Background. Neonatal acute kidney injury (AKI) is associated with increased morbidity and mortality, yet its timely identification remains challenging. Diagnostic criteria rely on serum creatinine and urine output, whose interpretation is complicated by the physiological changes occurring during the first days of life. Urinary biomarkers may provide complementary information, although their concentrations can vary with neonatal characteristics independently of AKI. Integrating biomarker measurements with clinical information through machine learning may improve early risk assessment. Aim. Our study aimed to develop and evaluate interpretable machine learning models combining early clinical and laboratory variables with urinary biomarkers for neonatal AKI risk stratification. Secondary objectives were to investigate associations of clinical variables, outcomes and biomarkers with AKI, and to explore biomarker variability according to sex, gestational age and birth weight. Methods. We conducted a preliminary analysis within an ongoing single-center prospective observational cohort study involving neonates admitted to the NICU of the University Hospital of Padua between June 2024 and September 2025. Clinical and laboratory data and urinary NGAL, TIMP-2, IGFBP7 and CCL14 measurements were assessed within the first 48 hours of life. AKI occurrence during the first week of life was defined using composite neonatal-modified KDIGO and additional serum creatinine-based criteria. Biomarkers were compared by AKI status and neonatal subgroups; adjusted associations were assessed using multivariable logistic regression. Two additive Kolmogorov–Arnold Network (KAN) models were developed using either the first or last available measurements within 48 hours. Discrimination was evaluated through repeated nested cross-validation. Results. Among 239 neonates, 34 (14.2%) developed AKI. The First and Last KAN models achieved comparable discriminative performance, with area under the receiver operating characteristic curve (AUROC) values of 0.832 (95% CI 0.752–0.901) and 0.835 (95% CI 0.759–0.899), respectively. Neonates with AKI showed greater early clinical instability, including more frequent inotropic therapy and invasive respiratory support, and greater nephrotoxic drug exposure. Median postnatal weight change was +0.66% in the AKI group versus −1.66% in the non-AKI group (p-value < 0.001). AKI was associated with longer invasive ventilation and NICU stay. All four urinary biomarkers and the [TIMP-2] × [IGFBP7] index were significantly higher in the AKI group. Adjusted associations remained significant for NGAL, TIMP-2, IGFBP7 and the [TIMP-2] × [IGFBP7] index, but not for CCL14. Biomarker levels differed by sex, gestational age and birth weight even among neonates without AKI. Conclusions. The encouraging discriminative performance of KAN-based models suggests that integrating early clinical information and urinary biomarkers may be useful for neonatal AKI risk stratification, with relevant information potentially available from the earliest postnatal assessments. Our findings also highlight the importance of considering individual clinical and developmental characteristics when interpreting urinary biomarkers. Further evaluation in larger, independent cohorts and through prospective temporal validation could pave the way for the clinical application of these models, enabling earlier recognition of neonates at increased risk and supporting more personalized monitoring and preventive strategies during the first days of life.
Development of a Machine Learning–Derived Clinical Risk Stratification Tool to Guide Urinary Biomarker Assessment for Neonatal Acute Kidney Injury Prediction
FRISO, SOFIA
2024/2025
Abstract
Background. Neonatal acute kidney injury (AKI) is associated with increased morbidity and mortality, yet its timely identification remains challenging. Diagnostic criteria rely on serum creatinine and urine output, whose interpretation is complicated by the physiological changes occurring during the first days of life. Urinary biomarkers may provide complementary information, although their concentrations can vary with neonatal characteristics independently of AKI. Integrating biomarker measurements with clinical information through machine learning may improve early risk assessment. Aim. Our study aimed to develop and evaluate interpretable machine learning models combining early clinical and laboratory variables with urinary biomarkers for neonatal AKI risk stratification. Secondary objectives were to investigate associations of clinical variables, outcomes and biomarkers with AKI, and to explore biomarker variability according to sex, gestational age and birth weight. Methods. We conducted a preliminary analysis within an ongoing single-center prospective observational cohort study involving neonates admitted to the NICU of the University Hospital of Padua between June 2024 and September 2025. Clinical and laboratory data and urinary NGAL, TIMP-2, IGFBP7 and CCL14 measurements were assessed within the first 48 hours of life. AKI occurrence during the first week of life was defined using composite neonatal-modified KDIGO and additional serum creatinine-based criteria. Biomarkers were compared by AKI status and neonatal subgroups; adjusted associations were assessed using multivariable logistic regression. Two additive Kolmogorov–Arnold Network (KAN) models were developed using either the first or last available measurements within 48 hours. Discrimination was evaluated through repeated nested cross-validation. Results. Among 239 neonates, 34 (14.2%) developed AKI. The First and Last KAN models achieved comparable discriminative performance, with area under the receiver operating characteristic curve (AUROC) values of 0.832 (95% CI 0.752–0.901) and 0.835 (95% CI 0.759–0.899), respectively. Neonates with AKI showed greater early clinical instability, including more frequent inotropic therapy and invasive respiratory support, and greater nephrotoxic drug exposure. Median postnatal weight change was +0.66% in the AKI group versus −1.66% in the non-AKI group (p-value < 0.001). AKI was associated with longer invasive ventilation and NICU stay. All four urinary biomarkers and the [TIMP-2] × [IGFBP7] index were significantly higher in the AKI group. Adjusted associations remained significant for NGAL, TIMP-2, IGFBP7 and the [TIMP-2] × [IGFBP7] index, but not for CCL14. Biomarker levels differed by sex, gestational age and birth weight even among neonates without AKI. Conclusions. The encouraging discriminative performance of KAN-based models suggests that integrating early clinical information and urinary biomarkers may be useful for neonatal AKI risk stratification, with relevant information potentially available from the earliest postnatal assessments. Our findings also highlight the importance of considering individual clinical and developmental characteristics when interpreting urinary biomarkers. Further evaluation in larger, independent cohorts and through prospective temporal validation could pave the way for the clinical application of these models, enabling earlier recognition of neonates at increased risk and supporting more personalized monitoring and preventive strategies during the first days of life.| File | Dimensione | Formato | |
|---|---|---|---|
|
Friso_Sofia.pdf
Accesso riservato
Dimensione
2.15 MB
Formato
Adobe PDF
|
2.15 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/120640