CRP‐To‐Albumin Ratio and CALLY Index for Sepsis Identification and Mortality Discrimination in Critically Ill Children

ABSTRACT

Background

We compared the diagnostic and prognostic performance of the C‐reactive protein‐to‐albumin (CRP/alb) ratio, neutrophil‐to‐albumin ratio (NAR), and C‐reactive protein‐albumin‐lymphocyte (CALLY) index in critically ill children.

Methods

This retrospective study included 302 children (1 month–18 years) admitted to a PICU between 2023 and 2025. Outcomes were PICU mortality and sepsis (IPSCC criteria). Discrimination was assessed via ROC analysis; independent associations were evaluated using adjusted logistic regression. Analyses assessed discrimination and adjusted associations rather than a clinical prediction model.

Results

Mortality was 19.2%; 43.0% had sepsis. For sepsis identification, the CRP/alb ratio had the highest area under the curve (AUC 0.868) with a higher AUC than the CALLY index (AUC 0.845; p = 0.012) but performed comparably to CRP alone (AUC 0.864; p = 0.217). NAR showed poor discrimination (AUC 0.571). For mortality, CRP/alb (AUC 0.757), CRP (AUC 0.751), and CALLY (AUC 0.743) demonstrated similar discrimination (all p > 0.05). Lower CALLY values independently predicted sepsis (aOR 0.94; 95% CI 0.90–0.98; p = 0.002). The CRP/alb ratio (aOR 1.17 per 10‐unit increase; 95% CI 1.06–1.29; p = 0.002) and lactate (aOR 1.34; 95% CI 1.16–1.55; p < 0.001) independently predicted mortality. NRI/IDI analyses showed that composite indices did not improve risk reclassification beyond CRP alone and occasionally worsened it, particularly for sepsis.

Conclusions

The CRP/alb ratio showed higher discrimination for sepsis than the CALLY index and was the only composite index independently associated with mortality. However, reclassification analyses did not support incremental value beyond CRP alone. Composite indices showed comparable rather than additive performance relative to CRP in this population.

1. Introduction

Sepsis remains a major cause of morbidity and mortality in pediatric intensive care units (PICUs). Multicenter and international observational studies indicate that pediatric sepsis is associated with significant morbidity and mortality, particularly among patients requiring intensive care, where mortality rates remain persistently high [1, 2]. The clinical course and outcomes of pediatric sepsis differ substantially from those in the adult population, driven by age‐dependent variations in the immune system, limited physiological reserves, and underlying comorbidities [3].

The uncontrolled inflammatory response that develops during sepsis is linked to multiple organ dysfunction, extended PICU hospitalization, and higher mortality rates [4]. Consequently, the use of easily accessible and reproducible biomarkers capable of reflecting inflammatory burden and physiological status in pediatric intensive care patients is becoming increasingly important. Recent studies suggest that ratios and indices evaluating both inflammation and nutritional status may offer greater prognostic utility than individual biomarkers used alone [5, 6].

C‐reactive protein (CRP) is an acute‐phase reactant synthesized in response to stimulation by various cytokines following infection, ischemia, trauma, and other inflammatory insults. While studies in adult populations have reported a link between elevated CRP levels and mortality [5], research in pediatric patients has demonstrated a comparatively weaker association with mortality than other biomarkers [7].

Serum albumin is a primary determinant of colloid osmotic pressure and plays a vital role in the transport of numerous hormones, medications, and bioactive elements. In studies involving critically ill pediatric patients, hypoalbuminemia has been shown to be associated with higher mortality rates, elevated Pediatric Risk of Mortality III (PRISM III) scores, increased mechanical ventilation requirements, and prolonged hospital stays [8, 9, 10, 11].

The Persistent Inflammation, Immunosuppression, and Catabolism Syndrome (PICS) framework describes three key domains of critical illness pathophysiology: systemic inflammation, catabolic or nutritional depletion, and immune dysfunction, commonly reflected by CRP, albumin, and lymphocyte count, respectively [12]. This framework has also been increasingly recognized in pediatric critical illness [13], providing a pathophysiological rationale for evaluating composite biomarkers that integrate these domains. Among such markers, the CRP/albumin ratio has been the most widely studied. A recent meta‐analysis of nine studies including 3224 patients confirmed its association with sepsis mortality [6], and a large pediatric study of 8000 patients reported that a higher CRP/albumin ratio was independently associated with increased 28‐day mortality [14]. The neutrophil/albumin ratio (NAR), which combines inflammatory burden and nutritional status, has emerged as a prognostic marker in adult traumatic brain injury and cardiovascular disease [15, 16], but has not been studied in pediatric critical care. The CALLY index, originally described in hepatocellular carcinoma [17], similarly integrates inflammatory, nutritional, and immune parameters through CRP, albumin, and lymphocyte count. Although it has shown promising results in adult sepsis cohorts [18, 19], it has not yet been evaluated in a pediatric ICU population. Accordingly, comparative evaluation of these composite biomarkers in critically ill children may help clarify their diagnostic and prognostic utility in pediatric sepsis.

The objective of this study was to evaluate and compare the diagnostic and prognostic utility of three inflammation‐nutrition indices—the CRP/albumin ratio, NAR, and CALLY index—for identifying sepsis and discriminating mortality risk in critically ill children admitted to the PICU. We hypothesized that the CRP/albumin ratio, as the most extensively studied index, would demonstrate superior discriminative performance compared to the other indices.

This study was designed to evaluate the discriminative performance and adjusted associations of these biomarkers rather than to develop or validate an individual‐level clinical prediction model.

2. Materials and Methods

This retrospective, single‐center observational study was conducted in a 15‐bed, Level III medical‐surgical pediatric intensive care unit (PICU). Patients aged 1 month to 18 years admitted to the PICU between June 1, 2023, and June 1, 2025, were eligible for inclusion. The primary outcome was PICU mortality, and the secondary outcome was sepsis, defined according to International Pediatric Sepsis Consensus Conference criteria.

Exclusion criteria were defined as a PICU stay of less than 24 h, insufficient medical records, and age outside the target range (< 1 month or > 18 years). The reasons for exclusion and the numerical distribution of these patients were systematically documented.

The study was conducted in accordance with the principles of the Declaration of Helsinki. Ethics approval was obtained from the Clinical Research Ethics Committee of University of Health Sciences, Istanbul Ümraniye Training and Research Hospital (Date: September 30, 2025; Approval No: 338).

Upon admission to the PICU, baseline laboratory parameters, including complete blood count (CBC), biochemical markers, C‐reactive protein (CRP), and procalcitonin (PCT) levels, were recorded. Demographic and clinical data, such as age (months), sex, body weight (kg), presence of comorbidities, length of stay (LOS) in the PICU, and the requirement and duration of mechanical ventilation (MV), were also documented alongside clinical outcomes.

To assess the severity of illness, the Pediatric Risk of Mortality (PRISM) and Pediatric Logistic Organ Dysfunction (PELOD) scores were calculated within the first 24 h of admission using official online calculators (PRISM: https://sfar.org/scores2/prism2.php; PELOD: http://www.sfar.org/scores2/pelod2.php). Additionally, the Pediatric Index of Mortality 3 (PIM3) and the Sequential Organ Failure Assessment (SOFA) scores were recorded. The SOFA score was utilized to determine the degree of organ dysfunction by evaluating six organ systems: respiratory, coagulation, hepatic, cardiovascular, central nervous, and renal. The PIM3 score served as a risk assessment tool based on clinical and physiological parameters at the time of admission to predict mortality risk.

Acute kidney injury (AKI) was diagnosed according to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria. Accordingly, AKI was defined as an increase in serum creatinine of ≥ 0.3 mg/dL, a ≥ 1.5‐fold increase from baseline, or a urine output of < 0.5 mL/kg/h for at least 6 h.

Sepsis was defined based on the International Pediatric Sepsis Consensus Conference (IPSCC) criteria as a systemic inflammatory response syndrome (SIRS) in the presence of, or as a result of, suspected or proven infection. The inflammation‐based indices were calculated using the following formulas:

$$\begin{array}{c}\text{CALLY Index}=\text{Albumin}\left(\mathrm{g}/\mathrm{dL}\right)\hfill \\ \times \text{Lymphocyte count}\left(\times {10}^3/\mathrm{\mu L}\right)/\mathrm{CRP}\left(\mathrm{mg}/\mathrm{L}\right)\hfill \end{array}$$

$$\mathrm{NAR}=\text{Neutrophil count}\left({10}^9/\mathrm{L}\right)/\text{Albumin}\left(\mathrm{g}/\mathrm{dL}\right)$$

2.1. Statistical Analysis

This study compared the discriminative performance of three inflammation–nutrition indices and examined their adjusted associations with sepsis and PICU mortality; it was not designed to develop or validate an individual‐level clinical prediction model. ROC AUC analyses with DeLong pairwise comparisons were the prespecified primary analyses; multivariable logistic‐regression models were secondary and were interpreted as adjusted associations; and the NRI/IDI analyses, the revised NAR Model C, and subgroup observations were exploratory.

Continuous variables are presented as mean ± SD for normally distributed data and median (IQR) for skewed data, while categorical variables are presented as frequencies and percentages. Group comparisons were performed using Student's t‐test, the Mann–Whitney U test, the chi‐square test, or Fisher's exact test, as appropriate. Complete‐case analysis was used for the primary predictors and outcomes.

Receiver operating characteristic (ROC) curve analyses provided area under the curve (AUC) values with 95% confidence intervals (CIs), optimal cut‐off points based on the Youden index, and sensitivity, specificity, positive predictive value, and negative predictive value at these cut‐offs. Pairwise comparisons of ROC curves were performed in R version 4.4.1 using the pROC package (version 1.18.4) with DeLong's test [20]. AUC 95% confidence intervals were calculated using the DeLong method. Sensitivity and specificity 95% confidence intervals at optimal cut‐off points were obtained using the Clopper–Pearson exact binomial method. Bootstrap internal validation with 1000 resamples was performed to estimate optimism‐corrected AUC values.

To evaluate the independent association of each composite index while minimizing conceptual multicollinearity due to shared component variables, three separate multivariable logistic regression models were constructed: Model A for the CRP/albumin ratio and Model B for the CALLY index for both outcomes, and Model C for NAR for sepsis only. The CRP/albumin ratio, CALLY index, and NAR were specified a priori as the primary predictors. Covariates were selected on the basis of clinical relevance while preserving an events‐per‐variable (EPV) ratio of at least 10; inclusion of hemoglobin in the mortality models would have reduced the EPV from 11.6 to 9.7. Mortality models were adjusted for age, comorbidity, lactate, and platelet count, whereas sepsis models were additionally adjusted for sex, white blood cell count, and hemoglobin. All models satisfied the EPV ≥ 10 criterion [21]. Collinearity was assessed using the variance inflation factor, with VIF < 5 considered acceptable [22].

Model performance was summarized by discrimination (area under the ROC curve), fit (Akaike Information Criterion and Nagelkerke R²), and calibration. Calibration was assessed formally rather than by visual inspection alone: by the Hosmer–Lemeshow goodness‐of‐fit test; by calibration plots that grouped patients into deciles of predicted risk and plotted observed against mean predicted probabilities (the conventional Hosmer–Lemeshow approach; Figure S1; ref. [23]); and by the calibration slope from each model's linear predictor, reported with bootstrap optimism correction. Because the apparent calibration‐in‐the‐large is zero by construction for in‐sample evaluation, we report the optimism‐corrected slope as the in‐sample metric and note that external validation is needed for a meaningful out‐of‐sample calibration‐in‐the‐large assessment. Collinearity was assessed by the variance inflation factor (VIF), with VIF < 5 considered acceptable [22]; in the sepsis NAR model (Model C), initial inclusion of total white‐blood‐cell (WBC) count produced collinearity between NAR and WBC (VIF > 5), because the neutrophil count is a component of NAR, so WBC was excluded and all VIFs in the final model were below 1.2 (maximum 1.150). Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated relative to CRP using identical covariates [24].

Primary analyses (descriptive statistics, ROC analysis with DeLong comparisons, multivariable logistic regression, the Hosmer–Lemeshow test, AIC, and Nagelkerke R²) were performed in IBM SPSS Statistics (version 30.0; IBM Corp., Armonk, NY); the bootstrap optimism‐corrected calibration slopes and the consolidated Table S6 diagnostics were computed in R (rms, pROC, ResourceSelection), as these are not standard SPSS outputs. A two‐sided p < 0.05 was considered statistically significant.

3. Results

The study cohort comprised 302 patients with a median age of 46.0 months (IQR, 8.0–123.8) and a median body weight of 14.5 kg (IQR, 6.0–26.0). Male patients accounted for 61.3% of the cohort, 62.3% had at least one comorbidity, and sepsis was present in 43.0% (n = 130). Among patients with available renal data (n = 130), AKI was identified in 20.8%. The overall survival rate was 80.8% (survivors, n = 244; non‐survivors, n = 58). Compared with survivors, non‐survivors had higher rates of comorbidity (91.4% vs. 55.3%, p < 0.001), sepsis (81.0% vs. 34.0%, p < 0.001), shock among those with available hemodynamic data (n = 145; 96.1% vs. 28.7%, p < 0.001), and AKI among those with available renal data (34.0% vs. 13.3%, p = 0.010). Admission PIM3 and SOFA scores were higher in non‐survivors (both p < 0.001), whereas age, sex, body weight, PICU length of stay, duration of invasive mechanical ventilation, PRISM‐III, and PELOD‐2 did not differ significantly between groups. Baseline demographic and clinical characteristics of the study population are presented in Table 1.

TABLE 1.

Baseline characteristics and comparison of survivors and non‐survivors.

Characteristic	Total (n = 302)	Survivors (n = 244)	Non‐survivors (n = 58)	p
Demographics
Age (months)	46.0 (8.0–123.8)	48.5 (7.0–123.2)	29.5 (9.2–125.0)	0.840
Male sex, n (%)	185 (61.3)	156 (63.9)	29 (50.0)	0.070
Body weight (kg)	14.5 (6.0–26.0)	15.0 (6.0–26.5)	12.0 (7.0–25.8)	0.850
Clinical characteristics
Comorbidity, n (%)	188 (62.3)	135 (55.3)	53 (91.4)	< 0.001
Sepsis, n (%)	130 (43.0)	83 (34.0)	47 (81.0)	< 0.001
Shock, n (%) a	76 (52.4)	27 (28.7)	49 (96.1)	< 0.001
AKI, n (%) b	27 (20.8)	11 (13.3)	16 (34.0)	0.010
Clinical outcomes
PICU length of stay (days)	7.0 (3.0–21.8)	7.0 (3.8–16.0)	14.5 (3.0–32.2)	0.160
Duration of IMV (days) c	8.0 (4.0–20.0)	8.0 (4.0–20.0)	7.0 (4.0–26.0)	0.700
Illness severity scores
PIM3 (n = 61)	4.3 (1.9–9.4)	2.5 (1.7–4.6)	10.9 (4.6–13.5)	< 0.001
PRISM‐III (n = 89)	17.0 (10.0–22.0)	16.5 (8.0–22.0)	17.0 (13.0–23.0)	0.190
PELOD‐2 (n = 89)	17.0 (11.0–22.0)	14.0 (10.8–21.2)	20.0 (11.0–23.0)	0.340
SOFA (n = 61)	6.0 (4.0–7.0)	4.0 (3.0–7.0)	7.0 (6.8–8.0)	< 0.001

Note: Data are median (IQR) or n (%). p‐values: Mann–Whitney U (continuous), χ²/Fisher's exact (categorical).

Abbreviations: AKI = acute kidney injury; IMV = invasive mechanical ventilation; PELOD‐2 = Pediatric Logistic Organ Dysfunction‐2; PICU = pediatric intensive care unit; PIM3 = Pediatric Index of Mortality 3; PRISM‐III = Pediatric Risk of Mortality III; SOFA = Sequential Organ Failure Assessment. Bold values indicate statistical significance (p < 0.05).

^a Calculated among 145 patients with available hemodynamic data (n = 94 survivors, n = 51 non‐survivors).

^b Calculated among 130 patients with available renal data (n = 83 survivors, n = 47 non‐survivors).

^c Duration of invasive mechanical ventilation was available for 159 patients.

Admission laboratory parameters are summarized in Table 2. Compared with survivors, non‐survivors had higher CRP [55.40 (23.12–104.25) vs. 5.98 (1.12–27.25) mg/L, p < 0.001], PCT [0.92 (0.30–3.75) vs. 0.23 (0.08–1.70) ng/mL, p < 0.001], and CRP/albumin ratio [17.83 (6.29–36.97) vs. 1.71 (0.30–8.36), p < 0.001]. Albumin, lymphocyte count, hemoglobin, and platelet count were lower in non‐survivors (all p ≤ 0.004). The CALLY index was also lower in non‐survivors [0.07 (0.03–0.45) vs. 1.55 (0.19–10.52), p < 0.001]. WBC count and NAR did not differ significantly between survivors and non‐survivors (p = 0.180 and p = 0.352, respectively).

TABLE 2.

Comparison of laboratory parameters and biomarker indices between survivors and non‐survivors.

Parameter	Total (n = 302)	Survivors (n = 244)	Non‐survivors (n = 58)	p
Inflammatory markers
CRP (mg/L)	9.55 (1.75–47.35)	5.98 (1.12–27.25)	55.40 (23.12–104.25)	< 0.001
Albumin (g/dL)	3.50 (3.00–3.90)	3.60 (3.10–3.92)	3.00 (2.60–3.60)	< 0.001
WBC (×103/μL)	10.50 (7.60–15.09)	10.52 (7.95–15.06)	9.59 (4.22–16.75)	0.180
PCT (ng/mL) (n = 300)	0.32 (0.09–2.00)	0.23 (0.08–1.70)	0.92 (0.30–3.75)	< 0.001
Lymphocyte (×103/μL)	2.00 (1.02–3.79)	2.14 (1.14–3.83)	1.44 (0.71–3.17)	0.004
Hematologic parameters
Hemoglobin (g/dL)	10.75 (9.43–12.17)	11.10 (9.78–12.40)	9.60 (8.53–10.95)	< 0.001
Platelet (×103/μL)	269 (160–370)	275 (186–371)	187 (39–338)	0.002
Biomarker indices
CRP/Albumin ratio	2.69 (0.47–15.81)	1.71 (0.30–8.36)	17.83 (6.29–36.97)	< 0.001
NAR	1.97 (1.17–3.14)	1.97 (1.26–3.12)	2.39 (1.25–3.77)	0.352
CALLY index	0.81 (0.07–7.00)	1.55 (0.19–10.52)	0.07 (0.03–0.45)	< 0.001

Note: Data are median (IQR). p‐values were calculated using the Mann–Whitney U test. CALLY index = Albumin (g/dL) × Lymphocyte count (×10³/μL)/CRP (mg/L). Bold values indicate statistical significance (p < 0.05).

Abbreviations: CALLY = CRP‐albumin‐lymphocyte index; CRP = C‐reactive protein; PCT = procalcitonin; NAR = neutrophil‐to‐albumin ratio; WBC = white blood cell count.

Laboratory parameters according to sepsis status are shown in Table 3. Patients with sepsis had higher CRP [54.90 (17.77–117.92) vs. 2.48 (0.64–9.22) mg/L, p < 0.001], PCT [1.00 (0.32–7.19) vs. 0.11 (0.06–0.65) ng/mL, p < 0.001], CRP/albumin ratio [16.67 (4.62–39.00) vs. 0.66 (0.16–2.62), p < 0.001], and NAR [2.27 (1.42–3.57) vs. 1.84 (1.19–2.83), p = 0.037] than those without sepsis. Albumin, lymphocyte count, hemoglobin, and platelet count were lower in the sepsis group (all p < 0.001). The CALLY index was also lower in patients with sepsis [0.07 (0.03–0.56) vs. 4.01 (0.64–19.19), p < 0.001]. WBC count did not differ significantly between patients with and without sepsis (p = 0.157).

TABLE 3.

Comparison of laboratory parameters and biomarker indices between patients with and without sepsis.

Parameter	Total (n = 302)	Sepsis (n = 130)	No sepsis (n = 172)	p
Inflammatory markers
CRP (mg/L)	9.55 (1.75–47.35)	54.90 (17.77–117.92)	2.48 (0.64–9.22)	< 0.001
Albumin (g/dL)	3.50 (3.00–3.90)	3.10 (2.80–3.58)	3.70 (3.30–4.10)	< 0.001
WBC (×103/μL)	10.50 (7.60–15.09)	10.29 (5.07–15.06)	10.60 (8.21–15.12)	0.157
PCT (ng/mL) (n = 300)	0.32 (0.09–2.00)	1.00 (0.32–7.19)	0.11 (0.06–0.65)	< 0.001
Lymphocyte (×103/μL)	2.00 (1.02–3.79)	1.44 (0.85–2.79)	2.48 (1.25–4.31)	< 0.001
Hematologic parameters
Hemoglobin (g/dL)	10.75 (9.43–12.17)	9.90 (8.60–11.40)	11.30 (10.00–12.72)	< 0.001
Platelet (×103/μL)	269 (160–370)	189 (98–350)	287 (207–379)	< 0.001
Biomarker indices
CRP/Albumin ratio	2.69 (0.47–15.81)	16.67 (4.62–39.00)	0.66 (0.16–2.62)	< 0.001
NAR	1.97 (1.26–3.18)	2.27 (1.42–3.57)	1.84 (1.19–2.83)	0.037
CALLY index	0.81 (0.08–7.00)	0.07 (0.03–0.56)	4.01 (0.64–19.19)	< 0.001

Note: Data are median (IQR). p‐values were calculated using the Mann–Whitney U test. Bold values indicate statistical significance (p < 0.05).

In multivariable logistic regression analyses, the primary indices were entered in separate models. For mortality, Model A adjusted for age, comorbidity, lactate, and platelet count showed that the CRP/albumin ratio was independently associated with mortality (aOR 1.015, 95% CI 1.006–1.026, p = 0.002), whereas in the corresponding mortality model replacing the CRP/albumin ratio with the CALLY index, CALLY was not independently associated with mortality (aOR 1.001, 95% CI 0.987–1.016, p = 0.856). Lactate was also independently associated with mortality in Model A (aOR 1.34, 95% CI 1.16–1.55, p < 0.001).

For sepsis, Models A and B adjusted for age, sex, comorbidity, lactate, WBC, platelet count, and hemoglobin showed independent associations for the CRP/albumin ratio (aOR 1.074, 95% CI 1.044–1.105, p < 0.001) and the CALLY index (aOR 0.940, 95% CI 0.903–0.978, p = 0.002), respectively. In the revised Model C excluding WBC, NAR was independently associated with sepsis (aOR 1.250, 95% CI 1.084–1.442, p = 0.002). In this model, comorbidity (aOR 4.008, 95% CI 2.224–7.224, p < 0.001), lactate (aOR 1.188, 95% CI 1.042–1.355, p = 0.010), platelet count (aOR 0.996, 95% CI 0.994–0.998, p < 0.001), and hemoglobin (aOR 0.771, 95% CI 0.672–0.884, p < 0.001) were also independently associated with sepsis.

Model A had the lowest AIC for sepsis (298.0), whereas the revised Model C had an AIC of 334.3. The mortality models showed AIC values of 244.1 for Model A and 243.5 for Model B. In the revised Model C, the maximum VIF was 1.150, Nagelkerke R² was 0.338, and the Hosmer–Lemeshow test was nonsignificant (χ² = 7.27, df = 8, p = 0.508). Bootstrap‐corrected calibration slopes were 0.854 for mortality Model A, 0.870 for sepsis Model A, and 0.895 for sepsis Model C, all below 1.0 and consistent with mild optimism. By the Hosmer–Lemeshow test, mortality Model A was adequately calibrated (χ² = 14.90, p = 0.061), sepsis Model A was significantly miscalibrated (p < 0.001) despite high discrimination, and the revised sepsis Model C was well calibrated (χ² = 7.27, df = 8, p = 0.508). Complete‐case denominators were 302 for CRP/albumin analyses, 300 for procalcitonin, 298 for CALLY, and 297 for NAR; multivariable mortality models included 302 (Model A) and 298 (Model B) patients, whereas multivariable sepsis models included 302 (Model A), 298 (Model B), and 297 (Model C) patients and DeLong comparisons used the 297‐patient matched sample. Detailed multivariable results and VIF values are in Tables S1–S3, calibration plots are in Figure S1, and a consolidated comparison of all multivariable models (discrimination, fit, Nagelkerke R², Hosmer–Lemeshow calibration, bootstrap calibration slope, and collinearity diagnostics) is provided in Supplementary Table S6.

In reclassification analyses, replacing CRP with the CRP/albumin ratio for mortality yielded an NRI of 0.068 (95% CI −0.211 to 0.349, p = 0.631) and an IDI of 0.007 (95% CI −0.000 to 0.014, p = 0.084). Replacing CRP with the CALLY index for mortality yielded an NRI of −0.240 (95% CI −0.511 to 0.024, p = 0.086) and an IDI of −0.030 (95% CI −0.060 to −0.004, p = 0.038). For sepsis, replacement of CRP with the CRP/albumin ratio yielded an NRI of −0.450 (95% CI −0.658 to −0.250, p < 0.001) and an IDI of −0.008 (95% CI −0.015 to −0.001, p = 0.036), whereas replacement with the CALLY index yielded an NRI of −0.499 (95% CI −0.733 to −0.290, p < 0.001) and an IDI of −0.168 (95% CI −0.219 to −0.117, p < 0.001). Comprehensive reclassification metrics, including bootstrap‐derived confidence intervals, are detailed in Supplementary Table S4.

ROC analyses for mortality are summarized in Table 4. The CRP/albumin ratio had an AUC of 0.757 (95% CI 0.688–0.827), followed by CRP [0.751 (0.682–0.821)], CALLY [0.743 (0.668–0.819)], and lactate [0.730 (0.658–0.803)]. At the optimal cut‐off of ≥ 5.92, the CRP/albumin ratio yielded a sensitivity of 79.3% and a specificity of 70.9%, whereas the CALLY index yielded a sensitivity of 65.5% and a specificity of 78.2% at a cut‐off of ≤ 0.16. NAR did not show significant discriminatory ability for mortality [AUC 0.540 (0.446–0.635), p = 0.403]. In paired DeLong analyses performed in the matched complete‐case sample, the AUC of the CRP/albumin ratio did not differ significantly from those of CRP (p = 0.124), CALLY (p = 0.551), albumin (p = 0.118), lactate (p = 0.800), or hemoglobin (p = 0.051), but was higher than those of NAR (p < 0.001), PCT (p = 0.008), and platelet count (p = 0.017).

TABLE 4.

Diagnostic performance of laboratory markers for predicting PICU mortality: ROC curve analysis with optimal cut‐off values determined by Youden's index.

Marker	AUC (95% CI)	Cut‐off	Sens., % (95% CI)	Spec., % (95% CI)	p
CRP/Alb	0.757 (0.688–0.827)	≥ 5.92	79.3 (66.6–88.8)	70.9 (64.8–76.5)	< 0.001
CRP	0.751 (0.682–0.821)	≥ 22.00	77.6 (64.7–87.5)	71.7 (65.6–77.3)	< 0.001
CALLY †	0.743 (0.668–0.819)	≤ 0.16	65.5 (51.4–77.8)	78.2 (72.5–83.2)	< 0.001
Lactate	0.730 (0.658–0.803)	≥ 1.68	81.0 (68.6–90.1)	60.2 (53.8–66.4)	< 0.001
Albumin	0.702 (0.624–0.781)	≤ 3.00	53.4 (39.9–66.7)	79.5 (73.9–84.4)	< 0.001
Hb	0.669 (0.589–0.748)	≤ 10.50	72.4 (59.1–83.3)	60.7 (54.2–66.8)	< 0.001
PCT ‡	0.653 (0.577–0.730)	≥ 0.33	73.7 (60.3–84.5)	56.0 (49.5–62.3)	< 0.001
PLT	0.632 (0.538–0.726)	≤ 133	43.1 (30.2–56.8)	86.5 (81.5–90.5)	0.006
NAR §	0.540 (0.446–0.635)	≥ 7.11	16.4 (7.8–28.8)	96.7 (93.6–98.6)	0.403

Note: Optimal cut‐off values were determined using Youden's index (J = sensitivity + specificity −1). AUC 95% CIs and p values (testing AUC ≠ 0.5) were calculated by the DeLong method; sensitivity and specificity 95% CIs by the Clopper–Pearson exact binomial method. Markers are ranked by descending AUC. Cut‐off values were determined by Youden index in SPSS Statistics 30.0; R pROC yields marginally different thresholds (e.g., CRP/Alb > = 5.86, CRP > = 20.70) due to algorithmic interpolation differences. Sensitivity and specificity are identical at both thresholds. All other markers n = 302.

Abbreviations: Alb, albumin; AUC, area under the receiver operating characteristic curve; CALLY, CRP–albumin–lymphocyte index (albumin × lymphocyte count/CRP); CI, confidence interval; CRP, C‐reactive protein; Hb, hemoglobin; NAR, neutrophil‐to‐albumin ratio; PCT, procalcitonin; PLT, platelet count; Sens., sensitivity; Spec., specificity.

^† n = 298.

^‡ n = 300.

^§ n = 297.

ROC analyses for sepsis are summarized in Table 5. The CRP/albumin ratio had an AUC of 0.868 (95% CI 0.826–0.910), followed by CRP [0.864 (0.822–0.907)], CALLY [0.845 (0.799–0.890)], and PCT [0.783 (0.732–0.834)]. At the optimal cut‐off of ≥ 2.84, the CRP/albumin ratio yielded a sensitivity of 83.1% and a specificity of 77.9%, whereas the CALLY index yielded a sensitivity of 69.0% and a specificity of 88.4% at a cut‐off of ≤ 0.28. In paired DeLong analyses performed in the matched complete‐case sample, the AUC of the CRP/albumin ratio did not differ significantly from that of CRP (p = 0.217), but was higher than that of CALLY (p = 0.012), PCT (p = 0.001), albumin (p < 0.001), hemoglobin (p < 0.001), platelet count (p < 0.001), lactate (p < 0.001), and NAR (p < 0.001). The receiver operating characteristic (ROC) curves for the evaluated biomarkers are shown in Figure 1. Detailed DeLong comparison statistics are provided in Supplementary Table S5.

TABLE 5.

Diagnostic performance of laboratory markers for predicting sepsis: ROC curve analysis with optimal cut‐off values determined by Youden's index.

Marker	AUC (95% CI)	Cut‐off	Sens., % (95% CI)	Spec., % (95% CI)	p *
CRP/Alb	0.868 (0.826–0.910)	≥ 2.84	83.1 (75.5–89.1)	77.9 (71.0–83.9)	Ref.
CRP	0.864 (0.822–0.907)	≥ 17.24	75.4 (67.1–82.5)	86.6 (80.6–91.3)	0.217
CALLY †	0.845 (0.799–0.890)	≤ 0.28	69.0 (60.2–77.0)	88.4 (82.6–92.8)	0.012
PCT ‡	0.783 (0.732–0.834)	≥ 0.21	83.7 (76.2–89.6)	64.3 (56.7–71.5)	0.001
Albumin	0.746 (0.691–0.802)	≤ 3.55	74.6 (66.2–81.8)	63.4 (55.7–70.6)	< 0.001
Hb	0.679 (0.618–0.740)	≤ 9.85	49.2 (40.4–58.1)	79.1 (72.2–84.9)	< 0.001
PLT	0.657 (0.592–0.722)	≤ 189.5	51.5 (42.6–60.4)	84.3 (78.0–89.4)	< 0.001
Lactate	0.584 (0.518–0.650)	≥ 1.67	57.7 (48.7–66.3)	59.9 (52.1–67.3)	< 0.001
NAR §	0.571 (0.503–0.638)	≥ 2.96	40.8 (32.1–49.9)	77.3 (70.3–83.4)	< 0.001

Note: All other markers n = 302. Optimal cut‐off values were determined using Youden's index (J = sensitivity + specificity −1). AUC 95% CIs were calculated by the DeLong method; sensitivity and specificity 95% CIs by the Clopper–Pearson exact binomial method. Pairwise p values compare each marker's AUC against CRP/Alb (DeLong test). Markers are ranked by descending AUC.

Abbreviations: Alb, albumin; AUC, area under the receiver operating characteristic curve; CALLY, CRP–albumin–lymphocyte index (albumin × lymphocyte count/CRP); CI, confidence interval; CRP, C‐reactive protein; Hb, hemoglobin; NAR, neutrophil‐to‐albumin ratio; PCT, procalcitonin; PLT, platelet count; Sens., sensitivity; Spec., specificity.

^*DeLong pairwise comparison p value versus CRP/Alb (reference).

^† n = 298.

^‡ n = 300.

^§ n = 297.

FIGURE 1.

ROC curves of inflammatory biomarkers for mortality prediction (A) and sepsis identification (B).

4. Discussion

In this retrospective cohort of 302 critically ill children, the CRP/albumin ratio showed the highest discrimination among the composite indices for both sepsis identification (AUC 0.868) and mortality (AUC 0.757), although it did not significantly exceed CRP alone for either outcome (DeLong p = 0.217 and p = 0.124). It had a nominally higher AUC than the CALLY index for sepsis (p = 0.012; not robust to conservative multiplicity correction) but not for mortality (p = 0.551), and NAR showed poor standalone discrimination (sepsis AUC 0.571; no useful mortality discrimination, AUC 0.540). Our hypothesis was therefore only partially supported: the CRP/albumin ratio led the composite indices for sepsis but offered no advantage over CRP alone, and neither composite index improved reclassification beyond CRP. These results describe marker discrimination and adjusted associations rather than a validated individual‐level risk model.

To our knowledge, this is the first direct comparison of the CRP/albumin ratio, CALLY index, and NAR within a single unselected pediatric intensive care cohort that included both septic and non‐septic admissions. The main contribution is comparative: in children, the diagnostic information carried by the CRP/albumin ratio for sepsis is essentially that of CRP itself (the two were statistically indistinguishable, p = 0.217), the ratio had a nominally higher AUC than the CALLY index (p = 0.012; not robust to conservative multiplicity correction), and NAR contributed little. The independent association of the CRP/albumin ratio with mortality (adjusted OR 1.015 per unit, or 1.166 per 10‐unit increment, approximately a 17% increase in the odds of mortality) should be interpreted as an adjusted association, not an individual‐level risk estimate. This is directionally consistent with the closest comparable pediatric report (Mohamed and Elhawary [25]), and our lower optimal cut‐off likely reflects broader inclusion and a more heterogeneous case mix (Arslan et al. [26]). Findings in adult oncological, cardiovascular, and acute‐care populations are concordant [6, 26, 27, 28, 29, 30, 31, 32, 33].

The CALLY index discriminated sepsis (AUC 0.845) better than mortality (AUC 0.743), and the same pattern appeared in the adjusted models: lower CALLY values were independently associated with sepsis (adjusted OR 0.940, p = 0.002) but showed no independent association with mortality after adjustment for comorbidity and lactate (adjusted OR 1.001, 95% CI 0.987–1.016, p = 0.856). This contrasts with adult sepsis cohorts that reported strong CALLY discrimination for mortality (Yılmaz and Ak [19], AUC 0.906 for 30‐day mortality; Zhang et al. [18]) and with the sepsis‐only pediatric cohort of Wang et al. [34], in which lower CALLY independently predicted death.

This divergence may be developmental. Leukocyte composition changes substantially with age in children: lymphocytes predominate in infancy, and the neutrophil‐to‐lymphocyte ratio approaches the adult ratio only at around 4 to 6 years of age, so adult‐derived thresholds for lymphocyte‐ and neutrophil‐based indices may transfer poorly to pediatric patients [35]. Severe pediatric sepsis is also associated with early and persistent lymphopenia, which has been linked to prolonged organ dysfunction and PICU mortality [36] and attributed to lymphocyte apoptosis and functional exhaustion [37]. These developmental and sepsis‐related changes may partly explain why the lymphocyte component of the CALLY index and the neutrophil‐based NAR discriminated less well in this cohort, and the inclusion of non‐septic admissions may have further reduced any mortality signal. Because this study was not powered for age‐stratified or immune‐phenotype analyses, these mechanisms are offered as plausible explanations rather than tested mediators.

NAR showed poor standalone discrimination on ROC analysis (sepsis AUC 0.571; no useful mortality discrimination, AUC 0.540) and did not differ between survivors and non‐survivors (p = 0.352). It reached independent significance only in the revised sepsis Model C (adjusted OR 1.250, 95% CI 1.084–1.442, p = 0.002), and only after WBC was removed to address structural collinearity (original VIF > 5). Model C also had the poorest overall fit of the three sepsis models (AIC 334.3, compared with 298.0 for Model A). This association should therefore be regarded as model‐dependent and hypothesis‐generating rather than as evidence that NAR is a useful bedside marker in children. In adults, by contrast, NAR has shown prognostic value in traumatic brain injury [15], cardiovascular disease [16], and COVID‐19 [38], which suggests that the null pediatric result reflects a population‐specific difference rather than a general limitation of the marker.

Discrimination and calibration were interpreted separately. The CRP/albumin‐based sepsis Model A had the highest multivariable discrimination and the best AIC, but it was significantly miscalibrated by the Hosmer–Lemeshow test. Conversely, the revised Model C showed acceptable Hosmer–Lemeshow calibration but weaker discrimination and poorer overall fit. Bootstrap‐corrected calibration slopes below 1.0 indicated mild optimism across models. In this cohort, therefore, high discrimination was not accompanied by adequate calibration, and the CRP/albumin ratio should be interpreted as a marker of discrimination and adjusted association rather than as a calibrated individual‐level risk model.

Neither composite index improved reclassification beyond CRP. For sepsis, both indices worsened it (CRP/albumin NRI = −0.450, IDI = −0.008; CALLY NRI = −0.499, IDI = −0.168; all p ≤ 0.036). For mortality, the CRP/albumin ratio was reclassification‐neutral, whereas the CALLY index did not improve NRI and significantly worsened IDI (IDI −0.030, 95% CI −0.060 to −0.004, p = 0.038). Negative NRI and IDI values indicate that adding albumin‐ or lymphocyte‐derived information to CRP did not improve patient classification and, for sepsis, worsened it. This does not mean that the composite indices were unrelated to sepsis; rather, their additional components did not add useful classification information beyond CRP alone. Albumin and lymphocyte counts may vary with nutritional status, fluid balance, steroid exposure, and other noninfectious factors, which may explain the lack of incremental classification benefit [39]. This negative incremental‐value finding is clinically relevant because it cautions against replacing CRP with more complex adult‐derived composite indices in pediatric critical care without prospective validation.

In this cohort, the CRP/albumin ratio showed higher sensitivity and the CALLY index higher specificity at their respective cut‐offs; these descriptive patterns are hypothesis‐generating and require prospective validation before any screening or confirmatory role can be proposed. Given the modest adjusted effect size (adjusted OR 1.015 per unit, or 1.166 per 10‐unit increment) and the negative reclassification results, neither index should be used as a stand‐alone indicator of mortality risk.

The present study has several limitations. First, its retrospective and single‐center design may limit generalizability. Second, biomarkers were evaluated at a single time point on PICU admission, and temporal trends were not assessed. Third, the limited number of non‐survivors (n = 58, EPV = 11.6) may have reduced the precision of the mortality model estimates, and PIM3 and PRISM‐III scores were not included as regression covariates. Among the sepsis models, Model A showed significant miscalibration (Hosmer–Lemeshow p < 0.001), and any probability‐based thresholds would require external validation [40] In the original Model C, multicollinearity between NAR and WBC (VIF > 5) necessitated exclusion of WBC; the revised model showed acceptable collinearity (VIF < 1.2) but higher AIC. We could not account for preadmission treatments or symptom duration. Because we used raw lymphocyte and neutrophil counts rather than age‐standardized values to compute the CALLY index and NAR, the discriminative performance of these indices may have been attenuated by developmental hematological variability across the pediatric age spectrum. Complete‐case analysis was used; although missingness for the composite indices was low (1.3%–1.7%), no sensitivity analysis with imputation was performed. Pairwise DeLong comparisons were not adjusted for multiplicity; the CRP/albumin ratio versus CALLY index comparison for sepsis (p = 0.012) would not remain significant after Bonferroni correction for eight pairwise tests (adjusted threshold = 0.006), and the reported p‐values should be interpreted as nominal. This study was not prospectively registered. A further limitation is our use of the 2005 International Pediatric Sepsis Consensus Conference (IPSCC) criteria rather than the 2024 Phoenix Pediatric Sepsis Criteria [41, 42]. The IPSCC framework is based on systemic inflammatory response syndrome (SIRS) criteria, which have limited specificity in critically ill children and may classify transient inflammatory responses as sepsis while incompletely capturing life‐threatening organ dysfunction. The Phoenix criteria instead define sepsis as infection with life‐threatening organ dysfunction (Phoenix Sepsis Score ≥ 2). Because our cohort predated the 2024 criteria, our SIRS‐based case definition may identify a broader and less specific sepsis phenotype, and the CRP/albumin ratio, CALLY index, and NAR should be re‐evaluated under the Phoenix definition in future studies. The IPSCC definition was the available historical case definition and supports comparison with prior pediatric biomarker studies, but the findings should not be generalized uncritically to Phoenix‐defined sepsis.

Taken together, these findings indicate that none of the evaluated composite indices demonstrated clear superiority over CRP alone. Despite these limitations, our findings support further validation in larger, multicenter pediatric cohorts.

5. Conclusions

In this single‐center retrospective cohort of critically ill children, the CRP/albumin ratio showed the highest discrimination among the evaluated composite indices for both sepsis identification (AUC 0.868) and mortality (AUC 0.757) and had a nominally higher AUC than the CALLY index for sepsis (DeLong p = 0.012; not robust to conservative multiplicity correction), by a small absolute margin and without exceeding CRP alone. Neither composite index improved reclassification or integrated discrimination beyond CRP, indicating limited added value over this widely available single marker. Lower CALLY values were independently associated with sepsis (adjusted OR 0.940) but not with mortality, and NAR showed poor standalone discrimination (sepsis AUC 0.571; no useful mortality discrimination, AUC 0.540); its significant Model C association was limited by poor standalone discrimination, inferior fit, and model dependence. These findings are exploratory and describe discrimination and adjusted associations rather than a validated prediction tool; prospective multicenter studies are needed to determine whether these composite indices add value beyond CRP in pediatric critical care.

Author Contributions

G.A. and S.E. contributed to the conception and design of the study. G.A., İ.P., M.E.P., and D.K. performed data collection and analysis. G.A. and S.E. drafted the manuscript. All authors reviewed and approved the final version of the manuscript.

Funding

The authors have nothing to report.

Ethics Statement

The study was approved by the Clinical Research Ethics Committee of the University of Health Sciences, İstanbul Ümraniye Training and Research Hospital (Approval No: 338, Date: September 30, 2025). The study was conducted in accordance with the principles of the Declaration of Helsinki. Since this was a retrospective study, the requirement for informed consent was waived by the ethics committee.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Calibration plots of predicted versus observed probabilities for Model A. (A) Mortality and (B) sepsis identification. The dashed 45° line indicates perfect calibration; points represent deciles of predicted risk. While the mortality model was well‐calibrated (HL χ² = 14.90, p = 0.061; AUC = 0.844), the sepsis model showed significant deviation from the ideal line (χ² = 36.59, p < 0.001), despite a high AUC of 0.874. These multivariable AUCs exceed the unadjusted performance of the CRP/albumin ratio reported in the main text.

Table S1: Multivariable Logistic Regression for PICU Mortality.

Table S2: Multivariable Logistic Regression for Sepsis.

Table S3: Variance Inflation Factors (VIF).

Table S4: Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI).

Table S5: DeLong Pairwise Comparisons of AUC.

Table S6: Consolidated multivariable model performance.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethics restrictions.