OBJECTIVES:

To explore heterogeneity in the effectiveness and cost-effectiveness of high flow nasal cannula therapy (HFNC) compared with continuous positive airway pressure (CPAP) as the first-line mode of noninvasive respiratory support for acute illness in children, using data from the FIRST-ABC Randomized Controlled Trial (RCT) (ISRCTN60048867).

METHODS:

Six hundred acutely ill children aged 0-15 years clinically assessed to require noninvasive respiratory support were randomized 1:1 to start either HFNC (n=301), or CPAP (n=299) in FIRST-ABC step down RCT. The primary outcome of the RCT was time from randomization to liberation from all respiratory support (excluding supplemental oxygen), assessed against a noninferiority margin of an adjusted hazard ratio (HR) of 0.75.

Causal Machine Learning (ML) approaches, allow for complex interactions between covariates while avoiding overfitting. We apply causal forests (CF), a causal ML method based on modified decision trees that can estimate subgroup- and individual-level treatment effects, without requiring correct pre-specification of the effect model. We consider CF alongside parametric approaches for estimating heterogeneity in Treatment Effects (HTE).

We considered pre-specified subgroups defined by age (<12 months vs >=12 months), co-morbidities (none vs neurological vs other), length of prior Invasive Mechanical Ventilation (IMV) (<5 days vs >=5 days), reason for IMV (cardiac vs non-cardiac), reason for respiratory support post-extubation (planned vs indeterminate vs rescue) and SpO2:FiO2 (SF) ratio at randomisation.

RESULTS:

HFNC was found to be noninferior to CPAP in the aggregate analysis, with a median time to liberation of 52.9 hours (95% CI, 46.0 to 60.9) versus 47.9 hours (95% CI, 40.5 to 55.7). At an individual level, considerable variation was found in the estimated effects of HFNC on median time to liberation and cost-effectiveness endpoints. Less variation was observed at the subgroup level.

CONCLUSIONS:

The ML driven HTE estimates may allow effective and cost-effective targeting of individualised treatments.