Unveiling the immuno-inflammatory endotype of sepsis-associated acute kidney injury: a machine learning-based approach

Article information

Korean J Nephrol. 2026;.j.krcp.26.112

Publication date (electronic) : 2026 August 28

doi : https://doi.org/10.23876/j.krcp.26.112

Jin Zhang, Pengchao Tian, Jingyi Wang, Wenxiong Li, Na Cui

Department of Critical Care Medicine, Beijing Chao-Yang Hospital, Capital Medical University, Beijing, China

Correspondence: Na Cui Department of Critical Care Medicine, Beijing Chao-Yang Hospital, Capital Medical University, 8 Gongren Tiyuchang Nanlu, Chaoyang District, Beijing 100020, China. E-mail: cyyy_cn@163.com

Received 2026 March 9; Revised 2026 June 16; Accepted 2026 July 13.

Abstract

Introduction

Sepsis is a life-threatening condition characterized by a dysregulated host response to infection leading to organ dysfunction [1]. The kidney is particularly vulnerable, and sepsis-associated acute kidney injury (SA-AKI) is common and associated with poor prognosis in chronic kidney disease [2,3]. In routine practice, SA-AKI is diagnosed using the Kidney Disease: Improving Global Outcomes (KDIGO) criteria based on serum creatinine and urine output; however, these functional markers often lag behind the onset of tissue injury and may miss the optimal window for early risk-guided intervention. Numerous models for early prediction of SA-AKI have been proposed, many of which rely solely on clinical variables and demonstrate limited predictive accuracy [4,5].

The pathophysiology of SA-AKI extends beyond simple hypoperfusion; it involves a complex interplay of microcirculatory dysfunction, metabolic reprogramming, and a dysregulated immuno-inflammatory response [6]. Sepsis triggers cytokine storms, endothelial damage, and thromboinflammation, which disrupt renal microvasculature. Concurrently, complement activation and lymphocyte dysfunction (e.g., apoptosis and exhaustion) heighten organ vulnerability [7,8]. Previous studies showed that isolated immuno-inflammatory biomarkers alone, such as inflammatory cytokines, complement components, and T lymphocyte subsets, were associated with a higher risk of acute kidney injury (AKI) and a worse prognosis [9,10]. Few studies have systematically integrated immuno-inflammatory phenotype with clinical variables in risk prediction, potentially missing the synergistic predictive value of combined immune and clinical factors.

To fill the gap, we integrated immuno-inflammatory features with clinical variables to develop an Extreme Gradient Boosting (XGBoost) model for SA-AKI prediction in septic patients without preexisting renal dysfunction and to assess its prediction performance and its usefulness in clinical decision-making.

Methods

Data processing and feature selection

A total of 17 variables with missing values exceeding 50% were initially excluded. The remaining 95 variables, comprising 63 clinical and 32 immune variables, were kept for further processing. For these variables, if they had missing values <10%, they were imputed using the median or mean, depending on their distribution. If they had missing values within 10%–20%, or if they were categorical variables, these variables were imputed using the missForest package (version 1.6) in R version 4.5.1 (R Foundation for Statistical Computing). After imputation, a three-step selection procedure was conducted. All variables were initially selected based on clinical significance and their inclusion in the risk scores, and then their associations with SA-AKI and collinearity were considered. In this procedure, 26 clinical variables and 19 immune variables were excluded, and a total of 37 clinical variables and 13 immune variables with p-values <0.2 in univariable logistic regression and no significant collinearity (variance inflation factor, <10 and Spearman’s coefficient ≤0.7) were retained for further selection via the support vector machine algorithm using 10-fold cross-validation (Fig. 1). Ultimately, five clinical features and four immune features were chosen for model development (Fig. 1).

Figure 1.

Flowchart for feature selection and model development.

APACHE II, Acute Physiology and Chronic Health Evaluation II; AUC, area under the curve; AUPRC, area under the precision-recall curve; C3, complement component 3; CD, cluster of differentiation; GAP, central venous-to-arterial carbon dioxide partial pressure difference; IgA, immunoglobulin A; IL-8, interleukin-8; IL-10, interleukin-10; PCT, procalcitonin; PD1, programmed cell death protein 1; PT, prothrombin time; ScvO₂, central venous oxygen saturation; SOFA, Sequential Organ Failure Assessment; TNF-α, tumor necrosis factor alpha; VIF, variance inflation factor; XGBoost, Extreme Gradient Boosting.

Results

Cohort characteristics and outcomes

Among 1,551 septic patients with preserved baseline renal function, 695 (44.8%) developed incident SA-AKI within 7 days. Patients who developed SA-AKI were in worse clinical condition, had higher inflammatory burden, and more coagulopathy (Table 1). They also had a higher 28-day mortality (20.0% vs. 15.1%) and longer ICU length of stay (Table 1). Renal function at sepsis diagnosis and AKI diagnosis, and time from sepsis to AKI onset were shown in Supplementary Table 1 (available online).

Table 1.

Baseline characteristics of septic patients stratified by SA-AKI

Immuno-inflammatory profile associated with sepsis-associated acute kidney injury

Patients who developed SA-AKI showed a coherent immuno-inflammatory pattern; thus, they were classified as a high-risk endotype combining hyperinflammation, complement consumption, and T-cell dysregulation (Table 2), such as lower total lymphocytes and total T cells, depletion of the naïve CD4⁺ T-cell fraction, and a small but significant relative increase in the CD28⁺CD4⁺ T-cell percentage.

Table 2.

Immuno-inflammatory features of septic patients

Model development and performance comparison

The comprehensive performance of the clinical and clinical-immune models is summarized in Table 3. The clinical-immune model showed higher prediction accuracy, precision, and better discrimination (Fig. 2A, B) and calibration (Fig. 3A), underscoring the incremental value of immune features. The combined model also showed better clinical performance than the clinical variables alone across a broad range of threshold probabilities (Fig. 2C).

Table 3.

Predictive performance of the clinical and clinical-immune XGBoost models in the overall cohort and subgroups

Figure 2.

Model prediction performance and clinical usefulness of the clinical-immune model and clinical model.

Clinical-immune model showed good discrimination performance by receiver operating characteristic (A) and precision-recall curves (B) and higher clinical net benefit (C).

AUC, area under the receiver operating characteristic curve.

Figure 3.

Calibration and distribution of predicted probabilities for the clinical and clinical-immune models.

(A) Calibration curves comparing the agreement between the predicted probabilities and observed incidence of sepsis-associated acute kidney injury (SA-AKI) for the clinical and clinical–immune models; the dashed diagonal line represents perfect calibration. (B) Distribution of SA-AKI probabilities predicted by the clinical model, with patients ranked in ascending order of predicted probability and colored according to their observed outcomes. (C) Distribution of SA-AKI probabilities predicted by the clinical–immune model, with patients ranked in ascending order of predicted probability and colored according to their observed outcomes.

To further visualize risk stratification, prediction distribution plots were generated (Fig. 3B, C). The clinical-immune model exhibited better separation (Fig. 3C), with patients who developed SA-AKI clustered at higher predicted-probability levels than in the clinical model (Fig. 3B).

Model explanation

As the clinical-immune model demonstrated superior predictive performance, its feature importance was further analyzed to enhance interpretability. The feature importance plot lists the enrolled variables in descending order according to their contribution to the model (Fig 4A). Procalcitonin (PCT) was identified as the strongest predictive feature, followed by the APACHE II score. Notably, immuno-inflammatory features played a crucial role in the model: IgA and naïve CD4⁺ T% ranked third and fourth, respectively, superior to the classic SOFA score.

Figure 4.

Feature importance and SHAP analysis of the clinical-immune model.

Feature importance plot of the clinical-immune model (A) and the explanation by SHAP analysis (B). SHAP analysis was helpful in understanding the direction of a specific feature to a higher or lower prediction, and the color indicates whether that feature is high (in purple) or low (in yellow) for that observation.

APACHE II, Acute Physiology and Chronic Health Evaluation II; C3, complement component 3; CD, cluster of differentiation; IgA, immunoglobulin A; PCT, procalcitonin; PT, prothrombin time; SHAP, SHapley Additive exPlanations; SOFA, Sequential Organ Failure Assessment.

SHAP analysis was used to better understand the individual contributions and directions of effect of each feature on SA-AKI risk (Fig. 4B). The analysis showed that higher levels of PCT, APACHE II score, SOFA score, IgA, and CD28⁺CD4⁺ T% were associated with a higher risk of SA-AKI. Interestingly, patients with lower naïve CD4⁺ T% and C3 levels were inclined to develop SA-AKI. The others were shown in Fig. 4B.

Discussion

In this study, we highlight the importance of immune dysregulation for predicting incident SA-AKI and demonstrate that incorporating early immune features provides incremental value for risk stratification. To our best knowledge, our study was the first to identify a distinct immuno-inflammatory phenotype—characterized by elevated PCT and IgA, reduced C3, depleted naïve CD4⁺ T cells, and expanded CD28⁺CD4⁺ T cells—as a critical determinant of SA-AKI risk. Our clinical-immune model achieved satisfactory discrimination (AUC, 0.914) and exceeded most previously reported SA-AKI prediction models [5,15,16]. Our work is also among the first to systematically integrate humoral, complement, and cellular immune features into a unified predictive framework for SA-AKI, highlighting the substantial added value of immune dysregulation in early risk stratification.

Our finding that lower C3 predicts SA-AKI could be partly explained by consumptive depletion resulting from systemic complement hyperactivation. Clinically, low C3 serves not merely as a marker of deficiency but as a severity signal associated with septic shock and mortality [17]. This “consumption hypothesis” is supported by a recent multicenter study by Henry et al. [18], which observed that, in coronavirus disease 2019-induced sepsis, signs of complement consumption (including lower C3 and higher activation ratios) significantly predicted the development of severe AKI. Furthermore, recent evidence confirms that urinary markers of alternative pathway activation correlate with AKI severity, linking systemic C3 consumption to local renal damage [10,19].

Beyond the complement system, immunoglobulins act as critical effectors in sepsis pathogenesis, orchestrating mucosal defense, pathogen clearance, and inflammatory regulation. Consequently, aberrant immunoglobulin levels are closely linked to disease severity and prognosis [20]. In our study, we observed significantly elevated levels of both IgG and IgA in patients with SA-AKI; notably, IgA emerged as a pivotal independent predictor. This finding mirrors the “high-Ig phenotype” identified in the ALBIOS trial, which was associated with increased mortality [20]. Mechanistically, this elevation signals the humoral response to sepsis: severe infection triggers massive polyclonal B-cell activation and gut barrier dysfunction, resulting in a systemic surge of immunoglobulins. Crucially, IgA is not inert; high levels of circulating IgA can directly activate the alternative complement pathway and serve as a potent inducer of neutrophil extracellular trap (NET) formation [21]. These NETs subsequently drive immunothrombosis and endothelial activation [22]. Therefore, the specific profile of “Low C3 and High IgA” likely represents a pathogenic synergy: aberrant humoral activation accelerates complement consumption and NET-driven microvascular stress, rendering the kidney highly vulnerable to inflammatory injury.

A unique contribution of our study is the identification of a distinct T-cell phenotype associated with SA-AKI: depleted naïve CD4⁺ T cells alongside an expanded proportion of CD28⁺CD4⁺ T cells. The reduction in naïve CD4⁺ T cells in our cohort is consistent with the well-described sepsis-induced lymphopenia, characterized by preferential loss of CD4^⁺ T cells—particularly naïve precursors—and long-lasting immune paralysis [23,24]. With respect to the kidney, Guo et al. [25] recently demonstrated that absolute counts of CD3⁺, CD4⁺, and CD8⁺ T cells are markedly diminished in septic patients with AKI, serving as independent risk predictors. More granularly, our finding of specific naïve cell depletion is strongly corroborated by Wang et al. [26], who identified a significant reduction in naïve CD4⁺ T subsets as a key feature distinguishing SA-AKI from non-SA-AKI patients. These studies reinforce the notion that naïve cell depletion may impair adaptive immunity and tissue repair, increasing renal vulnerability [26,27]. Furthermore, this depletion may remove a critical regulatory “brake” on innate immunity. As Zhang et al. [28] demonstrated, a sufficient naïve T-cell pool is essential to constrain toll-like receptor–mediated hyperinflammation; thus, their loss likely unleashes an uncontrolled cytokine storm that inflicts direct collateral damage on the kidney. In stark contrast to this depletion, we observed that a higher percentage of CD28⁺CD4⁺ T cells was associated with SA-AKI. In the acute phase of sepsis, high CD28 expression likely reflects hyper-responsiveness and an increased potential for activation [27,29]. Given that CD28 co-stimulation may play a critical role in maximal cytokine production, an expanded CD28⁺ pool may be associated with an exaggerated inflammatory response [30,31]. This hypothesis is strongly supported by Singbartl et al. [32], who demonstrated that CD28 co-stimulation is indispensable for mediating neutrophil-dependent acute renal failure during endotoxemia. Overall, these data suggested a complex and divergent cellular immune pattern in early SA-AKI. Thus, both profound immunosuppression (depleted naïve cells) and a hyper-activatable effector pool (high CD4⁺CD28⁺ T%) could contribute to kidney injury.

In addition to immune markers, we found that PCT, APACHE II, SOFA, and prothrombin time (PT) were all associated with SA-AKI. Consistent with previous studies [33,34], APACHE II and SOFA scores quantify the cumulative burden of organ dysfunction, serving as the pathophysiological “substrate” for kidney injury. However, in this study, PCT contributed most to SA-AKI prediction, surpassing the traditional severity score. It should not be unexpected, as PCT could not only serve as a surrogate for the severity of infection but also be associated with an intense systemic inflammatory response (i.e., “cytokine storm”) [35,36]. In this context, PCT likely quantifies the magnitude of the acute “infection hit” sustained by the kidney, complementing the prediction performance of the organ dysfunction scores [35,36]. Furthermore, the inclusion of prolonged PT implies the critical role of sepsis-associated coagulopathy. This elevation reflects a state of consumption coagulopathy and hints at “microvascular arrest” driven by disseminated fibrin deposition [37,38], consistent with the findings of Hoshino et al. [9], who had shown that elevated coagulation markers were strongly associated with poor sepsis outcomes. Thus, PT serves as an accessible potential hint for thromboinflammation and microcirculatory derangement within the renal vasculature.

To visualize the complex biological interactions within our model, we performed a correlation analysis among the selected features (Fig. 5). Complement C3 displayed consistent negative correlations with PCT (ρ = –0.29) and PT (ρ = –0.34). This inverse relationship suggests that complement depletion is not an isolated event but is integral to a pathogenic amplification network driven by systemic inflammatory load (PCT) and coagulopathy (PT). Most significantly, the heatmap visually encapsulates the “complex and divergent cellular immune pattern.” Interestingly, naïve CD4⁺ T% and CD28⁺ CD4⁺ T% showed a moderate positive correlation, likely reflecting that both subsets partially track the overall status of the T-cell pool. However, their associations with SA-AKI diverged sharply in direction: Naïve CD4⁺ T% was negatively correlated (ρ = –0.08, implying that depletion promotes injury), whereas CD28⁺ CD4⁺ T% was positively correlated (ρ = 0.06, implying that expansion promotes injury). This functional divergence amid quantitative correlation highlights a critical insight: SA-AKI risk is driven not just by lymphopenia, but by a qualitative shift in which a hyper-responsive effector pool (High CD28) persists or expands relative to a depleted regulatory naïve reserve. This aligns with experimental evidence suggesting that while naïve cell loss unleashes innate inflammation [28], acute CD28 co-stimulation actively amplifies renal tissue injury [32].

Figure 5.

Correlation heatmap of variables included in the clinical-immune model.

The analysis was performed on the overall study cohort of 1,551 septic patients with preserved baseline renal function. The matrix displays the correlation coefficients (ρ) between the outcome (SA-AKI) and the nine selected predictors (5 clinical and 4 immune features). The numerical values within each cell represent the correlation coefficient. Colors indicate the direction and strength of the correlation (Red indicates positive correlation; Blue indicates negative correlation).

APACHE II, Acute Physiology and Chronic Health Evaluation II; C3, complement component 3; CD, cluster of differentiation; IgA, immunoglobulin A; PCT, procalcitonin; PT, prothrombin time; SA-AKI, sepsis-associated acute kidney injury; SHAP, SHapley Additive exPlanations; SOFA, Sequential Organ Failure Assessment.

In summary, the superior performance of our clinical-immune model compared to classical scoring systems is grounded in its ability to encapsulate the multidimensional heterogeneity of sepsis, effectively intertwining the pathophysiological “substrate” of illness severity with specific “immune drivers.” Our results showed SA-AKI to be a composite outcome driven by convergence of “severity substrate,” the “inflammation-complement-coagulation axis,” and “T-cell remodeling.” In addition, the deployment of the XGBoost algorithm allowed us to capture the complex, nonlinear interactions among these heterogeneous biological systems, while SHAP analysis validated the major inputs of immune markers to the model’s decision-making and the fact that immune markers were central risk drivers, not just “bystanders.”

From a clinical standpoint, the proposed framework’s feasibility and ease of use are deeply rooted in existing intensive care infrastructure and contemporary digital health tools. Logistically, although the model incorporates multiple advanced immunological metrics, these variables can be effortlessly obtained via a single standard venipuncture upon ICU admission. In modern tertiary centers, these laboratory assays are fully automated, standardized, and routinely processed, imposing no additional procedural or monitoring burden on either clinicians or patients [8]. Furthermore, to optimize bedside usability and minimize the cognitive load for busy ICU physicians, the complex mathematical backend of the XGBoost model can be seamlessly translated into an intuitive, web-based graphical interface (such as an R Shiny application) or directly integrated into the hospital’s electronic health record (EHR) system [39]. Clinicians can instantly obtain a personalized, real-time risk probability at the bedside, either by manually entering values on a user-friendly webpage or by automated retrieval of EHR data. This high level of practical usability effectively converts complex machine-learning outputs into a direct, actionable decision-support tool, enabling timely, fluid optimization and proactive avoidance of nephrotoxic medications [2]. Nevertheless, we acknowledge that the availability and turnaround time of immune profiling may vary across institutions; therefore, the current model may be most readily implemented in tertiary or academic ICUs with established immune-monitoring capacity.

Our study has several limitations. First, although patients were prospectively screened and enrolled, the prediction model was developed retrospectively. To avoid significantly shrinking the training sample size and potential overfitting, only internal 5-fold cross-validation was conducted, without external validation. Future prospective multicenter studies with in-depth immuno-inflammatory profiling are needed to externally validate and confirm these results and assess the full spectrum of SA-AKI immunopathology. Second, SA-AKI was defined solely by serum creatinine criteria because hourly urine output data were not routinely collected, which may have underestimated the incidence as reported before. Third, our analysis was based on static immuno-inflammatory markers rather than on dynamic ones; evidence on the impact of dynamic changes on prognosis was lacking. Fourth, medication exposure represents a ubiquitous confounding factor in the intensive care setting. Although we evaluated and adjusted for broad nephrotoxic drugs at baseline, we could not further analyze the effect of different types of nephrotoxic drugs. Finally, our analysis was clinical and observational; although we observed a strong phenotype integrating these immune markers with SA-AKI, the underlying molecular mechanisms remain to be elucidated via animal or viral experiments.

In conclusion, immune dysregulation plays an important role in developing AKI in sepsis patients. Immune biomarkers could significantly improve early identification of high SA-AKI risk patients.

Supplementary Materials

Article information Continued