From glycemic variability to digital signal biomarker: a prognostic and precision medicine framework for sepsis-associated acute kidney injury

Abstract

Sepsis, a condition with substantial global morbidity and mortality, frequently leads to sepsis-associated acute kidney injury (SA-AKI). While glycemic variability (GV) correlates with adverse outcomes in critically ill populations, its prognostic value in SA-AKI remains underexplored. Using the MIMIC-IV database, this large-scale machine learning cohort study examined SA-AKI patients. Restricted cubic spline, Kaplan-Meier analysis, and Cox regression analyses were conducted to evaluate associations between GV, measured by glycemic coefficient of variation (CV) and 28- and 90-day mortality. Subgroup analyses stratified by age, sex, and diabetes status were performed. Prognostic models were developed using Cox proportional hazards (CoxPH), Least absolute shrinkage and selection operator (LASSO), and random survival forests (RSF). Among 12,268 eligible SA-AKI patients, the Boruta algorithm identified glycemic CV as a key prognostic determinant. When stratified by CV quartiles, higher CV quartiles exhibited significantly increased 28-day and 90-day mortality. Subgroup analyses revealed consistent associations except in diabetic patients, where increases in CV showed no correlation with mortality. Machine learning models exhibited strong predictive performance, with 28-day area under the curves (AUCs) of 0.822 (CoxPH), 0.822 (LASSO), and 0.845 (RSF), and 90-day AUCs of 0.819 (CoxPH), 0.820 (LASSO), and 0.837 (RSF). Elevated GV is associated with increased short- and long-term mortality in SA-AKI. Beyond prognostication, these findings position GV as a real-time, modifiable digital biomarker that may underpin a mechanistic Glycemic Oscillation–Induced Renal Injury (GO-RI) Axis. This framework supports future development of machine learning–enabled precision ICU nephrology strategies for dynamic risk stratification and phenotype-specific glycemic modulation in SA-AKI.

Introduction

Sepsis, a life-threatening organ dysfunction syndrome resulting from maladaptive host responses to infection, is characterized by pathophysiological features including systemic inflammatory dysregulation, immune compromise, and aberrant metabolic processes, often progressing to multiple organ system failure. This disorder represents a major contributor to the global disease burden, accounting for 19.7% of total international mortality [1,2], with pronounced impacts on critically ill patients in intensive care units (ICUs). An epidemiological study involving 22,748 ICU-admitted demonstrated sepsis prevalence and case fatality rates of 25.5% and 40%, respectively [3].

Sepsis-associated acute kidney injury (SA-AKI) is clinically defined as acute kidney injury (AKI) occurring within 7 days following sepsis onset [4] Among the various multiorgan injuries caused by sepsis, SA-AKI represents one of the most common complications [5,6], A multicenter retrospective study revealed that up to 60% of patients with septic shock progress to AKI [7]. In comparison to non-sepsis-related AKI, SA-AKI is associated with a significantly higher risk of adverse clinical outcomes [8], highlighting the critical importance of early assessment and standardized management in this patient group.

Although the pathophysiology of sepsis-associated acute kidney injury (SA-AKI) is not yet fully understood, several mechanisms, including inflammation, dysregulation of the renin-angiotensin-aldosterone system, complement activation, alterations in hemodynamics, and mitochondrial dysfunction, are believed to contribute to the development of AKI [4,9]. Recent studies have highlighted that glucose dysregulation is a significant component in the progression of both sepsis and SA-AKI. The proposed mechanisms are as follows: on the one hand, sepsis and AKI can induce stress-related hyperglycemia, hypoglycemia, and insulin resistance, leading to increased blood glucose instability and exacerbating kidney injury [10–12]. On the other hand, acid-base imbalances, the accumulation of uremic toxins, and a decline in renal insulin clearance caused by SA-AKI can further intensify glucose fluctuations [13].

Glycemic variability(GV) has been shown to be strongly associated with adverse outcomes in critically ill patients, including those with sepsis and AKI [14–17]. However, its prognostic value in the SA-AKI subgroup remains to be confirmed. Utilizing the Medical Information Mart for Intensive Care (MIMIC) repository, we performed a retrospective analysis to explore the relationship between GV during ICU stays and the 28-day and 90-day mortality risks in SA-AKI patients. Subsequently, we developed a predictive model for clinical prognosis by incorporating routine clinical biomarkers.

Methods

Data source

This retrospective cohort study utilized data from MIMIC-IV v3.1 database [18,19], which contains deidentified clinical records of patients admitted to emergency departments or intensive care units at Beth Israel Deaconess Medical Center (BIDMC), Boston, Massachusetts, between 2008 and 2019. The BIDMC Institutional Review Board granted a waiver of informed consent and approved secondary data analysis. Author X.M. successfully completed the Collaborative Institutional Training Initiative (CITI program) and obtained credentialed database access (ID:14232440) prior to data extraction.

Participants and definitions

This study used Navicat Premium software (version 15) for patient selection and variable extraction from MIMIC-IV v3.1 database. Patients meeting sequential inclusion criteria were enrolled [1]: diagnosed with sepsis according to the recommendations of the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3) [2,20] development of AKI within 48 h of sepsis diagnosis, with AKI defined and classified according to the KDIGO criteria [3,21]; age > 18 years. Exclusion criteria were [1]: patients with less than 48 h of ICU stay [2]; patients with multiple ICU admissions (only the first ICU stay was included). Extracted variables encompassed demographic characteristics (sex, age, race), vital signs (Systolic Blood Pressure [SBP], Diastolic Blood Pressure [DBP], heart rate (HR), temperature, body mass index, blood tests (white blood cells [WBCs]), hemoglobin (Hb), platelets (PLT), neutrophils, lymphocytes, monocytes), hepatic/renal biomarkers (albumin, alanine aminotransferase, aspartate aminotransferase, blood urea nitrogen (BUN), creatinine), electrolyte/metabolic parameters (sodium, potassium, calcium, glucose, pH, lactate), coagulation indices (international normalized ratio, prothrombin time [PT], activated partial thromboplastin time [PTT]), Hemoglobin A1c (HbA1c), comorbidities (diabetes mellitus, chronic kidney disease [CKD], congestive heart failure [CHF], chronic pulmonary disease, moderate/severe liver disease, malignancy, cerebrovascular disease), therapeutic interventions (statin use, aspirin administration, antidiabetic medications, mechanical ventilation, and continuous renal replacement therapy [CRRT]). For time-dependent variables with repeated ICU measurements, all recorded values were extracted, and maximum and minimum values were calculated for analysis. GV was quantified through the glycemic coefficient of variation (CV), which was calculated using the following formula: CV = (Standard Deviation (SD) of Blood Glucose/Mean Blood Glucose) × 100%.

Statistical analysis

Continuous variables are presented as median [interquartile range (IQR)] and were tested using the Wilcoxon rank-sum test. Categorical variables are presented as counts (percentages) and analyzed via the Chi-square test. Variables with a missing rate > 20% or a Variance Inflation Factor > 5 were excluded. For all variables (including both continuous and categorical) with less than 20% missingness, we employed Multiple Imputation by Chained Equations (MICE) to handle missing data.

Eligible variables were included in the Boruta algorithm [22] for feature selection. Boruta is a wrapper-based feature selection algorithm based on random forests, with the core idea of determining statistically significant features by comparing the importance of original features with randomly generated ‘shadow features.’ By introducing shadow features and statistical testing mechanisms, the Boruta algorithm provides an efficient and robust feature selection approach, particularly suited for identifying key variables in complex datasets.

Patients were stratified into quartiles based on CV values. Kaplan-Meier survival curves, with log-rank tests, were used to compare the 28-day and 90-day survival probabilities across the quartiles. Cox proportional hazards models were applied to estimate hazard ratios (HRs) with 95% confidence intervals (CIs). Model 1 assessed the unadjusted effects of CV, while Model 2 included full adjustment, incorporating Boruta-selected features. Subgroup analyses were stratified by age, sex, and diabetes status. Restricted cubic splines (RCS) were utilized to examine potential nonlinear relationships between continuous CV values and survival outcomes, with adjustment for all important variables identified by the Boruta algorithm.

Finally, the Cox proportional hazards (CoxPH), Least absolute shrinkage and selection operator (LASSO), and random survival forests (RSF)were developed to predict the 28-day and 90-day mortality risks in SA-AKI patients. All machine learning models were implemented using the R programming language, leveraging the mlr3 framework for survival analysis with hyperparameter optimization. To ensure robustness, a grid search strategy with 5-fold cross-validation was applied during hyperparameter tuning. A random seed was set prior to all random processes, including data splitting and model training. The specific implementation is detailed in the code. The dataset was randomly divided into a training set and a validation set in a 7:3 ratio. Model development was conducted exclusively on the training set, while all reported performance metrics were calculated on the held-out validation set. Model discrimination was assessed using the concordance index (C-index) and AUC values derived from ROC curves. Decision curve analysis (DCA) was employed to evaluate clinical utility, and calibration curves were used to assess the accuracy of model predictions for absolute risk.

The statistical analyses were conducted using R version 4.4.2. The following key packages were used: mrl3 (v1.2.0), mlr3verse (v0.3.1), mlr3proba (v0.7.5), mlr3tuning (v1.4.0), mlr3extralearners (v1.0.0), glmnet (4.1.0), Boruta (v9.0.0), survival (v3.8.3), rms (v8.0.0), and mice (v3.18.0). All analyses were performed using transparent, reproducible code under the FAIR principles (Findable, Accessible, Interoperable, Reusable) to enhance model interpretability and replicability [23].

Results

Baseline characteristics of study participants

A total of 12,268 patients were included in the final analysis (Figure 1). Among these, 3,353 (27.3%) and 4,199 (34.2%) experienced mortality within 28-day and 90-day follow-up periods, respectively. At the 28-day follow-up, CV in the survival group was significantly lower than that in the non-survival group (19.79 vs. 22.92, p < 0.001), with no differences between the two groups in terms of diabetes prevalence. Additionally, significant differences were observed between the survival and non-survival groups regarding age, sex, race, vital signs (e.g. blood pressure, respiratory rate, and temperature), maximum and minimum laboratory values (e.g. baseline creatinine and BUN), comorbidities (e.g. CKD), and treatment interventions (e.g. CRRT), as summarized in Table 1 and Supplementary Table S1.

Figure 1.

Flow Diagram of Study Population Selection. SA-AKI, Sepsis-associated acute kidney injury; ICU, intensive care units; MIMIC, Medical Information Mart for Intensive Care; VIF, Variance Inflation Factor.

Table 1.

The basic characteristics of the study participants according to the 28-day situation.

Characteristic	Overall N = 12,268	Survivor N = 8,915	Non-Survivor N = 3,353	P-Value	SMD
Age	69 (58, 79)	68 (57, 78)	72 (61, 82)	<0.001	0.225
Sex				0.042	0.042
Female	5,020 (41%)	3,598 (40%)	1,422 (42%)
Male	7,248 (59%)	5,317 (60%)	1,931 (58%)
AKI stage				<0.001	0.439
Stage 1	2,845 (23%)	2,261 (25%)	584 (17%)
Stage 2	5,821 (47%)	4,529 (51%)	1,292 (39%)
Stage 3	3,602 (29%)	2,125 (24%)	1,477 (44%)
Race				<0.001	0.158
Black	1,072 (8.7%)	782 (8.8%)	290 (8.6%)
Others	3,348 (27%)	2,262 (25%)	1,086 (32%)
White	7,848 (64%)	5,871 (66%)	1,977 (59%)
Scr_baseline	0.7 (0.5, 1.0)	0.7 (0.5, 0.9)	0.7 (0.5, 1.1)	<0.001	0.107
BUN_max	38 (23, 65)	34 (21, 57)	54 (33, 83)	<0.001	0.521
BUN_min	17 (11, 26)	15 (10, 23)	21 (13, 35)	<0.001	0.464
Glucose range	92 (55, 159)	85 (51, 149)	112 (70, 186)	<0.001	0.177
Glucose CV	20.67 (14.43, 29.40)	19.79 (13.84, 28.43)	22.92 (16.42, 31.90)	<0.001	0.202
Chronic kidney disease	3,035 (25%)	2,047 (23%)	988 (29%)	<0.001	0.148
Congestive heart failure	4,367 (36%)	3,011 (34%)	1,356 (40%)	<0.001	0.138
Chronic pulmonary disease	3,459 (28%)	2,464 (28%)	995 (30%)	0.027	0.045
Diabetes	4,057 (33%)	2,955 (33%)	1,102 (33%)	0.8	0.006
Tumor	1,779 (15%)	1,000 (11%)	779 (23%)	<0.001	0.322
Cerebrovascular disease	1,957 (16%)	1,299 (15%)	658 (20%)	<0.001	0.135
Liver disease	2,263 (18%)	1,374 (15%)	889 (27%)	<0.001	0.275
CRRT	1,544 (13%)	784 (8.8%)	760 (23%)	<0.001	0.388
Mechanical ventilation	10,244 (84%)	7,424 (83%)	2,820 (84%)	0.3	0.022

Categorical variables are presented as n (%), and continuous variables are presented as median (interquartile range). Scr: Serum creatinine; AKI: Acute kidney injury; BUN: Blood urea nitrogen; Glucose range, Glucose Fluctuation Range; Glucose CV: Glucose Coefficient of Variation; CRRT: Continuous Renal Replacement Therapy.

Feature selection

Following initial preprocessing, 19 covariates were excluded due to missingness exceeding 20%, and an additional 2 variables were removed due to multicollinearity (VIF > 5). The results of the Boruta algorithm identified 45 variables associated with both the 28-day mortality rate (Figure 2) and the 90-day mortality rate (Supplementary Fig. S1) in SA-AKI patients. These variables include the minimum and maximum values of SBP, DBP, HR, temperature, PT, PTT, serum potassium, serum calcium, serum sodium, pH, lactate, WBC, Hb, PLT, and BUN, as well as baseline serum creatinine, tumor, age, AKI stage, glucose fluctuation range, CV, comorbidities(such as diabetes mellitus, CKD, CHF, moderate/severe liver disease, and cerebrovascular disease), and treatment measures(including CRRT, mechanical ventilation, statin use, and aspirin administration). Notably, sex, race, and chronic pulmonary disease were not included.

Figure 2.

Feature selection for predicting 28-day mortality was performed using the Boruta algorithm.

Kaplan-Meier survival analysis

Participants were stratified into glycemic CV quartiles: Quartile 1 (Q1) (<14.43), Q2 (14.43–20.67), Q3 (20.67–29.40), and Q4 (>29.40). A significant inverse correlation was observed between ascending CV quartiles and survival outcomes, with sequentially declining 28-day (p < 0.001) and 90-day (p < 0.001) survival probabilities (Figure 3). Female participants demonstrated inferior survival rates compared to males (Supplementary Fig. S2). Participants aged <60 years exhibited superior survival outcomes relative to those ≥60 years (Supplementary Fig. S3). No statistically significant survival differences were observed between diabetic and non-diabetic subgroups at either 28-day (p = 0.75) or 90-day (p = 0.64) follow-up intervals (Supplementary Fig. S4).

Figure 3.

Kaplan-Meier survival curves for 28-day (A) and 90-day (B) mortality stratified by glycemic coefficient of variation (CV) quartiles: Q1 (<14.43%), Q2 (14.43–20.67%), Q3 (20.67–29.40%), and Q4 (>29.40%). The shaded areas represent the 95% confidence intervals.

Restricted cubic splines and Cox regression analysis

Restricted cubic spline (RCS) analysis revealed no nonlinear association between glycemic CV and 28-day mortality (P for nonlinear = 0.156), while demonstrating a nonlinear relationship with 90-day mortality (p = 0.015) (Figure 4). Cox proportional hazards models, with quartile 1 (Q1: CV < 14.43) as the reference group, showed progressively elevated risks of 28-day (Figure 5) and 90-day mortality (Supplementary Fig. S5) across ascending CV quartiles in both unadjusted and fully adjusted models. In the fully adjusted models, the hazard ratios (HRs) for mortality significantly increased across quartiles. For 28-day mortality, the adjusted HRs were 1.12 (95% CI 1.00–1.25) for Q2, 1.20 (95% CI 1.07–1.33) for Q3, and 1.22 (95% CI 1.07–1.38) for Q4 (P for trend < 0.001). Correspondingly, for 90-day mortality, the adjusted HRs were 1.12 (95% CI 1.01–1.23) for Q2, 1.24 (95% CI 1.12–1.36) for Q3, and 1.23 (95% CI 1.10–1.38) for Q4 (P for trend < 0.001). Subgroup analyses stratified by age, sex, and diabetes status demonstrated consistent associations: Q4 exhibited significantly increased mortality risks (HR > 1, p < 0.05). across all subgroups except diabetic patients, irrespective of covariate adjustment status. Notably, despite higher CV values in diabetic versus non-diabetic patients (26.51 [IQR 18.22–35.43] vs. 18.51 [IQR 13.37–25.81], p < 0.001), no significant relationship between CV elevation and mortality risk was observed in the diabetic subgroup at either 28-day (P for trend = 0.280) or 90-day (P for trend = 0.793) follow-up.

Figure 4.

Restricted cubic spline (RCS) analysis of the association between glycemic coefficient of variation (CV) and (A) 28-day and (B) 90-day all-cause mortality. The shaded areas represent the 95% confidence intervals. Adjusted for age, weight, sex, heart rate, respiratory rate, systolic blood pressure. 95% CI, confidence interval.

Figure 5.

Cox proportional hazards regression analysis of 28-day mortality across population-based quartiles of glycemic coefficient of variation (CV). CV quartile ranges: Q1 (≤14.43%), Q2 (14.44–20.67%), Q3 (20.68–29.40%), and Q4 (≥29.41%). Q1 was used as the reference group. 95% CI, confidence interval.

Establishment and validation of the machine learning prediction model

The ROC curves for CV in predicting 28-day and 90-day all-cause mortality in SA-AKI patients demonstrated limited discriminative capacity, with AUC values of 0.578 and 0.587, respectively (Supplementary Fig. S6). The machine learning algorithms—Cox proportional hazards (CoxPH), random survival forest (RSF), and Least absolute shrinkage and selection operator (LASSO)—were developed to predict 28-day and 90-day mortality risks, utilizing the 45 features selected by the Boruta algorithm for model training. ROC curve analysis demonstrated robust discriminative performance across models (Figure 6), with 28-day AUC values of 0.822 (CoxPH), 0.822 (LASSO), and 0.845 (RSF), and 90-day AUC values of 0.819 (CoxPH), 0.820 (LASSO), and 0.837 (RSF). Concordance indices (C-indices) further validated model accuracy: 0.766 (CoxPH), 0.767 (LASSO), and 0.784 (RSF). The Brier scores indicated strong overall predictive accuracy at 28 and 90 days, respectively: 0.140 and 0.159 (CoxPH), 0.140 and 0.159 (LASSO), 0.135 and 0.155 (RSF). Calibration curves revealed strong agreement between predicted and observed mortality probabilities across all models (Supplementary Fig. S7), which was quantitatively supported by calibration slopes near the ideal value of 1: 1.138 (CoxPH), 1.177 (LASSO), and 1.183 (RSF) at 28 days; 1.118 (CoxPH), 1.163 (LASSO), and 1.167 (RSF) at 90 days. DCA demonstrated positive net benefit and indicated robust clinical effectiveness (Supplementary Fig. S8).

Figure 6.

Receiver operating characteristic curve analysis of the machine learning algorithms. The 28-day AUC values were 0.822 for CoxPH, 0.822 for LASSO, and 0.845 for RSF. The 90-day AUC values were 0.819 for CoxPH, 0.820 for LASSO, and 0.837 for RSF. The C-index values were 0.766 for CoxPH, 0.767 for LASSO, and 0.784 for RSF. CoxPH, Cox proportional hazards; LASSO, Least absolute shrinkage and selection operator; RSF, Survival random forest; T, observation time; AUC: area under the curve.

Discussion

This investigation demonstrated an independent association between elevated glycemic CV and increased 28-day and 90-day mortality in SA-AKI. Stratified analyses by age (>60 vs ≤60 years), sex, and diabetes status revealed consistent CV-mortality gradients across all subgroups (P for trend < 0.05), with the exception of diabetic patients where no significant dose-response relationship was observed. The Boruta algorithm identified CV as a predictor variable within the SA-AKI mortality feature set. While CV demonstrated limited discriminatory capacity (AUC < 0.6), multivariable machine learning frameworks integrating CV with clinical biomarkers demonstrated superior predictive performance (AUC > 0.8). These composite models exhibited excellent calibration accuracy and clinical utility. Beyond demonstrating that CV is an independent predictor of mortality in SA-AKI, this study reframes glycemic variability as a digitally tractable biomarker that can be dynamically integrated into machine learning–driven precision nephrology workflows in the ICU. Rather than functioning solely as a prognostic variable, CV may act as a modifiable signal that informs phenotype-specific glucose management strategies.

Although there is a wide range of indices for quantifying glycemic variability, extensive research has established a significant association between elevated GV and adverse clinical outcomes in critically ill patients [24]. Hermanides et al. demonstrated that heightened GV, measured by the mean absolute glucose change per hour and SD, was independently correlated with increased ICU mortality and in-hospital mortality. Notably, among patients receiving intensive glycemic control therapy, a reduction in GV was associated with a protective effect against the progression of morbidity, even when mean glucose levels remained elevated [25]. In septic populations, comparative analyses revealed a nearly fivefold increase in hospital mortality risk for subjects with higher glycemic lability index (GLI) despite lower mean glucose values, relative to those with lower GLI (OR = 4.73, 95% CI:2.6–8.7) [16]. Furthermore, elevated GV, measured using the mean amplitude of glycemic excursions and CV, during the first 24 h post-ICU admission, was an independent predictor of 30-day mortality in sepsis cohorts (aHR 2.59, 95% CI:1.49–4.50) [15]. By specifically investigating SA-AKI patients, this study elucidated GV’s critical role in this high-risk subpopulation. Multivariable Cox regression revealed significant associations between increases in CV and 28-day mortality risk, with adjusted hazard ratios of 1.22 (95% CI: 1.07–1.38) for 28-day and 1.23 (95% CI: 1.10–1.38) for 90-day mortality (P for trend < 0.01).

The mechanisms through which GV impacts the prognosis of SA-AKI patients require further investigation. Potential pathways may involve multiple interrelated processes. First, GV exacerbates oxidative stress and endothelial injury. Despite differences in the methods used to calculate glycemic variability, numerous in vivo and in vitro studies demonstrate that intermittent glucose fluctuations induce greater oxidative stress and endothelial damage compared to sustained hyperglycemia [26,27]. GV promotes cellular apoptosis and is associated with increased reactive oxygen species production and vascular dysfunction [28,29]. GV promotes cellular apoptosis and is associated with increased reactive oxygen species production and vascular dysfunction [30,31]. Second, the likelihood of hypoglycemia increases with elevated glucose CV, Hypoglycemia may further aggravate organ damage in SA-AKI patients by promoting thrombotic events and enhancing the release of inflammatory cytokines and mediators [32].

We propose a conceptual framework termed the Glycemic Oscillation–Induced Renal Injury Axis, in which recurrent fluctuations in glucose levels amplify oxidative stress, endothelial dysfunction, and immune dysregulation, leading to microvascular instability and accelerated kidney injury in the context of sepsis. Within this axis, glycemic variability serves as a digital signature of metabolic instability, reflecting maladaptive host responses and impaired renal resilience. This aligns with accumulating evidence that oscillatory rather than sustained hyperglycemia triggers disproportionate cellular stress responses. Our data support positioning GV as a hallmark of a ‘Metabolic Instability Phenotype’ in SA-AKI, which may define a subgroup of patients who could benefit from targeted glycemic stabilization strategies.

Notably, while diabetic patients exhibited significantly greater glycemic fluctuations and higher glycemic variability compared to non-diabetic conterparts, stratified analyses revealed increased glycemic variability independently correlated with elevated mortality risk in non-diabetic patients, but showed no significant association in diabetic individuals, aligning with previous epidemiological findings [15,33]. This observation suggests diabetes status may confer protective effects on short-term critical illness outcomes [34]. In our multivariable Cox regression models, diabetic status was associated with reduced 28-day mortality risk and 90-day mortality risk compared to non-diabetic patients. This may be explained by several factors. First, chronic exposure to glycemic fluctuations in diabetes may induce a form of metabolic adaptation or tolerance to oxidative stress and inflammatory cytokine responses [35]. Second, confounding by medication use is plausible; for instance, insulin therapy itself is a source of CV but also a marker of illness severity, and some medications like SGLT2 inhibitors may confer renoprotective effects that confound the relationship. Unfortunately, our dataset lacked detailed outpatient medication records, which prevented a comprehensive exploration of this possibility. Therefore, it would be imperative for future studies to include well-designed clinical and mechanistic research to elucidate the underlying mechanisms.

The glycemic coefficient of variation may serve as a simple, data-driven biomarker for early mortality risk stratification in SA-AKI. Integrating CV-based monitoring into clinical decision-support systems could enable proactive, individualized interventions and enrich future trial designs. As a standardized metric, CV can be automatically computed in real-time from sequential glucose measurements using modern electronic health record systems and bedside monitors, allowing continuous assessment without imposing additional computational burdens on clinicians. By embedding CV-driven alerts within clinical decision support tools, care teams may receive early notifications of high-risk SA-AKI patients with significant glycemic variability, prompting timely actions such as adjusted insulin regimens or intensified hemodynamic surveillance. Furthermore, AI algorithms have demonstrated the ability to enhance early detection, improve risk prediction, personalize treatment strategies, and support clinical decision-making processes in AKI management, the integration of predictive models like ours into AI-driven ICU dashboards represents the next frontier in critical care nephrology [36,37].

As far as we know, this is the first study to investigate the relationship between glycemic CV and survival outcomes in the SA-AKI population. Unlike previous studies on sepsis or AKI cohorts, we employed machine learning predictive models, thereby enhancing the theoretical foundation for the clinical application of CV. Our research offers several notable strengths. First, this large-scale cohort study (n = 12,268) ensured robust statistical power. Multiple statistical approaches were employed to elucidate the significant association between higher CV and 28- and 90-day all-cause mortality, highlighting the critical importance of stringent CV control within the SA-AKI population. Second, CV was calculated using all glucose measurements obtained during ICU admission as an exposure variable. Finally, we developed machine learning predictive models that incorporated CV alongside time-varying clinical biomarkers routinely monitored in the ICU. These models demonstrated superior prognostic performance and clinical feasibility.

However, this study has several limitations. First, as a single-center retrospective investigation, it is inherently susceptible to residual confounding factors. We employed multivariable and stratified analyses to minimize the impact of confounding variables. Second, variables with missing values exceeding 20% were excluded, potentially introducing bias if these variables influenced outcomes. Third, potential for surveillance bias exists; patients with greater illness severity likely had more frequent glucose measurements, which could influence the calculation of CV and potentially confound its association with outcomes. Furthermore, while we adjusted for available covariates, potential confounders such as baseline HbA1c levels and specific details of antidiabetic medications were not fully available in the dataset and could influence the observed relationships. Future research should prioritize multicenter validation to establish generalizable CV thresholds and evaluate the robustness of machine learning models across heterogeneous populations.

Conclusion

This study has demonstrated that a higher CV is significantly associated with increased 28-day mortality and 90-day mortality in SA-AKI patients. However, no significant association was observed between increases in CV and mortality risk in the diabetic subgroup.CV is a dynamic digital biomarker enabling risk-responsive and phenotype-specific glycemic modulation strategies in SA-AKI. Machine learning prognostic models incorporating CV and clinical biomarkers demonstrated robust predictive performance. These findings highlight CV as more than an associated risk factor; it emerges as a real-time, modifiable digital biomarker that can be incorporated into machine learning–enabled precision ICU nephrology platforms to support dynamic risk stratification and individualized glucose control strategies in SA-AKI. Multicenter prospective studies and trials are warranted to validate these findings and establish generalizable CV thresholds across heterogeneous critical care populations.

Funding Statement

This study has received funding by The Guangdong Provincial Department of Science and Technology, Science and Technology Plan Project, Journal of Jinan University High-Level Science and Technology Journal Construction Project [No.2021B121020012].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Ethics approval and consent to participate

This study utilizes the publicly available MIMIC-IV database (version 3.1), which contains fully de-identified clinical records approved for research by the Beth Israel Deaconess Medical Center Institutional Review Boards. No additional ethics approval was required for this secondary analysis of preexisting anonymized data. This study complied with the Declaration of Helsinki and relevant local data protection regulations.

Data availability statement

The datasets generated and analyzed during the current study are available in the Medical Information Mart for Intensive Care (MIMIC)-IV database (https://physionet.org/content/mimiciv/3.1). The R script containing the core analytical code during this study are available in a GitHub repository at https://github.com/MiaWilson-Ma/machinelearning.