Monday, June 17, 2024

Sepsis: A Rational Definition Founded on Big Data


The diagnostic criteria for sepsis in adults have been in flux since the early 1990s, and are still a matter of debate.    Criteria have evolved via expert panels and consensus.   Newly, an international group has used "big data" from tens of thousands of cases to develop diagnostic criteria for pediatric sepsis.   It's called the "Phoenix" project and appeared recently in JAMA.

This blog includes, at bottom, a bit about the "AUPRC" stat used in Phoenix.


JAMA has a home page for collected sepsis articles and op eds (here).  The most recent international consensus document was "Sepsis-3," issued in 2016 by the "Third International Concensus Definitions" group, Singer et al. JAMA 315:801.  It was designed by a 19-member task force from the US Society of Critical Care Medicine and the European Society of Intensive Care Medicine.   They include a summary of prior work, such as "Sepsis-1" in 1991 and "Sepsis-2" in 2001.

There is a lot of spread between the Sepsis-2 and Sepsis-3 definitions (see Engoren 2020 here and Vermassen 2021 here.)    Debate spills over into related areas, like the CMS SEP-1 quality measure which was recently elevated to a financially impactful status under Medicare despite criticisms of the measure (IDSA here).

  • From a different field, see the colorful and stormy history of "consensus definitions" in the psychiatry handbook DSM; find the new book by Horwitz here.

Big Data Informs Pediatric Sepsis Definition

I was struck by the February 2024 paper by Sanchez-Pinto et al., in JAMA, "development and validation of Phoenix Criteria for pediatric sepsis and shock."   Here.   The data analysis project spanned 5 countries and used literally hundreds of thousands of patient records.   The resulting model out-performed prior 2005 international pediatric consensus criteria.   

The extremely large body of empirical data ["A," Phoenix study] then informed a separate pediatric consensus criteria process ("B," see Schlapbach et al., here) to adopt the Phoenix result.   These authors write: 

  • Prior pediatric sepsis criteria that used systemic inflammatory response syndrome (SIRS) criteria, which have poor predictive properties, and included the redundant term, severe sepsis. 
  • The SCCM task force recommends that sepsis in children be identified by a Phoenix Sepsis Score...

JAMA includes an op ed overview of the whole process by Carlton et al. here.

Nerd Note - Novel Statistics Right Here in JAMA

The Phoenix paper makes much use of the AUPRC metric - area under the precision recall score, which is an alternative to the far, far more common "AUROC" or area under the receiver operating curve.  I'm not sure I've got my head around this yet.   
  • Whereas the virtue of the AUROC is that it looks at clean simple sensitivity and specificity as if in a vacuum, and AUROC deliberately leaves base rate and PPV-NPV for a later day, the AUPRC builds PPV into itself, so it's only valid for a population with a similar base rate.
Read about AUPRC in the Phoenix paper, and they cite to a "super paper," Saito & Rehmsmeier 2015, in PLOS One, with some 2000 citations, for further understanding.

Beyond reading Saito & Rehmsmeier, I will leave a public link to a long Chat GPT dialog where I tried to teach myself more about PRC - here.  

You can also take yourself to a YouTube video - here.