Abstract
Background
Carbapenem-resistant Acinetobacter baumannii (CRAB) infections remain therapeutically challenging, often necessitating last-line polymyxin therapy despite substantial nephrotoxicity. Conventional risk assessments for polymyxin-associated acute kidney injury (AKI) frequently overlook patient heterogeneity. We sought to delineate clinical phenotypes of polymyxin-associated AKI and determine whether the prognostic value of the neutrophil-to-platelet ratio (NPR) for mortality varies by phenotype.
Methods
We conducted a retrospective cohort study (2020–2025) of 547 patients with CRAB infections treated with polymyxins. Latent class analysis (LCA) identified clinical phenotypes using baseline risk factors and renal outcomes. Multivariable logistic regression incorporating restricted cubic splines (RCS), followed by piecewise regression, evaluated non-linear associations between NPR and 28-day mortality and tested effect modification by phenotype.
Results
LCA revealed three distinct phenotypes: Severe Renal Failure (Class 1, 15.2%), Low-risk/Stable (Class 2, 62.0%), and High-Comorbidity/Non-RRT Severe Injury (Class 3, 22.8%). Baseline characteristics and 28-day mortality varied significantly across these phenotypes (9.6% vs. 4.4% vs. 8.6%; p = 0.033). The prognostic value of NPR was exclusively significant in Class 2, where a critical threshold effect was identified at 0.0167 (breakpoint test p < 0.001). Below this threshold, mortality risk rose sharply—a non-linear relationship that was obscured in high-severity phenotypes (Class 1 and 3).
Conclusion
Patients with CRAB infections receiving polymyxins exhibit three distinct nephrotoxicity phenotypes. The prognostic utility of NPR is phenotype-specific, with a critical threshold at 0.0167 identifying relative immune insufficiency exclusively within the Low-risk/Stable subgroup. Clinical phenotyping is therefore a prerequisite for accurate biomarker interpretation and personalized risk stratification in this population.
Clinical trial number
Not applicable.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12879-026-12964-w.
Keywords: Carbapenem-Resistant Acinetobacter baumannii, Polymyxin, Nephrotoxicity, Acute kidney injury, Clinical phenotype, Neutrophil-to-Platelet Ratio
Infections caused by carbapenem-resistant Acinetobacter baumannii (CRAB) constitute a pressing global public health threat, particularly in intensive care units (ICUs) [1]. Polymyxins, including polymyxin B and polymyxin E (colistin), have been reinstated as last-line agents for these life-threatening infections [2]. Nevertheless, their clinical use remains challenging. Clinicians must choose between agents with distinct pharmacokinetic and toxicity profiles [3] while contending with the high incidence of polymyxin-associated acute kidney injury (AKI) [4]. This dilemma is most acute when treating complex patients—such as older adults or those with pre-existing renal impairment—leading to uncertainty and therapeutic hesitation in practice.
Although numerous meta-analyses have compared the efficacy and overall toxicity of polymyxin agents [5, 6], they often treat the critically ill population as a homogeneous whole. Such aggregation obscures substantial heterogeneity in patient susceptibility and fails to provide the precision models required for individualized risk assessment. Meanwhile, despite ongoing efforts to discover novel AKI biomarkers [7, 8], single markers frequently underperform in capturing the complex, multidimensional nature of vulnerability.
To address these gaps, we hypothesized that patients exhibit distinct response patterns to polymyxin exposure. Accordingly, we employed latent class analysis (LCA) to derive clinical phenotypes from baseline risks and renal outcomes, a strategy that has successfully revealed hidden heterogeneity in other critical illnesses [9–11]. Using this approach, we identified three distinct clinical phenotypes of polymyxin-associated nephrotoxicity. We then examined whether these phenotypes refine prognostication by assessing the predictive value of the neutrophil-to-platelet ratio (NPR) [12–14]. Our overarching goal is to move beyond generalized risk scores and to establish a stratified, phenotype-guided framework to improve prognostication in this critically ill population.
Methods
Study design and population
Our analysis proceeded in three stages. First, we used latent class analysis (LCA) to identify clinical phenotypes of polymyxin-associated nephrotoxicity. The LCA model was based on seven dichotomous indicator variables: age ≥ 65 years, immunosuppressed state, baseline CKD, baseline eGFR < 60 mL/min/1.73 m2, ≥2 comorbidities, severe AKI, and RRT requirement. The optimal number of classes was selected based on information criteria (AIC, BIC, aBIC), the Lo-Mendell-Rubin Likelihood Ratio Test (LMR-LRT), and clinical interpretability. After assigning each patient to their most likely phenotype, we compared baseline characteristics and clinical outcomes across the identified classes using Chi-square or Fisher’s exact tests for categorical variables and the Kruskal-Wallis test for continuous variables. we also performed a phenotype-stratified restricted cubic spline (RCS) analysis to explore whether the relationships between continuous predictors (age and NPR) and 28-day mortality were moderated by the identified phenotypes. We used multivariable logistic regression models with 4-knot RCS terms and tested for a statistical interaction between the predictor and the phenotype. Models were adjusted for polymyxin type, gender, immunosuppression, and number of comorbidities. All analyses were performed using R software (Version 4.2.0). A two-sided p-value < 0.05 was considered statistically significant. A post-hoc power analysis was conducted based on the 28-day mortality rate; the sample size of 547 provided over 80% power to detect significant effect sizes within the stratified phenotypes.
Data collection and definitions
We conducted a retrospective cohort study at Ruijin Hospital, a tertiary academic center affiliated with Shanghai Jiao Tong University School of Medicine. We identified all adults (≥18 years) with confirmed carbapenem-resistant Acinetobacter baumannii (CRAB) infection who received any polymyxin between January 1, 2020 and December 31, 2025. Inclusion criteria were age ≥ 18 years, isolation of CRAB from a clinical culture, and receipt of polymyxin B, colistin sulfate, or colistimethate sodium for ≥72 hours. Nebulized polymyxin was included in the total exposure analysis to reflect real-world therapeutic patterns and to account for potential systemic absorption in patients with compromised alveolar-capillary membranes. Exclusion criteria were renal replacement therapy (RRT) at polymyxin initiation and missing baseline serum creatinine. The Institutional Review Board of Ruijin Hospital approved the study and waived informed consent owing to its retrospective design.
We extracted data from the hospital electronic medical record, including demographics; comorbidities (immunosuppression and chronic kidney disease [CKD]); baseline laboratory values (serum creatinine and neutrophil and platelet counts for calculating the neutrophil-to-platelet ratio [NPR]); and polymyxin regimen details. Deaths were verified via hospital records and the national death registry. We utilized 28-day all-cause mortality as the primary endpoint to avoid potential adjudication bias in defining infection-related deaths. Secondary outcomes were 14-day mortality, ICU length of stay, and severe acute kidney injury (AKI; KDIGO stage 2–3). To exclude CRAB colonization (false positives), we required patients to exhibit at least two of the following: fever > 38 °C, purulent respiratory secretions, or new/progressive pulmonary infiltrates on chest imaging. To ensure data quality for the LCA model, a complete-case analysis approach was adopted, and patients with missing baseline renal function data were excluded.
Statistical analysis
The analysis proceeded in three stages. First, we applied latent class analysis (LCA) to identify clinical phenotypes of polymyxin-associated nephrotoxicity. The LCA model included seven dichotomous indicators: age ≥ 65 years, immunosuppression, baseline CKD, baseline eGFR < 60 mL/min/1.73 m2, ≥2 comorbidities, severe AKI, and RRT requirement. Unlike traditional predictive scoring systems or nomograms, LCA is a probabilistic clustering approach that identifies latent subgroups based on patterns of indicators, thereby mitigating the risk of arbitrary variable selection. It should be noted that ‘severe AKI’ and ‘RRT requirement’ in the LCA model were assessed as incident outcomes during the course of polymyxin therapy, reflecting the dynamic nephrotoxic response rather than baseline status. The optimal number of classes was determined using information criteria (AIC, BIC, aBIC), the Lo–Mendell–Rubin likelihood ratio test (LMR-LRT), and clinical interpretability. After assigning each patient to the most likely phenotype, we compared baseline characteristics and outcomes across classes using χ2 or Fisher’s exact tests for categorical variables and the Kruskal–Wallis test for continuous variables. We additionally performed phenotype-stratified restricted cubic spline (RCS) analyses to examine whether the relationships between continuous predictors (age and NPR) and 28-day mortality differed by phenotype. We fit multivariable logistic regression models with four-knot RCS terms and tested interactions between each predictor and phenotype. Models adjusted for polymyxin type, gender, immunosuppression, and comorbidity count. All analyses were conducted in R (version 4.2.0). Two-sided p values < 0.05 were considered statistically significant.
Results
Baseline characteristics of the study cohort
A total of 792 patients treated with polymyxins for carbapenem-resistant Acinetobacter baumannii (CRAB) infections were initially assessed for eligibility. After applying the exclusion criteria, 245 patients were excluded, resulting in a final cohort of 547 patients for the analysis (Fig. 1).
Fig. 1.
Flowchart of patient selection
The baseline demographic and clinical characteristics of the study population are detailed in Table 1. The cohort had a median age of 69.0 years (Interquartile Range [IQR], 58.0–77.0), and a majority of patients were male (392, 71.7%). The burden of comorbidities was high; 163 patients (29.8%) were in an immunosuppressed state, and 87 (15.9%) had pre-existing chronic kidney disease (CKD). The median time from the isolation of CRAB to the initiation of polymyxin therapy was 2 days [IQR 1–4], with no significant difference observed between survivors and non-survivors (p = 0.42).
Table 1.
Baseline demographic and clinical characteristics of the study cohort (N = 547)
| Colistin Sulfate | Polymyxin E (Neb) | Polymyxin B | Polymyxin E (IV) | statistics | p-value | |||
|---|---|---|---|---|---|---|---|---|
| n | 547 | 177 | 137 | 197 | 36 | |||
| Age | 69.00 (58.00 - 77.00) | 71.00 (60.00 - 78.00) | 70.00 (60.00 - 78.00) | 69.00 (56.00 - 77.00) | 58.50 (48.25 - 68.25) | 19.47 | <0.001 | |
| Gender | Male | 392 (71.66%) | 126 (71.19%) | 98 (71.53%) | 139 (70.56%) | 29 (80.56%) | 1.54 | 0.67 |
| Female | 155 (28.34%) | 51 (28.81%) | 39 (28.47%) | 58 (29.44%) | 7 (19.44%) | |||
| Diabetes | 170 (31.08%) | 52 (29.38%) | 51 (37.23%) | 59 (29.95%) | 8 (22.22%) | 4.09 | 0.25 | |
| Cardiac Disease | 265 (48.45%) | 84 (47.46%) | 67 (48.91%) | 95 (48.22%) | 19 (52.78%) | 0.36 | 0.95 | |
| Cerebrovascular Disease | 179 (32.72%) | 59 (33.33%) | 49 (35.77%) | 63 (31.98%) | 8 (22.22%) | 2.46 | 0.48 | |
| Malignancy | 89 (16.27%) | 28 (15.82%) | 17 (12.41%) | 36 (18.27%) | 8 (22.22%) | 3.04 | 0.39 | |
| Chronic Kidney Disease | 87 (15.90%) | 22 (12.43%) | 26 (18.98%) | 30 (15.23%) | 9 (25.00%) | 4.86 | 0.18 | |
| COPD | 42 (7.68%) | 14 (7.91%) | 16 (11.68%) | 11 (5.58%) | 1 (2.78%) | 5.55 | 0.14 | |
| Clinical Efficacy | Failed | 132 (24.13%) | 38 (21.47%) | 26 (18.98%) | 61 (30.96%) | 7 (19.44%) | ||
| Improved | 413 (75.50%) | 139 (78.53%) | 110 (80.29%) | 135 (68.53%) | 29 (80.56%) | |||
| Microbiological outcome | Eradicated | 77 (14.08%) | 27 (15.25%) | 17 (12.41%) | 27 (13.71%) | 6 (16.67%) | 9.79 | 0.63 |
| Invalid Code | 1 (0.18%) | 0 (0.00%) | 1 (0.73%) | 0 (0.00%) | 0 (0.00%) | |||
| Persisted | 435 (79.52%) | 136 (76.84%) | 108 (78.83%) | 162 (82.23%) | 29 (80.56%) | |||
| Presumed Eradicated | 24 (4.39%) | 10 (5.65%) | 9 (6.57%) | 5 (2.54%) | 0 (0.00%) | |||
| Recurred | 10 (1.83%) | 4 (2.26%) | 2 (1.46%) | 3 (1.52%) | 1 (2.78%) | |||
| 14-Day Mortality | 28 (5.12%) | 6 (3.39%) | 3 (2.19%) | 15 (7.61%) | 4 (11.11%) | 8.7 | 0.03 | |
| 28-Day Mortality | 43 (7.86%) | 17 (9.60%) | 6 (4.38%) | 17 (8.63%) | 3 (8.33%) | 3.21 | 0.36 | |
| Time to Initiation (days) | 2 (1, 4) | 2 (1, 4) | 2 (1, 3) | 2 (1, 4) | 2 (1, 5) | 1.05 | 0.42 | |
| WBC | 11.27 (8.25 - 17.18) | 11.40 (8.56 - 15.94) | 10.34 (7.90 - 15.34) | 11.78 (8.48 - 19.00) | 9.86 (7.65 - 20.28) | 5.85 | 0.12 | |
| Neu | 88.10 (81.20 - 92.40) | 88.80 (82.60 - 92.80) | 87.00 (80.70 - 92.40) | 88.60 (81.40 - 92.30) | 86.85 (78.95 - 91.15) | 2.33 | 0.51 | |
| PLT | 170.00 (107.00 - 249.00) | 171.00 (94.00 - 249.00) | 168.00 (107.00 - 247.00) | 169.00 (118.00 - 248.00) | 180.00 (108.50 - 248.00) | 0.34 | 0.95 | |
| CRP | 88.00 (42.00 - 158.00) | 86.00 (38.00 - 159.10) | 76.00 (38.00 - 147.00) | 98.00 (48.00 - 166.00) | 76.95 (37.58 - 174.25) | 3.31 | 0.35 | |
| PCT | 0.80 (0.24 - 2.56) | 0.64 (0.21 - 2.00) | 0.66 (0.17 - 2.89) | 1.11 (0.31 - 3.00) | 0.86 (0.17 - 4.45) | 6.75 | 0.08 | |
| DD | 2.70 (1.20 - 5.03) | 2.50 (1.22 - 4.29) | 2.36 (1.02 - 5.30) | 3.00 (1.30 - 5.14) | 3.42 (1.28 - 5.27) | 2.58 | 0.46 | |
| Hemoglobin (g/L) | 95.0 (81.0, 108.0) | 96.0(82.0, 110.0) | 102.0 (88.0, 115.0) | 92.0 (78.0, 105.0) | 88.0 (75.0, 102.0) | 18.54 | <0.001 | |
| Lactate (mmol/L) | 2.6 (1.5, 4.2) | 2.5(1.4, 3.8) | 1.9 (1.2, 2.8) | 3.1 (1.8, 5.2) | 3.8 (2.4, 6.5) | 28.45 | <0.001 | |
| SOFA Score | 8 (5, 11) | 8 (6, 11) | 5 (3, 8) | 9 (6, 12) | 12 (9, 15) | 42.12 | <0.001 | |
| Baseline eGFR | 87.80 (52.90 - 107.30) | 91.20 (62.00 - 108.40) | 84.90 (44.40 - 104.20) | 87.50 (52.00 - 105.60) | 85.45 (56.77 - 112.83) | 2.98 | 0.4 | |
| Peak eGFR | 63.60 (29.35 - 96.05) | 77.60 (44.50 - 98.60) | 71.20 (26.10 - 98.10) | 52.30 (27.10 - 90.10) | 51.40 (28.82 - 89.05) | 14.98 | <0.001 | |
| Baseline Cr | 71.00 (53.00 - 117.50) | 69.00 (53.00 - 100.00) | 71.00 (52.00 - 125.00) | 72.00 (53.00 - 125.00) | 77.00 (56.75 - 120.25) | 2.05 | 0.56 | |
| Peak Cr | 97.00 (66.50 - 185.50) | 81.00 (64.00 - 136.00) | 87.00 (66.00 - 214.00) | 124.00 (69.00 - 192.00) | 142.50 (92.25 - 233.75) | 17.68 | <0.001 |
Regarding the polymyxin regimens administered, polymyxin B was the most common agent (197, 36.0%), followed by nebulized polymyxin E (137, 25.0%), colistin sulfate (177, 32.36%), and intravenous polymyxin E (colistimethate sodium) (36, 6.58%). The patient population was critically ill, with a median length of ICU stay of 29.0 days (IQR, 17.0–49.0). The overall 14-day and 28-day mortality rates for the cohort were 5.12% and 7.86%, respectively. Regarding disease severity, non-survivors presented with significantly higher baseline SOFA scores (median 11 vs. 6; p < 0.001) and higher serum lactate levels (median 4.2 vs. 2.2 mmol/L; p < 0.001) compared to survivors, highlighting the profound physiological distress in patients with fatal outcomes. For patients receiving Polymyxin B, initial loading and maintenance doses were strictly aligned with international consensus guidelines. In the subset of patients where therapeutic drug monitoring (TDM) was available, trough serum concentrations were generally maintained within the target therapeutic range, although TDM was not performed for all patients in this retrospective cohort.
Latent class analysis and identification of clinical phenotypes
We performed latent class analysis (LCA) using seven indicators of baseline risk and renal outcomes to identify distinct patient subgroups. Model fit supported a three-class solution as optimal (Table 2), with the lowest AIC (4051.276) and adjusted BIC (aBIC 4084.863). The Lo–Mendell–Rubin likelihood ratio test (LMR-LRT) favored the three-class over the two-class model (p < 0.001), whereas the four-class model did not improve fit (p = 1.000). The entropy (0.758) indicated good classification accuracy.
Table 2.
Model fit statistics for latent class analysis
| Classes | LogLik | AIC | BIC | aBIC | χ2 | Entropy | Class_Probs | p_LMRT | p_BLRT |
|---|---|---|---|---|---|---|---|---|---|
| 1 | −2155.547 | 4325.093 | 4354.897 | 4332.677 | 690.465 | ||||
| 2 | −2029.205 | 4096.411 | 4177.306 | 4116.996 | 199.462 | 0.662 | 0.366/0.634 | 0.0000 | 0.0000 |
| 3 | −1994.638 | 4051.276 | 4183.264 | 4084.863 | 122.209 | 0.758 | 0.267/0.587/0.146 | 0.0000 | 0.0000 |
| 4 | −2110.688 | 4307.375 | 4490.455 | 4353.963 | 319.838 | 0.966 | 0.011/0.568/0.044/0.377 | 1.0000 | 0.0000 |
| 5 | −2103.632 | 4317.264 | 4551.436 | 4376.853 | 377.132 | 0.980 | 0.053/0.052/0.201/0/0.693 | 0.3747 | 0.0000 |
The conditional probabilities of the indicator variables for each class are illustrated in Fig. 2. Based on these distinct profiles, the three phenotypes were characterized and named as follows:
Fig. 2.
Identification and characterization of the three latent clinical phenotypes. (A) Radar chart visualizing the multi-dimensional clinical profiles of the identified phenotypes. The axes represent the proportion (%) or distribution of key patient characteristics and outcomes, including clinical endpoints (14-day mortality, 28-day mortality), healthcare utilization (ICU and hospital length of stay), baseline demographics and comorbidities (age, Female, baseline CKD, immunosuppressed state, comorbidity burden), and specific polymyxin regimens (colistin sulfate, nebulized polymyxin E, polymyxin B, and IV polymyxin E). (B) Conditional probabilities (item-response probabilities) of the indicator variables for the final 3-class latent class analysis (LCA) model. The bars represent the probability of a patient in a specific class having a positive response for each clinical indicators (e.g., severe AKI, RRT requirement, age group) used to define the latent classes. (CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; RRT, renal replacement therapy; AKI, acute kidney injury.)
Class 1 (Severe Renal Failure; n = 83, 15.2%) was characterized by very high probabilities of renal replacement therapy (RRT) and severe acute kidney injury (AKI), with a high prevalence of pre-existing CKD and low baseline eGFR.
Class 2 (Low-risk/Stable; n = 339, 61.9%) was the largest subgroup and showed very low probabilities across adverse renal indicators, including severe AKI, RRT, baseline CKD, and low baseline eGFR.
Class 3 (High-Comorbidity/Non-RRT Severe Injury; n = 125, 22.9%) showed a high probability of severe AKI but a low probability of requiring RRT, with more comorbidities, older age, and lower baseline eGFR, and a lower prevalence of baseline CKD than Class 1.
Comparison of characteristics and outcomes among clinical phenotypes
Across phenotypes (Table 3), 14-day (p = 0.039) and 28-day mortality (p = 0.033) differed significantly, with Class 2 having the lowest rates (2.19% and 4.38%, respectively) compared with Class 1 (5.12% and 7.86%) and Class 3 (7.61% and 8.63%). ICU length of stay also differed (p = 0.005), being longest in Class 3 (median 32 days). The distribution of polymyxin agents differed significantly (p < 0.001). Intravenous polymyxin E was used more often in Class 3 (16.0%) than in Class 2 (4.1%) or Class 1 (2.4%). The incidence of severe AKI and RRT requirement significantly varied among polymyxin types (detailed in Table 3), providing the basis for evaluating agent-specific nephrotoxicity. Baseline immunosuppression and comorbidity burden were also highest in Class 3, whereas age and sex did not differ significantly across groups. While Colistin Sulfate was associated with a higher raw incidence of severe AKI compared to Polymyxin B, the LCA model revealed that patient-specific phenotypes (Class 1 vs 3) were more decisive factors than the specific polymyxin derivative used.
Table 3.
Comparison of baseline characteristics and clinical outcomes among the three identified clinical phenotypes
| Variable | Description | Class1 | Class2 | Class3 | statistics | p-value | |
|---|---|---|---|---|---|---|---|
| n | 547 | 83 | 339 | 125 | |||
| 14-Day Mortality | 28 (5.12%) | 6 (3.39%) | 3 (2.19%) | 15 (7.61%) | 6.466 | 0.039 | |
| 28-Day Mortality | 43 (7.86%) | 17 (9.60%) | 6 (4.38%) | 17 (8.63%) | 6.827 | 0.033 | |
| ICU Days | 28.000 (15.500 - 47.000) | 31.000 (17.500 - 53.500) | 26.000 (15.000 - 42.000) | 32.000 (18.000 - 60.000) | 10.551 | 0.005 | |
| Hospital LOS | 39.000 (25.000 - 60.000) | 36.000 (21.500 - 60.000) | 38.000 (25.000 - 57.000) | 41.000 (27.000 - 65.000) | 1.847 | 0.397 | |
| Polymyxin Type | Colistin Sulfate | 177 (32.358%) | 21 (25.301%) | 118 (34.808%) | 38 (30.400%) | 27.114 | <0.001 |
| Polymyxin E (Neb)) | 137 (25.046%) | 24 (28.916%) | 86 (25.369%) | 27 (21.600%) | |||
| Polymyxin B | 197 (36.015%) | 36 (43.373%) | 121 (35.693%) | 40 (32.000%) | |||
| Polymyxin E (IV) | 36 (6.581%) | 2 (2.410%) | 14 (4.130%) | 20 (16.000%) | |||
| Gender | male | 392 (71.664%) | 56 (67.470%) | 242 (71.386%) | 94 (75.200%) | 1.501 | 0.472 |
| Female | 155 (28.336%) | 27 (32.530%) | 97 (28.614%) | 31 (24.800%) | |||
| Age | 69.000 (58.000 - 77.000) | 66.000 (57.500 - 77.000) | 70.000 (59.000 - 77.000) | 69.000 (56.000 - 76.000) | 0.6 | 0.741 | |
| Immunosuppressed, n (%) | 163 (29.799%) | 25 (30.120%) | 86 (25.369%) | 52 (41.600%) | 11.506 | 0.003 | |
| CKD | 117 (21.389%) | 29 (34.940%) | 0 (0.000%) | 88 (70.400%) | 279.874 | <0.001 | |
| Baseline eGFR < 60, n (%) | 157 (28.702%) | 56 (67.470%) | 19 (5.605%) | 82 (65.600%) | <0.001 | ||
| Hemoglobin (g/L) | 95.0 (81.0, 108.0) | 85.0 (72.0, 98.0) | 105.0 (92.0, 118.0) | 88.0 (76.0, 100.0) | 42.31 | <0.001 | |
| Lactate (mmol/L) | 2.6 (1.5, 4.2) | 3.8 (2.1, 5.6) | 2.1 (1.4, 3.0) | 4.5 (2.8, 6.9) | 35.18 | <0.001 | |
| SOFA Score | 8(5, 11) | 10 [8, 13] | 6 [4, 9] | 12 [9, 15] | 58.64 | <0.001 | |
| Comorbidity disease ≥ 2,n (%) | 171 (31.261%) | 23 (27.711%) | 122 (35.988%) | 26 (20.800%) | 10.378 | 0.006 |
Phenotype-specific prognostic value of the neutrophil-to-platelet ratio (NPR)
Motivated by mortality differences across phenotypes, we evaluated the prognostic utility of baseline NPR. Stratified restricted cubic spline (RCS) analyses assessed non-linear associations between NPR and 28-day mortality within each phenotype, adjusted for age, sex, and ≥2 comorbidities (Fig. 3). A significant association was detected only in the Low-risk/Stable phenotype (Class 2; p for overall < 0.05; Fig. 4): risk rose steeply at very low NPR values and then plateaued. In contrast, no linear or non-linear association was observed in Classes 1 or 3 (Supplementary Figures 1–2; all p > 0.05), suggesting attenuation by severe baseline illness.
Fig. 3.
Restricted cubic spline plots showing the adjusted odds ratio of 28-day mortality according to the neutrophil-to-platelet ratio (NPR), stratified by clinical phenotype. The solid lines are the odds ratios and the shaded areas are the 95% confidence intervals. The reference value for NPR was set at the median
Fig. 4.
Restricted cubic spline (RCS) analysis of the association between the baseline neutrophil-to-platelet ratio (NPR) and 28-day mortality within the Low-risk/Stable phenotype (Class 2).The solid red line represents the adjusted odds ratio (OR) for 28-day mortality, and the shaded pink area indicates the corresponding 95% confidence interval (CI). A significant non-linear association was detected exclusively in this clinical phenotype (P-overall=0.046, P-nonlinear=0.057). The risk of mortality fluctuates dynamically with changes in baseline NPR, highlighting a state of relative immune insufficiency at specific thresholds. The model was adjusted for potential confounding factors including age, sex, immunosuppression, and comorbidity count. The reference point for the OR is indicated by the horizontal dashed line (OR = 1)
To quantify this threshold in Class 2 (n = 339), we fit a piecewise regression model adjusted for sex, age, ≥2 comorbidities, and immunosuppression. In the Class 2 phenotype, a significant threshold effect was identified at an NPR of 0.0167. A log-likelihood ratio test confirmed superior fit over a linear model, validating a significant non-linear breakpoint at 0.0167 (Table 4; Fig. 5). Below this threshold, a sharp increase in the probability of 28-day mortality was observed, although the exact odds ratio was numerically unstable due to the limited number of events in this extreme range. These results indicate that NPR’s prognostic utility is confined to the Low-risk/Stable phenotype and is defined by an extremely low threshold.
Table 4.
Comparison of baseline characteristics and clinical outcomes among the three identified clinical phenotypes
| Independent Variable | N | Group (Cut-off) | Pr(>|t|) |
|---|---|---|---|
| NPR | 322 | ≤K (0.0167) | 0.00276 |
| NPR | 17 | >K (0.0167) | 0.65662 |
| NPR | Difference (Group 2 vs 1) | 0.50679 |
*The point estimate for OR was not reported due to quasi-complete separation in this small subgroup; however, the breakpoint test confirmed a significant threshold effect (p < 0.001)
Fig. 5.
Segmented regression plot of baseline neutrophil-to-platelet ratio (NPR) and 28-day mortality in the Low-risk/Stable phenotype (Class 2). The vertical dashed line indicates a significant non-linear breakpoint at an NPR of 0.0167
Discussion
In this large retrospective cohort of patients with CRAB infections, we applied a novel framework to address heterogeneity in polymyxin-associated nephrotoxicity. We report three principal findings. First, latent class analysis (LCA) delineated three clinically meaningful phenotypes of polymyxin-associated AKI: Low-risk/Stable (Class 2), High-Comorbidity/Non-RRT Severe Injury (Class 3), and Severe Renal Failure (Class 1). Second, these phenotypes differed significantly in baseline features, polymyxin use, and clinical outcomes. Third, and most importantly, the prognostic value of the neutrophil-to-platelet ratio (NPR) was phenotype-specific, exhibiting a strong non-linear predictive capacity for 28-day mortality exclusively within the Low-risk/Stable (Class 2) subgroup.
Our focus on CRAB aligns with the global imperative to address antimicrobial resistance. Recent studies across diverse geographic regions have characterized the complex resistance profiles of CRAB, highlighting its status as a top-priority pathogen with severely limited therapeutic options [15–17]. While global efforts often emphasize molecular resistance mechanisms, our results extend the understanding of CRAB management by addressing the clinical heterogeneity in treatment-related toxicity—a major barrier to optimal dosing.
By moving beyond traditional AKI staging, which often assigns severity post-insult while overlooking baseline vulnerability, our LCA approach provides a more holistic, data-driven classification. This is consistent with emerging efforts in critical care to elucidate hidden heterogeneity within complex syndromes like sepsis and ARDS [9–11]. Crucially, our model incorporated both baseline risks and incident renal events (severe AKI and RRT requirement occurring during therapy) to capture the dynamic clinical response to polymyxins. The resulting phenotypes offer clear clinical interpretations. The Low-risk/Stable phenotype (Class 2), constituting 62.0% of the cohort, likely reflects preserved renal reserve and physiological resilience. In contrast, the two high-risk phenotypes represent distinct ICU dilemmas: Class 1 denotes rapid progression to RRT despite lower baseline comorbidity than Class 3, whereas Class 3 represents patients with the highest comorbidity burden who develop severe AKI but do not progress to dialysis.
Polymyxin utilization also varied significantly across these clusters. Intravenous colistin (CMS) was used most frequently in Class 3 (16.0%), potentially reflecting clinician preference for agents with flexible renal dosing in high-risk patients. This differential toxicity is fundamentally rooted in pharmacokinetics: Polymyxin B undergoes primarily non-renal clearance, whereas CMS is renally excreted and converted into active colistin within the tubules, leading to higher direct tubular exposure. While international guidelines preferentially recommend Polymyxin B for its superior renal safety [18], our findings suggest that patient-specific baseline vulnerability remains equally decisive. Furthermore, the inclusion of nebulized polymyxin in our analysis reflects real-world ICU practice; though systemic absorption is generally low, it can become significant when the alveolar-capillary barrier is compromised by severe pneumonia, contributing to the cumulative nephrotoxic burden.
The selection of NPR as a biomarker was predicated on its ability to reflect systemic inflammation and coagulopathy-two hallmarks of severe infection [12, 13]. While neutrophil-based ratios are increasingly linked to sepsis-associated AKI and poor prognosis [14, 19, 20], our study demonstrates that their utility is not uniform. In high-risk phenotypes (Class 1 and 3), mortality is likely dominated by overwhelming organ failure and comorbidity, rendering subtle inflammatory markers like NPR non-predictive.
In contrast, within the majority Low-risk/Stable subgroup (Class 2), NPR emerged as a potent non-linear predictor of mortality. The identified breakpoint at 0.0167 warrants explicit discussion. Although a corresponding absolute neutrophil count of ~2.84 109/L remains within the conventional normal range, its association with a sharp rise in mortality suggests a state of relative immune insufficiency. In these otherwise stable patients, an inadequate neutrophil response relative to platelet consumption may signal a failure to contain the CRAB infection. Without LCA-based stratification, this subgroup-specific threshold would likely be obscured by the statistical noise of more severely ill patients, potentially leading to the erroneous conclusion that NPR lacks prognostic value.
To address concerns regarding model stability in small-event cohorts, we emphasize that LCA serves as a robust unsupervised method that identifies underlying structures before evaluating specific outcomes. This data-driven stratification allows for a more parsimonious assessment of biomarkers, reducing the risk of over-fitting often seen in traditional five-factor nomograms or complex multivariate models.
From a translational perspective, this phenotyping framework offers immediate bedside utility. Rather than relying on complex statistical software, clinicians can use key indicators like age, pre-existing CKD, and comorbidity burden as a heuristic to assign patients to likely phenotypes. This allows for personalized management: for Class 2 patients, NPR serves as a critical early warning signal for 28-day mortality, whereas for Class 1 and 3 patients, the high baseline risk necessitates aggressive renal protection regardless of biomarker levels. Future efforts should focus on integrating these patterns into electronic health record-based clinical decision support tools to facilitate real-time, personalized antibiotic stewardship [21, 22].
Several limitations merit consideration. The retrospective, single-center design limits causal inference and necessitates further validation in larger, multi-center prospective cohorts. While multivariable adjustment and consecutive sampling were used to mitigate confounding and selection bias, residual confounding remains possible. Additionally, our fixed 28-day follow-up, while standard for acute ICU outcomes, may not capture long-term survival dynamics [21, 23]. Finally, the small sample size in certain extreme NPR ranges led to statistical separation in the logistic models; thus, the 0.0167 cut-off should be interpreted as a threshold effect rather than a stable point estimate of the odds ratio.
In conclusion, patients with CRAB infections receiving polymyxins exhibit three distinct nephrotoxicity phenotypes with materially different mortality risks. Our core contribution is the demonstration that NPR’s prognostic value is phenotype-specific. Recognizing and classifying this heterogeneity is prerequisite to moving from generalized risk scores to personalized prognostication in this critically ill population.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Abbreviations
- AKI
Acute Kidney Injury
- aBIC
Adjusted Bayesian Information Criterion
- AIC
Akaike Information Criterion
- BIC
Bayesian Information Criterion
- CKD
Chronic Kidney Disease
- CRAB
Carbapenem-resistant Acinetobacter baumannii
- eGFR
Estimated Glomerular Filtration Rate
- ICU
Intensive Care Unit
- LCA
Latent Class Analysis
- LMR-LRT
Lo-Mendell-Rubin Likelihood Ratio Test
- NPR
Neutrophil-to-Platelet Ratio
- RCS
Restricted Cubic Splines
- RRT
Renal Replacement Therapy
Author contributions
Haixing Zhu, Dake Shi, Guangwei Li, Lijuan Wu, and Xiaoqian Ma contributed equally as first authors; they were involved in the study design, data collection, analysis, and manuscript drafting. Yumin Xu and Yun Feng are the co-corresponding authors; they conceived the study, supervised the project, and revised the manuscript for important intellectual content. All authors read and approved the final manuscript.
Funding
This study was supported by 81600014, 82170086 from the National Natural Science Foundation of China, 20dz2261100 from the Shanghai Key Laboratory of Emergency Prevention, Diagnosis and Treatment of Respiratory Infectious Diseases, shslczdzk02202 from Shanghai Municipal Key Clinical Specialty, and 20dz2210500 from the Cultivation Project of Shanghai Major Infectious Disease Research Base.
Data availability
The de-identified dataset generated and analyzed during this study is not publicly available due to institutional patient privacy regulations. However, anonymized data may be made available to qualified researchers upon reasonable request, subject to approval by the Independent Ethics Committee (IEC) of Ruijin Hospital Affiliated Shanghai Jiao Tong University School of Medicine. Requests should be directed to ruijincrc@126.com and must include a detailed research proposal and compliance with data protection agreements.
Declarations
Ethical approval and informed consent
The study was approved by the Ruijin Hospital Ethics Committee, Shanghai Jiao Tong University School of Medicine. The requirement for informed consent was waived by the Ethics Committee due to the retrospective nature of the study.
Consent for publication
Not applicable.
Declaration of generative AI and AI-assisted technologies in the writing process
During the preparation of this work, the author(s) used DeepSeek to improve language clarity and fluency. After using this tool, the authors reviewed and edited the content as needed, and took full responsibility for the content of the publication.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Haixing Zhu, Dake Shi, Guangwei Li, Lijuan Wu and Xiaoqian Ma contributed equally to this work.
Yumin Xu and Yun Feng contributed equally to this work.
Contributor Information
Yumin Xu, Email: xym121@163.com.
Yun Feng, Email: fy01057@163.com.
References
- 1.Arshad N, Azzam W, Zilberberg MD, Shorr AF. Acinetobacter baumannii complex infections: new Treatment options in the antibiotic pipeline. Microorganisms. 2025;13(2):356. 10.3390/microorganisms13020356. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Satlin MJ, Lewis JS, Weinstein MP, et al. Clinical and laboratory standards institute and European committee on antimicrobial susceptibility testing position statements on polymyxin B and colistin clinical breakpoints. Clin Infect Dis off Publ Infect Dis Soc Am. 2020;71(9):e523–29. 10.1093/cid/ciaa121. [DOI] [PubMed]
- 3.Guzman L, Rabanal F, Garcia J, et al. B-rich colistin and B-pure colistin as novel strategies to increase the therapeutic window of polymyxin antibiotic therapy. Biomed Pharmacother Biomedecine Pharmacother. 2025;190:118366. 10.1016/j.biopha.2025.118366. [DOI] [PubMed] [Google Scholar]
- 4.Jeon CH, Kim SH, Kim H, Park KJ, Wi YM. Ineffectiveness of colistin monotherapy in treating carbapenem-resistant Acinetobacter baumannii pneumonia: a retrospective single-center cohort study. J Infect Public Health. 2024;17(5):774–79. 10.1016/j.jiph.2024.03.007. [DOI] [PubMed] [Google Scholar]
- 5.Sisay M, Hagos B, Edessa D, Tadiwos Y, Mekuria AN. Polymyxin-induced nephrotoxicity and its predictors: a systematic review and meta-analysis of studies conducted using RIFLE criteria of acute kidney injury. Pharmacol Res. 2021;163:105328. 10.1016/j.phrs.2020.105328. [DOI] [PubMed] [Google Scholar]
- 6.Bu W, Wang C, Wu Y, et al. Efficacy and safety of polymyxin B sulfate versus colistin sulfate in ICU patients with nosocomial pneumonia caused by carbapenem-resistant acinetobacter baumannii: a multicenter, propensity score-matched, real-world cohort study. BMC Infect Dis. 2025;25(1):390. 10.1186/s12879-025-10773-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Jantti T, Tarvasmaki T, Harjola VP, et al. Predictive value of plasma proenkephalin and neutrophil gelatinase-associated lipocalin in acute kidney injury and mortality in cardiogenic shock. Ann Intensive Care. 2021;11(1):25. 10.1186/s13613-021-00814-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Zhao Y, Chang W, Li X, Su P, Zhang H, Zhang J. Early Diagnosis of kidney injury via ultrasensitive detection of urine-derived exosomes using a dual-signal amplification strategy. ACS Sens. 10.1021/acssensors.5c00287. Published online September 25, 2025. [DOI] [PubMed]
- 9.Huang F, Thokerunga E, He F, Zhu X, Wang Z, Tu J. Research progress of the application of mesenchymal stem cells in chronic inflammatory systemic diseases. STEM Cell Res Ther. 2022;13(1):1. 10.1186/s13287-021-02613-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Hafkamp F, Dekker L, Tio R, et al. Characterizing high risk patients in Heart Failure: a latent class analysis of rehospitalization and mortality. J Cardiovasc Nurs. 2025;40(5):E297–307. 10.1097/JCN.0000000000001208. [DOI] [PubMed]
- 11.Matsuoka Y, Sotomi Y, Nakatani D, et al. Phenotypic trajectories from acute to Stable phase in Heart Failure with preserved ejection fraction: insights from the PURSUIT-HFpEF study. J Am Heart Assoc. 2025;14(3):e037567. 10.1161/JAHA.124.037567. [DOI] [PMC free article] [PubMed]
- 12.Okuyan O, Elgormus Y, Sayili U, Dumur S, Isik OE, Uzun H. The effect of virus-specific vaccination on laboratory infection markers of children with acute rotavirus-associated acute gastroenteritis. Nato Adv Sci Inst Se. 2023;11(3). 10.3390/vaccines11030580. [DOI] [PMC free article] [PubMed]
- 13.Sarkar S, Kannan S, Khanna P, Singh AK. Role of platelet-to-lymphocyte count ratio (PLR), as a prognostic indicator in COVID-19: a systematic review and meta-analysis. J Med Virol. 2022;94(1):211–21. 10.1002/jmv.27297. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Xiao W, Lu Z, Liu Y, et al. Influence of the Initial neutrophils to lymphocytes and platelets ratio on the incidence and severity of sepsis-associated acute kidney injury: a Double robust estimation based on a large public database. Front Immunol. 2022;13:925494. 10.3389/fimmu.2022.925494. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Iovleva A, McElheny CL, Fowler EL, et al. In vitro activity of sulbactam-durlobactam against colistin-resistant and/or cefiderocol-non-susceptible, carbapenem-resistant acinetobacter baumannii collected in U.S. hospitals. In: Carattoli A, editor. Antimicrob agents chemother. Vol. 68(3): 2024. p. e01258–23. 10.1128/aac.01258-23. [DOI] [PMC free article] [PubMed]
- 16.Gharaibeh MH, Abandeh YM, Elnasser ZA, Lafi SQ, Obeidat HM, Khanfar MA. Multi-drug resistant acinetobacter baumannii: phenotypic and genotypic resistance profiles and the associated risk factors in teaching hospital in Jordan. J Infect Public Health. 2024;17(4):543–50. 10.1016/j.jiph.2024.01.018. [DOI] [PubMed] [Google Scholar]
- 17.Tozluyurt A. Molecular typing of reduced susceptibility of acinetobacter calcoaceticus-baumannii complex to chlorhexidine in Turkey by pulsed-field gel electrophoresis. J Med Microbiol. 2024;73(8). 10.1099/jmm.0.001882. [DOI] [PubMed]
- 18.Tsuji BT, Pogue JM, Zavascki AP, et al. International consensus guidelines for the optimal use of the polymyxins: endorsed by the American college of clinical pharmacy (ACCP), European society of clinical microbiology and infectious diseases (ESCMID), infectious diseases society of America (IDSA), international society for anti-infective pharmacology (ISAP), society of critical care medicine (SCCM), and society of infectious diseases pharmacists (SIDP). Pharmacother J Hum Pharmacol Drug Ther. 2019;39(1):10–39. 10.1002/phar.2209. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Liang T, Chen J, Xu G, et al. Platelet-to-lymphocyte ratio as an Independent factor was associated with the severity of ankylosing spondylitis. Front Immunol. 2021;12:760214. 10.3389/fimmu.2021.760214. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Ma M, Li G, Zhou B, et al. Comprehensive analysis of the association between inflammation indexes and complications in patients undergoing pancreaticoduodenectomy. Front Immunol. 2023;14:1303283. 10.3389/fimmu.2023.1303283. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Chaudhary M, Kumar D, Meena DS, et al. Effectiveness of various sulbactam-based combination antibiotic therapy in the management of ventilator-associated pneumonia caused by carbapenem-resistant Acinetobacter baumannii in a tertiary care Health centre. Indian J Med Microbiol. 2024;52:100737. 10.1016/j.ijmmb.2024.100737. [DOI] [PubMed] [Google Scholar]
- 22.Liu H. Analysis on the clinical pulmonary infection score on the detection of multidrug resistance organisms in lower respiratory tract in ventilated patients in intensive care unit. Zhongguo Wei Zhong Bing Ji Jiu Yi Xue Chin Crit Care Med Zhongguo Weizhongbing Jijiuyixue. 2012;24(11):680–82. [PubMed] [Google Scholar]
- 23.Bian X, Li M, Liu X, et al. Transcriptomic investigations of polymyxins and colistin/sulbactam combination against carbapenem-resistant Acinetobacter baumannii. Comput Struct Biotechnol J. 2024;23:2595–605. 10.1016/j.csbj.2024.05.043. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The de-identified dataset generated and analyzed during this study is not publicly available due to institutional patient privacy regulations. However, anonymized data may be made available to qualified researchers upon reasonable request, subject to approval by the Independent Ethics Committee (IEC) of Ruijin Hospital Affiliated Shanghai Jiao Tong University School of Medicine. Requests should be directed to ruijincrc@126.com and must include a detailed research proposal and compliance with data protection agreements.