CORRESPONDENCE| VOLUME 7, ISSUE 10, OCTOBER 01, 2019

Secondary re-analysis of the FEAST trial

Kathryn Maitland, Diana M Gibb, Abdel Babiker

Lancet Respir Med 2019; 7: e29 DOI:https://doi.org/10.1016/S2213-2600(19)30272-3

The article by Michael Levin and colleagues¹ rekindles the discussion about adverse effects of boluses in children with febrile illness and shock, as reported in the FEAST trial.² However, unfortunately, results of this re-analysis of FEAST data fall very short of providing mechanisms by which boluses increase mortality; differences in physiological scores and biochemistry results between groups by no means imply causality. Furthermore, some re-analyses and statements about our previous papers are erroneous and contradict work by the FEAST team aimed at understanding these mechanisms.

We firmly stand by our previous analyses of terminal clinical events (TCEs) in FEAST showing that the main reason for excess deaths with boluses was cardiovascular collapse, and not respiratory or neurological causes.³ Adverse events and deaths, possibly related to fluid overload, were actively solicited and reviewed throughout the trial by an endpoint review committee, masked to the group. The committee reviewed structured clinical narratives detailing clinical presentation, progression, and terminal events reported by trial clinicians trained on standardised proforma, together with serial bedside clinical observations and laboratory data. As reported in our previous study,² only about 2% of children had possible respiratory or neurological fluid overload events. Furthermore, based on pre-specified criteria, the endpoint review committee adjudicated TCE to cardiovascular, respiratory, or neurological causes without knowing the treatment. These data showed excess cardiovascular (collapse) TCEs in the bolus groups with no excess deaths in respiratory or neurological TCEs. We believe this evaluation is a much more robust analysis of mechanisms than equations that use surrogate risk scores. Levin¹ criticises this paper, indicating that lactate was included in the definition of cardiovascular TCE, which is incorrect.³

Furthermore the authors have 100% imputed base excess, chloride, bicarbonate and haemoglobin values 1 h after the randomisation; per design, none were actually measured at 1 h. The imputed values were derived from published literature of studies in individuals who had received larger 1 h fluid volumes and from data on the unselected (two centres) minority (32%) of FEAST survivors at 24 h, which anyway show normalised values (figure). Treating the imputed values as observed is misleading. Of note, the median bolus in FEAST was only 20 mL/kg (IQR 20–40), which was previously reported to normalise base excess⁴ and is modest compared with volumes recommended in current guidelines.

Secondary re-analysis of the FEAST trial

Matteo Quartagno, Bianca De Stavola, Richard Emsley, et al

Lancet Respir Med 2019; 7: e30 DOI:https://doi.org/10.1016/S2213-2600(19)30270-X

We read with interest the article by Michael Levin and colleagues,¹where various secondary analyses were done on data from the FEAST trial to explore mechanisms by which fluid resuscitation therapy led to increased mortality in children with shock in Africa.

Unfortunately, the statistical methods used have major errors which invalidate their conclusion that “Hyperchloraemic acidosis and respiratory and neurological dysfunction induced by saline or albumin bolus explained the excess mortality due to bolus”. The first key error relates to the imputation of base excess, chloride, and bicarbonate at 1 h after randomisation for all patients (Article's appendix). The authors use two approaches to impute a single value at 1 h for each patient. The first approach is based on a separate linear interpolation in each treatment group, calculated from incomplete data on these parameters for survivors at 24 h, and applied to all patients; however, 87% of deaths occurred before 24 h²In the second approach, the 1 h values for base excess, bicarbonate, and chloride were derived from changes after saline infusion reported in other studies. These errors generate false conclusions. First, a single value has been imputed, and then treated as observed in the analysis; this statistically invalid procedure is known to hugely overestimate the information.³ Second, in the first approach this single value rests on the untested assumption of linear trend over 24 h and on extrapolation from trends seen in children who survived up to that time. Finally, both approaches result in a constant change from baseline within each treatment group: they impute the same value of change for all patients in the same treatment group. Thus, the imputed concentrations at 1 h are confounded with the treatment assignment (bolus or no bolus). Because of this confounding, the Cox models fitted by the authors erroneously suggest that differences in acidosis and chloride parameters (together with physiological scores) explain the treatment difference in mortality between bolus and no bolus.

There is a further major statistical flaw. Even if these variables are measured in all children at 1 h after the randomisation with pinpoint accuracy, the authors' analysis is not sufficient to support their causal conclusion about the mediating role of hyperchloraemic acidosis and respiratory and neurological dysfunction. Such analyses would require controlling for confounding factors of the relationship between abnormal physiology and biochemistry and mortality (mediators–outcome relationships). We are not protected against such confounding by randomisation, because these confounders might arise after the randomisation. Besides, changes in the estimate of the hazard ratio after adjusting for a putative mediator might not be indicative of mediation even if no confounding issues were present. Extensive literature exists on methods for mediation analyses that can be given a causal interpretation,^4, 5 but none of these methods are mentioned and none of the assumptions under which the authors' conclusions hold are discussed.