Low serum magnesium linked to early sepsis-related kidney injury risk

Critically ill patients with sepsis who arrive in the intensive care unit with lower serum magnesium concentrations may face a higher risk of developing sepsis-associated acute kidney injury within their ICU stay, according to a large retrospective cohort study published in May 2026. Using two independent critical care databases, researchers applied an interpretable machine learning framework and found that serum magnesium measured within the first 24 hours of admission ranked consistently among the most important predictors of the outcome, alongside urine output, SOFA score, and serum creatinine. The study, based on the MIMIC-IV and eICU Collaborative Research Database, included 19,805 adult patients who met the Sepsis-3 criteria.

Sepsis-associated acute kidney injury is a frequent complication in critically ill patients and remains difficult to identify at an early stage. It is independently associated with increased short-term mortality, prolonged hospitalization, and an elevated risk of long-term renal dysfunction. Current diagnostic approaches rely primarily on changes in serum creatinine and urine output, both of which often occur late in the disease course and may fail to capture early renal stress or subclinical injury. Global severity scores such as the Sequential Organ Failure Assessment (SOFA) score mainly reflect established organ dysfunction and provide limited information about early, potentially modifiable risk states. This diagnostic lag is precisely the gap the study set out to probe with routinely measured laboratory parameters.

Magnesium, the second most abundant intracellular cation, plays a critical role in enzymatic reactions, mitochondrial metabolism, and the modulation of inflammatory signaling. Electrolyte disturbances are common in critically ill patients and may influence organ function through effects on vascular tone, cellular energetics, and immune regulation. Altered magnesium homeostasis has been associated with adverse outcomes in critically ill populations, and both hypomagnesemia and hypermagnesemia have been linked to increased morbidity and mortality. Experimental data also suggest that magnesium may modulate pathways involved in oxidative stress and microvascular dysfunction, which are key components of sepsis-related organ injury. Yet the relationship between serum magnesium concentration and the development of sepsis-associated acute kidney injury had remained incompletely characterized, particularly during the early phase of intensive care when preventive strategies may still be feasible.

The retrospective cohort study drew on the Medical Information Mart for Intensive Care IV (MIMIC-IV) and the eICU Collaborative Research Database, both of which contain de-identified data from adult ICU admissions and are approved for research use with a waiver of informed consent. Adult patients aged 18 years or older who met the Sepsis-3 criteria were included, with sepsis defined as suspected infection accompanied by an increase in the SOFA score of at least two points. Only the first ICU admission for each patient was analyzed, and patients with end-stage kidney disease or chronic dialysis dependence before admission were excluded. The primary outcome was the development of acute kidney injury during the ICU stay, defined according to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria based on serum creatinine and urine output.

To minimize the risk of reverse causation, all candidate predictors were extracted from data obtained within the first 24 hours after ICU admission. Variables included demographic characteristics, comorbidities, vital signs, laboratory measurements including serum magnesium, severity scores, and early supportive therapies, with missing data handled using multiple imputation where appropriate. Feature screening combined six machine learning algorithms with Shapley Additive exPlanations (SHAP): CatBoost, Decision Tree, Extra Trees, LightGBM, Random Forest, and XGBoost. A multi-seed strategy for dataset splitting, followed by independent training and testing, was adopted to ensure reproducibility. Although LightGBM showed strong performance during initial screening, CatBoost demonstrated more consistent performance across multiple validation strategies and was selected as the final model. Twelve classification algorithms were then compared across three training-to-test split ratios of 8:2, 7:3, and 6:4, using both cross-validation and independent test sets, with hyperparameters tuned through randomized grid search.

The results converged on a consistent picture. Model performance improved with the sequential inclusion of top-ranked features and plateaued after approximately 10 variables, a pattern that held across repeated data splits. Across repeated analyses, urine output, SOFA score, serum creatinine, and serum magnesium were consistently ranked among the most important predictors, with additional contributions from weight, invasive ventilation, and phosphate. The final model achieved an area under the receiver operating characteristic curve (AUC) of approximately 0.84 when evaluated using all candidate features. Under cross-validation with a reduced feature set, the AUC ranged from approximately 0.67 to 0.72, with similar estimates in the corresponding independent test sets. Performance metrics were similar between training and test datasets, suggesting no evidence of substantial overfitting, although the authors note that the overall discriminative performance of the model was moderate, reflecting the multifactorial nature of the condition.

The most striking finding concerned the shape of the magnesium association. Using a generalized additive model to examine the relationship between individual variables and their SHAP values, the researchers found that serum magnesium demonstrated a distinctly non-linear pattern. Lower magnesium concentrations were associated with higher SHAP values, indicating higher model-predicted risk of acute kidney injury, whereas the contribution to model output appeared to attenuate above approximately 2.0 to 2.1 mg/dL, suggesting a plateau. This pattern was consistently observed when serum magnesium was measured within the first 24 hours of ICU admission. Although mean magnesium levels differed only modestly between patients who developed kidney injury and those who did not, the model-based analyses suggest the association may be driven by non-linear, range-specific patterns rather than differences in central tendency, consistent with a threshold-dependent relationship rather than a simple linear effect.

The machine learning findings were corroborated by conventional statistics. In multivariable logistic regression, serum magnesium remained independently associated with acute kidney injury after adjustment for potential confounders, with an estimated effect size of similar magnitude to several comorbid conditions and a confidence interval that did not include unity. Increasing age and multiple comorbidities, including hypertension, diabetes mellitus, chronic pulmonary disease, cerebrovascular disease, renal disease, peripheral vascular disease, and severe liver disease, were also associated with higher odds of the outcome. Patients who developed kidney injury exhibited lower urine output, higher SOFA scores, and higher serum creatinine levels, and differences were observed in serum magnesium, phosphate, and clinical characteristics such as invasive mechanical ventilation. The team also developed an online prediction tool implementing the final model, allowing users to input variables and obtain model-based predictions.

Despite the convergence of methods, the authors are emphatic that the findings describe a risk association, not a causal or protective effect. The retrospective observational design precludes causal inference, and residual confounding cannot be excluded. Factors known to influence magnesium levels, including diuretic use, fluid resuscitation, nutritional support, and medication exposure, were not fully captured. The use of admission creatinine as a surrogate baseline in some patients may have led to underestimation of AKI incidence, and even with predictors restricted to the early ICU period, reverse causation cannot be completely excluded. The analytical pipeline, while designed to enhance robustness, is also relatively complex and may limit direct clinical interpretability.

The clinical implications are correspondingly measured. Serum magnesium is routinely measured in ICU practice and may provide complementary information when interpreted alongside other clinical variables, potentially reflecting aspects of physiological disturbance during the early phase of critical illness that conventional indicators miss. However, the researchers explicitly caution that the present findings do not support interpreting magnesium as a therapeutic target. Serum magnesium levels may reflect underlying disease severity, renal handling, or treatment-related factors, and should therefore be considered primarily as a risk marker in this context. The observed non-linear relationship suggests that linear models may inadequately capture the association, and the authors conclude that further prospective studies are needed to clarify the role of serum magnesium in risk assessment and its potential clinical relevance for patients with sepsis.

Subject of Research: The association between early serum magnesium levels and the risk of sepsis-associated acute kidney injury in critically ill patients

Article Title: Serum magnesium levels and early risk of sepsis-associated acute kidney injury: an interpretable machine learning analysis based on the MIMIC-IV and eICU databases

Article References: Sun, Q., Sun, C., Sun, X., Xie, H., Ji, D., & Zheng, Y. (2026). Serum magnesium levels and early risk of sepsis-associated acute kidney injury: an interpretable machine learning analysis based on the MIMIC-IV and eICU databases. Critical Care: Sepsis and Severe Infection, 1(1), Article 4. https://doi.org/10.1186/s44541-026-00004-8

Image Credits: AI Generated

DOI: 10.1186/s44541-026-00004-8

Keywords: sepsis, acute kidney injury, serum magnesium, machine learning, MIMIC-IV, eICU, critical care, SHAP, CatBoost, KDIGO criteria, SOFA score, risk prediction