“there are holes:
- the assumption that blood cultures at 48 hours are some sort of gold standard (at least with specificity; ability to detect absence of infection)”

Now that looks like the start of an Archimedes question …

Now as to UTI/sepsis. Fair enough. But were they septic, or are these contaminants/perioperative local infections?

And the ‘looks unwell’ test — that might also be an Archi - how reliable/reproducible is the ‘looks sick’ test? (I have vague memories of the Rochester criteria looking at kappa values for this …)

Bob - I take your point about the nature of the infections, but it is worth remembering that UTI is a very serious infection in neonates, not least because it is associated with more generalised sepsis in a way which seems to be inversely proportional to age.

One of the things that is interesting to me here is the inclusion of “looks unwell”. My practice currently, right or wrong, is to to rely quite heavily on what the admitting registrar says about the state of the child at the point of admission. I try to get them to say “Look, this one was sick, and needs a full five days of IV antibiotics” or “I was just following protocol and you could probably stop IVs if the cultures are negative at 48 hours”. Superficially this works. However, there are holes:
- the assumption that blood cultures at 48 hours are some sort of gold standard (at least with specificity; ability to detect absence of infection)
- the assumption that what the doctor feels about the patient at the point of admission means something - and this paper here seems to undermine this a bit.

This is particularly odd when pretty much the entire aim of the paediatric early warning systems under development is to get a score which as good as an experienced doctor standing at the end of the bed and saying “that one looks pretty sick”

I will give you the 18d old with Pneumococcal sepsis as an unequivocal miss though. Well, almost give you that one, as they did a tympanocentisis and found the bug there too. Now, I may just be admitting to being less thorough that this lots, but when was the last time you undertook a tympanocentisis on anyone, let alone an 18d old? Does this say that the girl wasn’t really looking too well?

And there’s something niggling about the methodology. The ‘decision rule’ that has been tested is a bit odd, I agree. In the discussion the authors comment other rules are ‘essentially similar’ … but this might be ‘effectively different’. Is this paper really proposing their criteria as a new (and not-very-effective) rule for LR of infection, or trying to test a rule previously described (but not quoted)? Why didn’t they use the different criteria (like Boston & Phili) and tell us how those rules performed? And where does the ‘23 cells’ come from? (The Harriet Lane Handbook might be very reputable, but if I were building a clinical decision rule, I’d want to see the hard science behind it first.)

You see, what I think we really need to do is take all these data sets in a meta-analysis, and test the rules that exist on them: in large sets are they effective or not? And with larger sets, can we build better rules?