Three markers on a rapid platform reproduced critical care subphenotypes

Design

retrospective multicentre cohort study with classifier derivation and bootstrap internal validation

Population

269 critically ill children with acute cardiorespiratory failure and hyperglycaemia, 2012–2016

Primary outcome

agreement of a rapid immunoassay classifier with latent class analysis derived subphenotypes

Effect

AUROC 0.90 (95% CI 0.86–0.95), 89.6% agreement; mortality 33.3% vs 11.8% (p=0.009)

Hyperinflammatory and hypoinflammatory subphenotypes derived by latent class analysis predict outcome and modify treatment response, but they have stayed in research because deriving them needs a 13-marker multiplex assay that no unit runs in real time. This study asked whether a rapid immunoassay could stand in, using 269 critically ill children with acute cardiorespiratory failure and hyperglycaemia from multicentre paediatric intensive care units.

A parsimonious classifier using interleukin-6, interleukin-8 and soluble tumour necrosis factor receptor 1 was applied to rapid immunoassay measurements, giving an area under the receiver operating characteristic curve of 0.90 (95% CI 0.85 to 0.95) — but with poor calibration, because the rapid platform systematically underestimated soluble tumour necrosis factor receptor 1. Rebuilding the classifier on rapid immunoassay data directly gave the same discrimination, 0.90 (0.86 to 0.95), with excellent calibration and agreement with the latent class assignment in 241 of 269 cases (89.6%). The subphenotypes retained their meaning: mortality was 33.3% in the hyperinflammatory class against 11.8% in the hypoinflammatory (p=0.009), and the response to intensive insulin management differed between them (interaction p=0.024).

The calibration failure is the transferable lesson. A classifier built on one assay platform cannot simply be run on another, even when it discriminates just as well — the numbers mean different things. This is not yet a bedside test, and the population is narrow: children with both cardiorespiratory failure and hyperglycaemia, recruited a decade ago.

• Treat a classifier as assay-specific — recalibrate rather than port it between platforms
• Note this is retrospective and confined to children with hyperglycaemia and cardiorespiratory failure
• Do not use subphenotype classification to guide insulin management outside a trial
• Expect any prospective implementation to need its own calibration study
• Discrimination and calibration are separate properties — a model can have one without the other

Why it matters

It removes the practical obstacle that has kept critical care subphenotypes in research rather than trials.

Don't overread it

This is retrospective classifier development, not a trial of subphenotype-guided treatment.

The statistics, in plain English

Area under the curve measures whether the model ranks patients correctly; calibration measures whether the probability it outputs is the right number. The first classifier had excellent discrimination at 0.90 and still failed, because a systematically low input made every probability wrong in the same direction — a good illustration of why reporting area under the curve alone is not enough. The mortality difference of 33.3% against 11.8% comes from 269 children split across two classes, so the underlying counts are small.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.