the importance of diligent efforts to de-escalate. In order to facilitate rapid initiation of antibiotics, remember that sepsis is defined clinically. That is, the definition requires only that infection be suspected, not established. It is a clinical syndrome more than it is a diagnosis (in much the same way ARDS is a syndrome rather than a specific diagnosis). The findings in this paper provide one more illustration of how unrealistic it is to expect clinicians to assign a specific diagnosis on the front end even though that is what the coders would like us to do.