Current descriptions of risk factors for acute kidney injury (AKI) are often qualitative and do not quantify how specific biomarker values or their interactions influence risk. The authors aimed to synthesize interpretable machine learning (ML) results developed independently across multiple health systems to identify generalizable, value-specific risk drivers and bivariate interactions that contribute to AKI susceptibility.
The analysis used de-identified electronic health records from nine U.S. academic medical centers participating in PCORnet networks. The cohort comprised 785,497 adult inpatients with encounters from 2010 through 2019. The underlying EHR data are governed by the Greater Plains Collaborative (GPC) and PaTH clinical research networks and are not publicly available; access requires institutional agreements and coordinating center approval.
Each participating health system independently trained interpretable gradient boosting machine models to quantify predictor–outcome associations for AKI. Site-level predictor–risk relationships were then integrated using meta-regression to (1) synthesize cross-system evidence, (2) characterize nonlinear value–risk curves for continuous biomarkers, and (3) identify bivariate interactions between predictors that jointly modulate AKI risk. This cross-system meta-analytic approach was intended to produce clinically interpretable, generalizable thresholds rather than site-specific black-box outputs.
The pooled results revealed several consistent, value-specific risk drivers across health systems. Notable quantitative associations reported include:
An increase in glucose from 100 mg/dL to 140 mg/dL was associated with a 1.46-fold higher risk of AKI.
Anion gap showed an elevated AKI risk across a portion of its range, with a 1.14-fold increase reported across 4–12 mmol/L.
Chloride demonstrated risk increases overlapping portions of the reference range, with a 1.28-fold AKI risk increase reported across 96–100 mEq/L.
These findings emphasize that meaningful AKI risk changes can occur within or near conventional laboratory reference intervals for some biomarkers.
Electrolytes including potassium, calcium, and sodium exhibited quadratic (U-shaped) associations with AKI risk in the meta-analysis. This pattern indicates higher AKI risk at both lower and higher values of these electrolytes rather than a strictly monotonic relationship.
Bivariate meta-regression further identified interactions between key predictors, highlighting physiologic pathways and combined effects that modulate AKI risk beyond single-variable associations. The source reports that such interactions clarify how pairs of biomarkers jointly influence susceptibility, though specific pairwise interaction effect sizes and all interacting variable combinations were described in the full study outputs rather than fully enumerated in the abstract.
By translating ML-derived associations into quantitative, value-specific thresholds and exposing predictor interactions, this cross-system meta-analysis provides actionable insights for clinical risk stratification. The approach moves beyond surface-level associations to highlight physiologic ranges where even modest shifts in biomarker values are associated with meaningful changes in AKI risk. These results could inform personalized prevention strategies in hospital care, electronic risk alerts, or incorporation into decision support where generalizable thresholds are preferred over single-site cutoff points.
The data used were de-identified EHRs from GPC and PaTH and are not publicly available due to institutional and regulatory restrictions. Researchers seeking access must contact the coordinating centers and obtain institutional data use agreements and approvals to ensure patient privacy and compliance with ethical requirements. Institutional Review Boards at the University of Florida, the University of Pittsburgh, and the University of Missouri waived ethical approval for this work as reported by the authors.
The abstract reports that the data are restricted and that certain author disclosures exist: one author reported paid consulting unrelated to the present work and full-time employment at a company, and other authors reported unrelated industry-supported research or advisory roles. The abstract does not provide a detailed list of all statistical parameters, full interaction effect estimates, or all modeling hyperparameters; those details are available in the full manuscript and supplementary materials. The study emphasizes synthesis of ML-derived associations but does not claim prospective validation in clinical trials within the abstract.
This cross-system meta-analysis integrates interpretable gradient boosting models across nine academic centers and 785,497 inpatient records to define value-specific risk drivers and interactions underlying AKI. Quantified relationships—such as the glucose, anion gap, and chloride associations and the quadratic patterns for key electrolytes—demonstrate that clinically relevant AKI risk changes can occur within conventional biomarker ranges. These generalized, quantitative thresholds and identified interactions offer a foundation for improved risk stratification and tailored preventive strategies in hospitalized patients.