Lower limits of CRRT dose-intensity: what does the evidence support?

Dear Editor,

We thank Mei and colleagues for their careful reading of our systematic review [1] and for raising an interesting point regarding the lower threshold of CRRT dose-intensity and association with mortality. Mei et al. [2] are correct that both inputs to this subgroup are Cox hazard ratios, the multivariable adjusted HR of 1.73 (95% CI 1.19–2.51) from Okamoto et al. [3] and the baseline-adjusted HR of 1.70 (95% CI 1.16–2.49) from Faria and Libório [4], and that the pooled estimate is therefore a hazard ratio rather than an odds ratio. The label reflects the default output of our forest-plot routine and was carried through into the text; the pooled value is unaffected.

Our protocol [5] specified extraction of univariable or multivariable results but did not pre-specify how to choose among several adjusted models reported by a single study, and we accept that PRISMA item 10a asks for that process to be described [6]. At extraction we took the estimate that Faria and Libório report numerically for the 13 mL/kg/hr threshold: the only effect estimate given for that threshold in their abstract, and the only one stated identically in the abstract, the Results text and Fig. 3C (baseline-adjusted HR of 1.70 (95% CI 1.16–2.49). We appreciate the use of marginal structural model (MSM) to incorporate longitudinal average treatment effects, along with time-dependent confounding factors, treatment crossover, and informative censoring despite the retrospective nature of the data analysis. However, in Fig. 3 of Faria and Libório, panel E (MSM, 13 mL/kg/hr) is annotated HR 1.08 (95% CI 0.74–1.58) and panel F (20 mL/kg/hr) HR 0.87 (95% CI 0.66–1.14), whereas the Results text assigns 0.87 to the 13 mL/kg/hr threshold and 1.08 to 20 mL/kg/hr. We agree with Mei et al. that this should be clarified, though the clarification properly belongs to the original report. In the meantime, we do not consider a quantity that cannot be read unambiguously from the published article to be a defensible meta-analytic input.

In addition, we do not believe that the estimands from Okamoto et al. and Faria and Libório can be compared. Okamoto et al. classified patients by whole-course mean delivered dose, dichotomised at the cohort median of 13.2 mL/kg/hr, and estimated a hazard ratio from a multivariable Cox model adjusted for variables recorded at CRRT initiation [3]. Faria and Libório’s traditional Cox model is constructed in the same way: whole-course mean delivered dose, baseline-only adjustment, threshold of 13 mL/kg/hr [4]. These two estimates evaluate the same treatment comparison under the same exposure definition, and they are the only matched pair available at this threshold. The MSM targets an as-treated longitudinal effect of a time-varying exposure under time-dependent confounding, with weights accumulated over 3-day blocks. No analogous estimate exists, or can be derived, for Okamoto’s cohort. Substituting it would pool a longitudinal average treatment effect with a point-treatment effect, which is an example of an estimand mismatch that the guidance cited by Mei et al. advises against [7]. Pooling the matched estimates gives I² = 0%; substituting the marginal structural estimate raises I² to 66.8% or 88.2%. That heterogeneity is generated by the substitution rather than by any clinical difference between the two cohorts. We would also note that the ≥ 50% heterogeneity rule from our protocol would preclude reporting either of their alternative pooled estimates.

Therefore, Faria and Libório’s study was rated at serious risk of bias under ROBINS-I, with the selection-of-reported-result domain flagged precisely because different models were reported; the Discussion states that the effect estimates were susceptible to sensitivity analyses and warrant cautious interpretation, and that a minimum effective and tolerable dose-intensity may exist, but this needs to be prospectively confirmed. We agree that results from the < 13 mL/kg/hr subgroup analysis should be read as an exploratory, estimate-sensitive observation.

Our search closed on 19 September 2025. Tran et al. have since reported a further cohort drawn from MIMIC-IV, the same source data as Faria and Libório [8]. Their primary finding that low (< 20 mL/kg/hr) and moderate (20–25 mL/kg/hr) dose-intensity were associated with reduced 90-day mortality relative to > 25 mL/kg/hr is concordant with our primary analysis. In a sensitivity analysis using 15–19.9 mL/kg/hr as the reference category, doses < 15 mL/kg/hr carried an adjusted HR of 1.74 (95% CI 1.17–2.59). We note that the analysis used the same time-fixed, whole-course mean exposure definition and rest on 67 patients below 15 mL/kg/hr, so it cannot be treated as independent confirmation. (Table 1) That study also brings out a feature that the comparator is not a fixed quantity: the ≥ 13 mL/kg/hr group had a median delivered dose of 16.2 mL/kg/hr in Okamoto’s cohort but approximately 29 mL/kg/hr in Faria and Libório’s, and Tran et al. report mortality rising across successive dose bands above 20 mL/kg/hr. In our view, this is a further, and more fundamental, reason to treat the 13 mL/kg/hr threshold as exploratory.

Table 1.

Comparison of studies with available data for < 13 mL/kg/hr of delivered CRRT dose-intensity

	Okamoto et al. [2]	Faria and Libório [3]	Tran et al. [7]
Setting	• Retrospective study • Single center • Japan	• MIMIC IV database • USA	• MIMIC-IV database • USA
N	494	1329	2265
Prescribed dose (mL/kg/hr)	< 13.2 (median 11.8) vs. ≥13.2 (median 16.4)	-	-
Delivered dose (mL/kg/hr)	• 11.3 vs. 16.2	• < 20 vs. ≥20 • < 13 vs. ≥13	• Low (< 20) • Moderate (20–25) • High (> 25) • Subgroup analysis: Very low (< 15) vs. other groups
Definition of dose	Whole-course mean delivered dose	• Traditional Cox model: whole-course mean delivered dose • Marginal structural model: targeting dose-intensity as as-treated time-varying exposure with weights accumulated over 3-day blocks	Whole-course mean delivered dose
Estimates	Multivariable Cox regression model: Adjusted HR 1.73 (95% CI 1.19–2.51)	• Multivariable Cox regression model: Adjusted HR 1.70 (95% CI 1.16–2.49) • MSM adjusted for variables collected every 3 days during CRRT: marginal structural HR 0.87 (95% CI 0.66–1.14)	Multivariable Cox regression model: Adjusted HR 1.74 (95% CI 1.17–2.59) (< 15 mL/kg/h vs. 15-19.9 mL/kg/h group as the reference)

In conclusion, we agree with Mei and colleagues that no threshold should be regarded as established. We are grateful for the opportunity of this exchange and would further highlight the three ongoing randomised trials, KETZEREI (NCT06021288), LIMIT (NCT06014801) and WISDOM (NCT06446739), which will explore the lower bound of CRRT dose-intensity.

Abbreviations

BMI

Body mass index

CI

Confidence interval

CRRT

Continuous renal replacement therapy

HR

Hazard ratio

IQR

Interquartile range

MIMIC-IV

The Multiparameter Intelligent Monitoring in Intensive Care 4

MSM

Marginal structural model

ROBINS-I

Risk Of Bias In Non-randomized Studies - of Interventions

USA

United States of America

SD

Standard deviation

Author contributions

NL: Conceptualization; Writing – original draft, review and editing. SMB: Conceptualization; Writing – review and editing. All authors read and approved the final manuscript.

Funding

None.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable. This article uses only aggregate data reported in previously published studies and does not involve human participants, individual-level human data, human tissue, or animals.

Consent for publication

Not applicable.

Competing interests

SMB has received fees for scientific advisory from BioPorto, CalciMedica, Novartis, and Vantive; and has received grant support from Vantive. NL has received speaker fees from Vantive.