Diagnostic Validation of the Updated Pediatric Sepsis Biomarker Risk (PERSEVERE-II) for Acute Kidney Injury Prediction Model in Pediatric Septic Shock

Abstract

Objective:

We previously derived the Updated Pediatric Sepsis Biomarker Risk for Acute Kidney Injury (PERSEVERE-II AKI) prediction model, which had robust diagnostic test characteristics for severe AKI (at Day 3 of septic shock (D3 severe AKI). We now sought to validate this model in an independent cohort of children to the one in which the model was developed.

Design:

A secondary analysis of a multicenter, prospective, observational study carried out January 2019 to December 2022.

Setting:

10 PICUs in the United States.

Patients:

Children with septic shock aged 1 week to 18 years admitted to the PICU.

Interventions:

None.

Measurements and Main Results:

Seventy-nine of 363 (22%) patients had D3 severe AKI, defined as Kidney Disease Improving Global Outcomes (KDIGO) stage 2 or higher. Patients were assigned a probability of D3 severe AKI using the PERSEVERE-II AKI model. The model predicted D3 severe AKI with an area under the receiver operating characteristic curve (AUROC) 0.89 (95% confidence interval 0.85–0.93), sensitivity 77% (95%CI 66–86), specificity 88% (95%CI 84–92), positive predictive value 65% (95%CI 54–74), and negative predictive value 93% (95%CI 89–96). These data represent an increase in post-test probability of D3 severe AKI with a positive test from 22% to 65%, and a prevalence threshold of 28%. On multivariable regression, the PERSEVERE-II AKI prediction model demonstrated greater adjusted odds ratio (aOR) for D3 severe AKI (aOR 11.2 [95%CI 4.9–25.3]) and lesser aOR for failure of D3 renal recovery from early AKI (aOR 0.31 [95%CI 0.13–0.69]).

Conclusion:

The PERSEVERE-II AKI model demonstrates consistently robust performance for prediction of new or persistent D3 severe AKI in children with septic shock. A major limitation is that actual D3 severe AKI prevalence is below the prevalence threshold for the test, and thus future work should focus on evaluating use in enriched populations.

INTRODUCTION

Epidemiological data from 2013 to the present demonstrate that septic shock occurs frequently in the PICU, and it is associated with poor outcomes (1, 2). Severe, persistent acute kidney injury (AKI) – defined as Kidney Disease Improving Global Outcomes (KDIGO) stage 2 or higher on Day 3 of illness – occurs in 20% of children with septic shock and is associated with increased risk for morbidity and death (3–5). Unfortunately, there are no disease-modifying therapies for severe sepsis-associated AKI once present, requiring clinicians to rely solely on supportive measures that may be more successful if implemented early in a population at high risk for AKI development or persistence (6, 7). Developing strategies for early identification of this at-risk population (i.e., prognostic enrichment) is a highlighted research priority of the recent Pediatric Acute Disease Quality Initiative consensus conference on AKI diagnosis and management in children, both to facilitate timely intervention and enrich future clinical trials (8, 9).

Septic shock is caused by a dysregulated host response to infection that is heterogeneous at the individual patient level (7, 10). It is now well-recognized that this dysregulated immune response to infection is a key driver of sepsis-associated AKI (11), and thus should be considered when developing prognostic enrichment strategies for its development or persistence. In pediatric septic shock, the Pediatric Sepsis Biomarker Risk prediction model (PERSEVERE) – using levels of C-C chemokine ligand 3, granzyme B (GZMB), heat shock protein 70 kD 1B HSPA1B), interleukin-8 (IL-8), and matrix metallopeptidase 8 – has been demonstrated to show consistent associations with organ dysfunction, including sepsis-associated AKI (12–17). Our research group previously developed and reported the PERSEVERE-II AKI model utilized on Day 1 of pediatric septic shock for prediction of severe AKI on Day 3 (D3 severe AKI) (14). This model was the first pediatric sepsis-associated AKI prognostic enrichment tool to consider the host inflammatory response to infection in risk-stratification, and it demonstrated robust diagnostic test performance with an area under the receiver operating characteristic curve (AUROC [95% confidence interval]) of 0.95 (95%CI 0.92–0.98), sensitivity of 92%, specificity of 89%, positive predictive value (PPV) of 64%, and negative predictive value (NPV) of 98%. Therefore, we sought to validate this model in an independent cohort of children to the one in which the model was developed. We hypothesized that the PERSEVERE-II AKI model would predict D3 severe AKI in children with septic shock and outperform context-free degree of serum creatinine (SCr) elevation on Day 1 of septic shock, the current gold-standard for early risk assessment.

MATERIALS AND METHODS

The original study protocol for the PERSEVERE work was approved by the central Institutional Review Board at Cincinnati Children’s Hospital Medical Center (IRB Nos: 2008–0558 [prior] and 2022–0721 [updated], titled “Genomics”, last approved 9/7/2022) prior to study enrollment and has been previously published in detail (18). All research involving human subjects was performed in accordance with the ethical standards outlined in the 1975 Declaration of Helsinki and its later amendments.

We performed a secondary analysis of an ongoing multicenter, prospective observational cohort study of children with septic shock aged one week to 18 years admitted to 10 PICUs in the United States from January 2019 to December 2022 (“Genomics of Pediatric Septic Shock”) (19). The methodology for this study remains unchanged from its last published iteration in 2019 (14). Children were enrolled in the original study if they met the 2005 International Pediatric Sepsis Consensus Conference criteria for pediatric septic shock (20) (see Supplemental Digital Content [SDC], Methods), and informed consent was obtained. After enrollment, clinical and laboratory data were collected daily during PICU admission for up to 7 days and outcomes were tracked for 28 days. All patients enrolled in the original study were considered for inclusion in this secondary analysis and excluded if they: 1) were missing SCr data from either Day 1 or Day 3; 2) had a diagnosis of end stage kidney disease; or 3) had pre-existing kidney disease without a known baseline SCr.

Outcomes and Definitions

The primary outcome was performance of the PERSEVERE-II AKI Model to predict D3 severe AKI, defined as KDIGO Stage 2 AKI or higher using SCr criteria (SCr ≥2x baseline or requiring renal replacement therapy [RRT]) (21). The previously reported PERSEVERE-II AKI prediction model (15) was used in each patient to assign a probability of D3 severe AKI based on the following information available on Day 1 of septic shock: 1) KDIGO AKI stage (21); 2) PERSEVERE-II mortality probability (see SDC Methods); and 3) values of GZMB, HSPA1B, and IL-8 (15). Baseline SCr was assigned using the lowest measured value within three months of PICU admission; unknown values were imputed using an estimated glomerular filtration rate of 120 mL/min per 1.73 m², as previously validated (22, 23). In patients without either documented baseline SCr or height for body surface area calculation (n= 11), the age-based Pottel method was used to assign a baseline SCr (24).

Secondary outcomes included the following: 1) D3 renal recovery, defined as improvement in AKI stage at D3 in patients with early AKI (≥KDIGO Stage 1 on Day 1–2; n= 152); 2) need for RRT in the first week; 3) PICU-free days, calculated as 28 minus PICU length of stay (patients who died were assigned “0”); 4) vasoactive-free days, calculated as 28 minus number of days in PICU requiring vasoactive medications (patients who died while receiving vasoactive medications were assigned “0”); 5) and 28-day mortality. Severity of illness was quantified at admission using the validated Pediatric Risk of Mortality III score (25) and in the first 24 hours using the vasoactive-inotropic score (VIS) (26). The degree of SCr elevation above baseline on Day 1 of septic shock was calculated as a ratio of Day 1 measured SCr divided by the baseline SCr, as previously reported (19, 27, 28).

Statistical Analysis

We used the Standards for Reporting of Diagnostic Accuracy Studies checklist (see SDC Appendix 1) to guide our statistical analyses. Data were described using medians, interquartile ranges (IQR), frequencies, and percentages. Comparisons between groups were performed using relative risks (RR) and 95% confidence intervals (95%CI), Wilcoxon rank sum, Chi-square, or Fisher’s exact test, as appropriate. Receiver operating characteristic (ROC) curves were generated and area under the ROC curve (AUROC) were assessed to evaluate the performance of both the PERSEVERE-II AKI Prediction Model and the degree of SCr elevation above baseline on Day 1 for prediction of D3 severe AKI. AUROCs were compared using the DeLong test. Sensitivities, specificities, PPVs, NPVs, positive likelihood ratios (+LRs) and negative likelihood ratios (-LRs) were generated for the PERSEVERE-II AKI Prediction Model (with patients predicted to have D3 severe AKI if assigned to terminal nodes [TNs] 3, 6 or 7 (15)), and Day 1 SCr/Baseline SCr (with patients predicted to have D3 severe AKI if they had KDIGO Stage 1 AKI or worse). Test characteristics were used to compute pre- and post-test probabilities (29) and a prevalence threshold (30). Multivariable logistic regression was performed to assess the independent association between being predicted to have D3 severe AKI by the PERSEVERE-II AKI Prediction Model and both 1) presence of D3 severe AKI and 2) Day 3 renal recovery from early AKI, after adjustment for significant covariates identified on univariate analysis (p<0.15). Data are summarized using adjusted odds ratios (aOR) and 95%CI. We did not include the PERSEVERE-II mortality probability – which is included in the PERSEVERE-II AKI Model – as a covariate of interest due to the risk of collinearity.

A p-value of <0.05 was considered statistically significant. All statistical analyses were performed using Sigmaplot 15.0 (Systat Software Inc., Palo Alto, CA) and R Version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria).

RESULTS

The validation dataset included 384 children with septic shock. After exclusion of patients missing SCr data (n=7) and end stage kidney disease or pre-existing kidney disease without a known baseline SCr (n=14), 363 patients were included in this analysis (see SDC Figure S1). Seventy-nine patients (22%) developed the primary outcome of D3 severe AKI. Table 1 outlines clinical, demographic and outcome data for the cohort by the presence versus absence of D3 severe AKI. The presence of D3 severe AKI was associated with higher severity of illness on admission by PRISM III score and on Day 1 of septic shock by PERSEVERE-II mortality probability and VIS. Presence of D3 severe AKI was also associated with higher concentrations of each of the individual PERSEVERE biomarkers included in our model (GZMB, HSPA1B, and IL-8). D3 Severe AKI was also associated with higher rates of immunocompromise (39% vs 17%, p <0.001), known baseline SCr (71% vs. 52%, p=0.003), and requirement of vasoactive medications (89% vs. 70, p<0.001) or mechanical ventilation (76% vs. 50%, p<0.001). We failed to identify an association between D3 severe AKI and age. D3 severe AKI, versus not, was associated with lower baseline SCr (0.30 mg/dl [0.20–0.46] vs. 0.37 mg/dl [0.26–0.50], p=0.01). Finally, D3 severe AKI was associated with increased RRT requirement (58% vs. 1%, p<0.001), fewer PICU-free days (4 [0–20] vs. 24 [6–26], p<0.001), fewer vasoactive-free days (23.5 [14–26] vs. 26 [25–28], p<0.001), and higher risk of mortality (RR 4.9 [95%CI 2.3–10.2], p<0.001).

Table 1.

Clinical features and outcomes of the cohort according to the presence or absence of severe acute kidney injury (AKI) on day 3 of septic shock.

Variable	All (n=363)	No D3 Severe AKI (n=284)	D3 Severe AKI (n=79)	p
Age, years	9.6 (3.3–16.3)	9.6 (3.8–16.3)	10.3 (2.1–16.5)	0.54
Sex, Female (%)	176 (48)	140 (49)	36 (46)	0.56
PRISM III Score	8 (5–13)	8 (4–12)	13 (8–17)	<0.001
PERSEVERE-II Mortality Probability (%)	0.7 (0.7–18.9)	0.7 (0.7–16.7)	18.9 (0.7–33.3)	<0.001
Low Risk, n (%)	237 (65)	207 (73)	30 (38)
Intermediate Risk, n (%)	73 (20)	52 (18)	21 (27)
High Risk, n (%)	53 (15)	25 (9)	28 (35)
PERSEVERE Biomarkers
GZMB, pg/mL	20.7 (8.4–49.6)	18.7 (8.2–42.5)	25.9 (9.3–126)	0.014
HSPA1B, pg/mL	233337 (143151–366335)	213910 (129856– 329822)	300242 (221669–524574)	<0.001
IL-8, pg/mL	239 (86–845)	189 (70–473)	1078 (250–5242)	<0.001
Immunocompromised, yes (%)	78 (21)	47 (17)	31 (39)	<0.001
Baseline SCr, mg/dL	0.36 (0.24–0.49)	0.37 (0.26–0.50)	0.30 (0.20–0.46)	0.01
Known Baseline, yes (%)	205 (56)	149 (52)	56 (71)	0.003
Day 1 Vasoactives, yes (%)	268 (74)	198 (70)	70 (89)	<0.001
Day 1 VIS	7 (0–18)	5 (0–14)	17 (5.5–36.5)	<0.001
Day 1 MV, yes (%)	202 (56)	142 (50)	60 (76)	<0.001
Day 1–7 RRT, yes (%)	49 (13.5)	3 (1)	46 (58)	<0.001
28-day PICU-free days	22 (10–26)	24 (16–26)	4 (0–20)	<0.001
28-day Vasoactive-free days	26 (24–27)	26 (25–28)	23.5 (14–26)	<0.001
28-day Mortality, yes (%)	27 (7.2)	11 (3.9)	15 (19)	<0.001

Continuous variables reported as median (IQR)

SCr-serum creatinine; VIS- vasoactive inotropic score; MV-mechanical ventilation; RRT-renal replacement therapy

Figure 1 illustrates the distribution of the cohort into one of seven TNs using the PERSEVERE-II AKI Model. Patients assigned to the low risk TNs (TN1, TN2, TN4, TN5) were predicted not to have D3 severe AKI while patients assigned to the high risk TNs (TN3, TN6, TN7) were predicted to have D3 severe AKI. Compared to the original derivation cohort (15), the distribution of patients in each TN differed (see SDC Table S1). To attempt to delineate reasons for these differences, Table S2 (see SDC) outlines and compares demographic characteristics, PERSEVERE data, and outcomes between the original derivation cohort (15) and this cohort. Patients in the derivation cohort were younger, had higher severity of illness, and had higher concentrations of HSPA1B and GZMB, though we failed to identify any association between timing of cohort an outcomes (see SDC Table S2).

Figure 1: Classification of the cohort according to PERSEVERE-II AKI Prediction Model.

All patients (n=363) are included in the root node at the top, with corresponding number of patients with D3 severe AKI noted below. Patients are subsequently allocated to daughter nodes using biomarker or clinical data-based criteria, as indicated in the top row of each node. Subsequent daughter nodes are generated, ending in terminal nodes (TNs) indicated in green or red font. The TNs are used to assign a baseline D3 severe AKI risk to patients assigned to that node, with corresponding risk from originally derived model indicated in parentheses. These baseline severe SA-AKI risks are used for AUROC assessment. For diagnostic test characteristic calculation, D3 severe AKI risk are dichotomized into those predicted to have D3 severe AKI (TN 3, 6 and 7 in red) and not predicted to have D3 severe AKI (TN 1,2,4 and 5 in green). All biomarker data shown are in pg/mL.

Based on the model, 94/363 (26%) patients were predicted to have D3 severe AKI. A comparison of the clinical, demographic and outcome variables of patients predicted to have D3 severe AKI by the PERSEVERE-II AKI Model to those who were not is shown in Table S3 (see SDC). Comparing those predicted to have D3 severe AKI, versus not, the severe AKI group had an associated higher severity of illness by PRISM III score, PERSEVERE-II mortality probability, and Day 1 VIS. These patients also had an associated higher proportion with immunocompromise, use of vasoactive medications and/or mechanical ventilation, and known baseline SCr (64% vs. 54%, p=0.038). However, we failed to identify an association between such grouping and baseline SCr value. Predicted D3 severe AKI was associated with greater risk of actual D3 severe AKI (RR 9.7 [95%CI 6.1–15.5], p<0.001), lower risk of D3 renal recovery from early AKI (RR 0.51 [95%CI 0.37–0.71], p<0.001), and higher percentage of RRT use in the first week (40% vs. 4%, p<0.001). This predicted D3 severe AKI grouping, versus not, was also associated with fewer PICU-free days, fewer vasoactive-free days, and higher percentage of 28-day mortality (see SDC Table S3).

Table 2 outlines the test characteristics of the PERSEVERE-II AKI Model for prediction of D3 severe AKI in contrast to context-free Day 1 SCr elevation above baseline. The PERSEVERE-II AKI Model predicted D3 severe AKI with an AUROC of 0.89 (95%CI 0.85–0.93), sensitivity of 77% (IQR 66–86%), specificity of 88% (84–92%), PPV of 65% (54–74%) and NPV of 93% (89–96%). The pre- and post-test probabilities of D3 severe AKI in patients predicted to have this outcome by the PERSEVERE-II AKI Model were 22% and 65%, respectively. The actual prevalence of D3 severe AKI was 22%, which was below the calculated prevalence threshold for the PERSEVERE-II AKI Model for D3 severe AKI of 28%. PERSEVERE-II AKI Model predictive performance was superior to Day 1 SCr/Baseline SCr (p=0.004), and when Day 1 KDIGO Stage 1 AKI was used as the threshold for prediction of D3 severe AKI, the PERSEVERE-II AKI Model demonstrated higher specificity (88% vs. 71%), PPV (65% vs. 45%) and +LR (6.6 [4.7–9.4] vs. 3.0 [2.4–3.6]) without decrement in NPV (93% vs. 95%).

Table 2.

Comparison of the PERSEVERE-II AKI Model to SCr-based acute kidney injury (AKI) on day 1 for prediction of severe AKI on day 3 of septic shock.

Variable	Day 1 PERSEVERE-II Model D3 Severe AKI Predicted	Day 1 SCr Elevation KDIGO Stage 1 AKI	p
N (% cohort)	94 (26)	152 (42)	--
Day 3 Severe AKI, yes (%)	61 (65)	69 (45)	0.003
Day 3 Severe AKI Prediction
AUROC	0.89 (0.85–0.93)*	0.82 (0.76–0.88)**	0.004
Sensitivity, %	77 (66–86)	87 (78–93)
Specificity, %	88 (84–92)	71 (65–76)
PPV, %	65 (54–74)	45 (37–54)
NPV, %	93 (89–96)	95 (91–98)
True Negatives, n (%)	251 (69)	201 (55)
False positives, n (%)	33 (9)	83 (23)
True positives, n (%)	61 (17)	69 (19)
False negatives, n (%)	18 (5)	10 (3)
+LR	6.6 (4.7–9.4)	3.0 (2.4–3.6)
−LR	0.3 (0.2–0.4)	0.17 (0.1–0.3)
Day 1–2 AKI, yes (%)	82 (87)	152 (100)	--
Day 3 Renal recovery, yes (%)	30 (37)	80 (53)	0.019
Day 1–7 RRT, yes (%)	38 (40)	40 (26)	0.021

^*AUROC for PERSEVERE-II AKI Model probability;

^**AUROC for degree of SCr elevation above baseline (Day 1 SCr divided by Baseline SCr)

On multivariable logistic regression, PERSEVERE-II AKI Model prediction of D3 severe AKI showed the strongest independent predictor of greater odds for D3 severe AKI (aOR 11.2 [95%CI 4.9–25.3], p<0.001) and lesser odds of failure of D3 renal recovery from early AKI (aOR 0.31 [95%CI 0.13–0.69], p=0.005) (Table 3). While Day 1 SCr/Baseline SCr retained some association with greater odds of D3 severe AKI (aOR 1.5 [95%CI 1.1–1.9], p=0.007), similar findings were not seen with D3 renal recovery from early AKI (Table 3).

Table 3.

Univariate and multivariable logistic regression to assess for associations with kidney outcomes of interest.

Variable	Univariate Analysis			Multivariable Analysis
	OR	95%CI	p	aOR	95%CI	p
Day 3 Severe AKI
Age, years	0.99	0.96–1.02	0.46	--	--	--
PRISM III	1.1	1.06–1.1	<0.001	1.0	0.95–1.06	0.98
Baseline SCr Used, mg/dl	0.26	0.06–1.2	0.08	0.79	0.12–5.04	0.80
PERSEVERE-II AKI Model Predicted, yes	25.3	13.4–48.1	<0.001	11.2	4.9–25.3	<0.001
Day 1 SCr Elevation	2.5	1.95–3.2	<0.001	1.5	1.1–1.9	0.007
Day 1 VIS	1.04	1.02–1.05	<0.001	1.02	1.005–1.04	0.011
Day 1 Mechanical Ventilation, yes	3.2	1.8–5.6	<0.001	2.3	1.06–5.0	0.034
Day 3 Renal Recovery from Early AKI*
Age, years	1.04	0.99–1.08	0.08	0.99	0.95–1.05	0.88
PRISM III	0.93	0.88–1.98	0.003	0.98	0.93–1.04	0.53
Baseline SCr Used, mg/dl	0.82	0.16–4.2	0.81	--	--	--
PERSEVERE-II AKI Model Predicted, yes	0.24	0.12–0.47	<0.001	0.31	0.13–0.69	0.005
Day 1 SCr Elevation	0.81	0.67–0.99	0.04	0.98	0.81–1.2	0.82
Day 1 VIS	0.97	0.95–0.99	0.002	0.98	0.96–0.99	0.03
Day 1 Mechanical Ventilation, yes	0.26	0.13–0.52	<0.001	0.38	0.17–0.83	0.016

Variables with p<0.15 on univariate analysis were included in the multivariable model.

^*Analyses only performed in the 152 patients with early (Day 1–2) AKI

SCr-serum creatinine; VIS-vasoactive inotropic score; Day 1 SCr elevation-degree of SCr elevation above baseline (Day 1 SCr divided by Baseline SCr)

DISCUSSION

In this report we have used a cohort (independent of the original test development cohort) to validate the performance of the PERSEVERE-II AKI model for prediction of new or persistent D3 severe AKI in critically ill children with septic shock. Our results show that this demonstrates consistently robust diagnostic performance, similar to the originally derived model (15). The PERSEVERE-II AKI model was again superior to using Day 1 context-free degree of SCr elevation above baseline, with higher specificity (88% vs. 71%) and PPV (65% vs 45%). Furthermore, out of these tests, the PERSEVERE-II AKI model was the stronger independent predictor of D3 severe AKI and failure of D3 renal recovery from early AKI, after adjustment for severity of illness and other potential confounders. However, while the post-test probability of D3 severe AKI was higher than the pre-test probability when a patient was predicted to have D3 severe AKI by the PERSEVERE-II AKI Model (65% vs. 22%, respectively), the actual prevalence of D3 severe AKI (22%) was lower than the prevalence threshold (28%), introducing uncertainty regarding the validation of our results in an unselected (i.e., low prevalence) population.

While SCr elevation and AKI are common early in septic shock, several studies have now demonstrated that the associations of AKI with poor outcomes are stronger when it is severe and/or persistent (i.e., KDIGO Stage 2 or higher and present ≥ 48 hours) (4, 31). For this reason, early prediction of both patients with SCr-defined AKI who will continue to have D3 AKI and those who will develop new severe D3 AKI is important for bedside risk stratification to inform prognosis and guide interventions (8, 9). For example, the recently published TAKING FOCUS 2 (Trial in AKI using Neutrophil Gelatinase-associated lipocalin [NGAL] and Fluid Overload to optimize CRRT Use) study demonstrated that a stepwise approach to AKI risk stratification 12 hours after PICU admission with subsequent clinical decision support surrounding fluid accumulation thresholds at which to consider CRRT resulted in improved survival to PICU discharge and decreased PICU length of stay following CRRT discontinuation compared to historical controls (32). Similarly, in critically ill adults following surgery, standardized AKI risk stratification using the urinary biomarker [TIMP-2]●[IGFBP7] and application of an AKI prevention bundle of care in high risk patients reduced the incidence of severe AKI and decreased ICU length of stay compared to standard care (33, 34). The PERSEVERE-II AKI model provides a similar opportunity for early AKI risk refinement (i.e,, via nearly a 3-fold increase in post-test probability of D3 severe AKI) in children with septic shock that could allow for targeted implementation of kidney protective measures or informed enrollment in clinical trials aimed at treating sepsis-associated AKI, both of which are relatively low risk interventions (Figure 2). While rapid bedside assessment of the PERSEVERE biomarkers using a widely available platform remains an important obstacle to model translation to practice, small pilot studies have demonstrated feasibility of this approach in as little as 20 minutes (35). Implementation of this tool in real-time remains an important next step to understand what, if any, impact it may have on outcomes for children with and at risk for severe sepsis-associated AKI.

Figure 2: Proposed use of the PERSEVERE-II AKI Model for prognostic enrichment children with septic shock.

A heterogeneous group of children with septic shock are risk stratified for new or persistent severe AKI at Day 3 (D3 severe AKI) using the PERSEVERE-II AKI Model (i.e., prognostic enrichment), with high-risk patients undergoing general high-risk care and considered for targeted enrollment in clinical trials.

While the PERSEVERE-II Model performed reasonably well in this cohort, its test characteristics coupled with the relatively low prevalence of D3 severe, persistent AKI in pediatric septic shock suggests that future work examining its use and implementation would likely benefit from increasing the pre-test probability of D3 severe AKI through the use of additional risk stratification tools (30). For example, using the now well-validated Renal Angina Index (RAI) (28) or its sepsis-modification (sRAI) (19, 27) would increase the pre-test probability of D3 severe AKI above the desired prevalence threshold of 28%. As these tools are easily calculable at 12 hours following ICU admission using readily available clinical data (19, 27, 28), this represents a feasible way to enrich the population and enhance the validity of our results in the future, similar to the pragmatic approach used to guide urine NGAL measurement in TAKING FOCUS 2 (32, 36). Importantly, given that the PERSEVERE-II Model relies on multi-biomarker measurement, this sort of risk stratification would also lend itself to a more cost-effective and resource-conscious approach to management. Additionally, it is also important to note that other pediatric sepsis populations (i.e., those in low- and middle-income countries) may have a higher prevalence of severe AKI than that seen in the United States (37, 38). Given that the PERSEVERE-II mortality probability has been recently validated in a low-middle income country (39), it will be important to similarly assess the performance of the PERSEVERE-II AKI model in these unique populations.

This study has several strengths. The cohort was large and encompassed children from multiple centers around the United States, increasing its generalizability. A substantial portion of the cohort (56%) had documented baseline SCr, improving the validity of SCr-defined AKI; this is highlighted by the fact that more children with D3 severe AKI had known baseline SCr (71%) compared to those without (52%). Importantly, it validates the performance of a highly predictive model for severe sepsis-associated AKI in pediatric septic shock, a key step in translating this tool to the bedside.

This work also has limitations. Accurate urine output data were not available for AKI staging, and thus AKI rates may be underestimated (40). Septic shock was not defined using the newly published consensus guidelines (2) due to the timeframe of original cohort enrollment used in this secondary analysis (2019–2022). However, most patients would meet the new Phoenix criteria given the requirement for cardiovascular failure and the co-incidence of other organ failures (i.e., mechanical ventilation). While it did not impact the performance of the model, discrepancies existed between the derivation and validation cohort in terms of TN distribution and biomarker values, especially HSPA1B. While this may be secondary to differences in severity of illness of the cohorts (i.e., the derivation cohort appears sicker), there is also the possibility of batch-to-batch variability in biomarker assays, an ongoing limitation to biomarker studies. Finally, the prevalence of D3 severe AKI in this cohort was below the prevalence threshold dictated by the test characteristics of our model. Future work should focus on examining its use in an enriched population with a high pre-test probability for D3 severe AKI.

CONCLUSIONS

The PERSEVERE-II AKI Model demonstrates robust and consistent performance for prediction of new or persistent D3 severe AKI in children with septic shock, outperforming the current gold standard of Day 1 context-free SCr elevation alone. This model showed stronger independent prediction of D3 severe AKI and failure of D3 renal recovery from early AKI, after adjustment for severity of illness and other potential confounders. Future work is needed to develop a mechanism for rapid bedside PERSEVERE biomarker measurement to facilitate real-time implementation of this tool in an enriched population (i.e., with higher pre-test probability of D3 severe AKI), allowing for assessment of the impact of prognostic enrichment for pediatric sepsis-associated AKI.

Research in Context:

• Acute kidney injury (AKI) is associated with poor outcomes in pediatric septic shock, particularly when severe and/or persistent (i.e., lasting ≥72 hours).

• Because there are no disease-modifying therapies once present, being able to predict who will have severe AKI on Day 3 of septic shock may help facilitate early, targeted intervention and inform future clinical trial enrollment.

• We previously developed the PERSEVERE-II AKI model for the prediction of severe AKI on Day 3 of septic shock and now seek to validate this model in a multicenter dataset, which is a population independent of the one in which the model was developed.

At the Bedside:

• The PERSEVERE-II AKI model maintained acceptable, consistently robust performance for prediction of severe AKI on Day 3 of septic shock (D3 severe AKI) in a separate cohort to the one in which the model was developed.

• The PERSEVERE-II AKI model outperforms the use of context-free serum creatinine elevation for the prediction of D3 severe AKI in pediatric septic shock.

• In our dataset, the actual prevalence of D3 severe AKI in pediatric septic shock is below the prevalence threshold of the PERSEVER-II AKI model, and future work is needed to identify appropriately enriched populations for use.

Data Availability:

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.