Urine NGAL as a biomarker for septic AKI: a critical appraisal of clinical utility-data from the observational FINNAKI study

Abstract

Background: Neutrophil gelatinase-associated lipocalin (NGAL) is released from kidney tubular cells under stress as well as from neutrophils during inflammation. It has been suggested as a biomarker for acute kidney injury (AKI) in critically ill patients with sepsis. To evaluate clinical usefulness of urine NGAL (uNGAL), we post-hoc applied recently introduced statistical methods to a sub-cohort of septic patients from the prospective observational Finnish Acute Kidney Injury (FINNAKI) study. Accordingly, in 484 adult intensive care unit patients with sepsis by Sepsis-3 criteria, we calculated areas under the receiver operating characteristic curves (AUCs) for the first available uNGAL to assess discrimination for four outcomes: AKI defined by Kidney Disease: Improving Global Outcomes (KDIGO) criteria, severe (KDIGO 2-3) AKI, and renal replacement therapy (RRT) during the first 3 days of intensive care, and mortality at day 90. We constructed clinical prediction models for the outcomes and used risk assessment plots and decision curve analysis with predefined threshold probabilities to test whether adding uNGAL to the models improved reclassification or decision making in clinical practice.

Results: Incidences of AKI, severe AKI, RRT, and mortality were 44.8% (217/484), 27.7% (134/484), 9.5% (46/484), and 28.1% (136/484). Corresponding AUCs for uNGAL were 0.690, 0.728, 0.769, and 0.600. Adding uNGAL to the clinical prediction models improved discrimination of AKI, severe AKI, and RRT. However, the net benefits for the new models were only 1.4% (severe AKI and RRT) to 2.5% (AKI), and the number of patients needed to be tested per one extra true-positive varied from 40 (AKI) to 74 (RRT) at the predefined threshold probabilities.

Conclusions: The results of the recommended new statistical methods do not support the use of uNGAL in critically ill septic patients to predict AKI or clinical outcomes.

Keywords: Acute kidney injury; Critical illness; Intensive care; Neutrophil gelatinase-associated lipocalin; Sepsis.

Fig. 1

Study flowchart

Fig. 2

Risk assessment plots showing model enhancement in a AKI, b severe (KDIGO 2–3) AKI, c RRT, and d 90-day mortality. Dashed lines (baseline model) represent clinical risk models and solid lines represent new risk models with uNGAL. The gray areas between the solid and the dashed lines represent IDI_events (area between black lines) and IDI_nonevents (area between red lines). a Visually estimated from the curves, adding uNGAL to the clinical risk model improves separation of patients who will develop AKI when the risk of the event is more than ≈ 45%, and discrimination of patients who will not develop AKI when the risk of the event is less than ≈ 50%. b With severe AKI, uNGAL added to the clinical risk model improves distinguishing KDIGO 2–3 patients when the risk of the event (= severe AKI) is more than ≈ 25% and helps separating those with KDIGO stage 0–1 when the risk of the event is less than ≈ 30%. c Adding uNGAL to the clinical risk model improves the performance for assigning individuals that will end up with RRT when the risk of the event is lower than ≈ 40%, and enhances discrimination of those not ending up with RRT when the risk of the event is lower than ≈ 10%. d Corresponding statistics in Table 2, RAPs for the clinical 90-day mortality risk model and for the new model with uNGAL added illustrate that uNGAL offers only minimal enhancement separating those who will die by day 90 when the risk of the event is > 40%

Fig. 3

Decision curve analysis for a AKI, b severe (KDIGO 2–3), AKI, c RRT, and d 90-day mortality. Dashed black lines (baseline model) represent clinical risk models and dashed red lines represent new models with uNGAL. Black solid line: assume no patient has the outcome. Gray solid line: assume all patients have the outcome. a As the new model curve runs higher than the baseline curve, DCA shows a net benefit (NB) in identifying patients who will develop AKI at threshold probabilities of ≈ 0.25–0.35. The magnitude of the NB is 2.5% (95% CI 0.2–4.6%) at the predefined threshold probability of 0.30. However, at a threshold probability of 0.4, there is no NB at all. Note that if the models do not diverge from the gray line of “all expected positive”, neither of them adds anything to the strategy of expecting all to be positive at that threshold probability and should not be used. b With severe AKI, there is a 1.4% (95% CI 0.4–4.1%) NB at a threshold probability of 0.2. As with AKI, the NB does not persist within the area of clinically relevant threshold probabilities. c Adding uNGAL to the clinical RRT risk model gives a NB of 1.4% (95% CI 0.1–2.8%) in identifying patients who will end up in RRT at a threshold probability of 0.10. Note that at a threshold probability of ≈ 0.35 the curves intersect. d Decision curves for the clinical 90-day mortality risk model and for the clinical model including uNGAL do not diverge at a risk threshold of 0.05 thus showing no NB for adding uNGAL to the clinical risk model