GNRI Falls Below 75 as 28-Day Mortality Risk Rises in Sepsis-Associated AKI
In critically ill patients with sepsis-associated acute kidney injury, nutritional status may be doing more than reflecting frailty at the bedside. A new explainable machine learning model built from 773 patients receiving continuous renal replacement therapy linked the Geriatric Nutritional Risk Index to short-term mortality, with risk rising sharply once GNRI dropped below 75. The analysis focused on 28-day all-cause mortality and used a feature-selection pipeline before comparing candidate algorithms, a design meant to balance prediction strength with interpretability.
The final model was not a deep neural network or a black-box ensemble. Logistic regression outperformed the other candidate approaches on the test set and achieved the highest AUC, 0.765, with a 95% confidence interval of 0.706 to 0.824. Nine predictors survived LASSO-based selection, and the model was then unpacked with SHAP values to show which variables were carrying the most weight. In an intensive care setting where rapid decisions are made under uncertainty, that combination of discrimination and interpretability matters because clinicians need to know not only who is high risk, but why the prediction is moving in that direction.
SOFA-CRRT and Creatinine Drove the Model, but Nutrition Stayed in the Signal
SHAP analysis placed SOFA_CRRT and serum creatinine at the top of the predictor hierarchy. GNRI ranked seventh, which may sound modest until its nonlinear behavior is examined more closely. Below 75, mortality risk climbed steeply; above 90, the association shifted toward protection. The curve suggests GNRI is not acting as a simple linear marker of “more malnutrition, more death.” Instead, it appears to behave like a threshold variable, where nutritional reserve may become biologically meaningful only after a critical depletion point is crossed.
That pattern fits the physiology of sepsis-associated kidney injury. Severe infection drives catabolism, capillary leak, and inflammatory stress, while acute kidney injury and CRRT add another layer of metabolic instability. Albumin, body weight, and overall nutritional reserve are not just background characteristics in that context; they may influence tolerance to vasopressors, tissue repair, immune competence, and the ability to recover from multi-organ failure. A GNRI threshold of 75, if replicated, could therefore help identify patients whose nutritional depletion is no longer a secondary issue but part of the acute risk state.
The cohort itself was clinically focused and relatively homogeneous: critically ill patients with S-AKI who were already receiving CRRT. That narrow framing is a strength for model development because it reduces noise from mixed populations. It also means the findings should not be casually exported to less sick ward patients, to non-CRRT populations, or to AKI without sepsis. Risk behavior in the ICU often changes once renal support begins, because the clinical trajectory is shaped by fluid balance, inflammatory burden, and the severity of organ failure rather than by a single laboratory abnormality.
Discrimination Was Acceptable, Calibration Was Not Clean
The model’s predictive performance was accompanied by a warning sign that cannot be ignored. Calibration was suboptimal, with a Hosmer–Lemeshow _p_ value of 0.004. In plain clinical terms, the model separated higher-risk from lower-risk patients reasonably well, but the predicted probabilities did not align as tightly as they should with observed outcomes. A tool can still be useful for stratification when discrimination is fair, yet miscalibration limits bedside confidence when the output is treated as an individual-level probability rather than a research-use signal.
Decision curve analysis suggested net clinical benefit across threshold probabilities from 0.1 to 0.6. This indicates the model may be helpful across a relatively broad range of decision thresholds, especially in settings where teams need a rapid way to identify patients at materially elevated risk. Still, decision curve benefit does not erase the calibration gap. A model that is easier to explain is not automatically ready for direct clinical deployment, particularly when the same dataset is doing the work of both development and testing.
The web-based prototype makes the project more accessible and more immediately usable for research exploration. It also creates a practical interface for comparing patient-level inputs against the model’s estimate. Even so, the authors were explicit that prospective external validation is required before any clinical application. That is the central boundary here: a research prototype can organize risk, but it cannot yet direct care.
Why an Explainable Model Matters in ICU Nutrition Risk
Explainability is not a cosmetic feature in sepsis care. ICU decisions are shaped by multiple simultaneous threats: shock, renal failure, inflammatory burden, and rapidly changing metabolic status. A model built around SHAP values gives clinicians a way to see whether creatinine, severity scores, or nutritional status are driving the output, which can improve trust and also expose when a prediction is being generated by an unexpected variable pattern. In a population as unstable as CRRT-treated S-AKI, the ability to inspect the model’s reasoning is clinically meaningful because the consequences of overcalling or undercalling risk can be immediate.
GNRI’s nonlinear behavior is the most clinically provocative part of the analysis. A steep risk inflection below 75 suggests a potential intervention window for nutritional assessment and support, although no causal claim can be made from this design. What the data do support is a strong association between marked nutritional compromise and death within 28 days in a cohort already facing high organ failure burden. That makes GNRI more than a descriptive score; in this setting, it may function as a marker of physiologic reserve that the ICU team can no longer ignore.
Research-Use Only Until External Validation Closes the Gap
The model was built from de-identified retrospective data drawn from a publicly accessible database, and the analysis was ethically approved as a secondary study. Funding came from regional research grants, and the authors reported no competing interests. Those details matter because they clarify the study’s scope: this is a retrospective, explainable prediction effort, not a prospective intervention trial.
The most actionable result is the threshold signal. GNRI below 75 tracked sharply higher mortality, and GNRI above 90 tracked a protective association. The most limiting result is the calibration failure. Between those two findings lies the real clinical message: the score may help sort risk in critically ill patients with sepsis-associated AKI, but it should stay in the research lane until it is tested prospectively in new cohorts and shown to predict reliably beyond the population in which it was built.