A machine learning-based model for predicting the postoperative risk of acute kidney injury in neonates Authors: Liping He, Tianyin Gao, Yanli Tang, Saifen Jin, Manli Zhuang Published: 2025 ABSTRACT AbstractBackgroundAcute kidney injury (AKI) is a serious postoperative complication in hospitalized neonates. We aimed to develop and evaluate a machine learning (ML) model for predicting the risk of postoperative AKI in neonates.MethodsThe clinical records of 2,025 neonates were collected, and the patients were randomly divided into training and test sets. The outcome variable was the occurrence of postoperative AKI, and the models incorporated 25 predictive variables, including demographics, intraoperative infusions, and postoperative indicators. ML models were developed using six different algorithms on the training set, and their performance was assessed on the test set using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The model with the best AUC was selected for validation in the test set. The association between the risk factors and postoperative AKI was interpreted using the SHapley Additive exPlanations (SHAP) method.ResultsA total of 110 neonatal patients (5.43%) developed AKI following surgery. Patient age, operation duration, and urine output were the three most important predictors of AKI. Among the tested models, the logistic regression (LR) algorithm was the best predictor of postoperative AKI, achieving the highest AUC [median, 0.807; 95% confidence interval (CI): 0.701–0.897] and the highest sensitivity (median, 0.733; 95% CI: 0.5–0.938). The SHAP method was used to illustrate the prediction process of the LR model for neonatal postoperative AKI at the level of individual patients.ConclusionsThe ML model that uses the LR algorithm with eight commonly measured variables could serve as a tool to predict postoperative AKI in neonates.Keywords: Acute kidney injury (AKI), machine learning (ML), postoperative AKI complications, support vector machines (SVMs) Highlight box Key findings • This study developed and validated a logistic regression machine learning (ML) model using eight common clinical variables to predict postoperative acute kidney injury (AKI) in neonates. • The ML model had high accuracy (area under the curve: 0.807; sensitivity: 0.733) and outperformed five other ML models. What is known, and what is new? • AKI has a high global incidence among hospitalized children (26% in a meta-analysis and 20% in a Chinese population), and a mortality rate of up to 11%. Neonates are vulnerable to AKI after surgery because of their immature kidney function, hemodynamic instability, and reduced repair capacity. Current reliance on serum creatinine and urine output lacks sensitivity, and adult urinary biomarkers are unreliable for neonates. • This study was the first to use the SHapley Additive exPlanations (SHAP) method for the individualized and interpretable prediction of postoperative AKI in hospitalized neonates. What is the implication, and what should change now? • Our ML model could help clinicians to predict postoperative AKI in neonates; the most important predictors were patient age, operation duration, and urine output. • The SHAP method provides model transparency and enables the visualization and interpretation of the effects of different risk factors for individual patients. Introduction Acute kidney injury (AKI) is characterized by a rapid decline in renal filtration, accompanied by a sudden increase in serum creatinine or a decrease in urine output (1). AKI affects approximately 13 million hospitalized patients worldwide each year and contributes to nearly two million deaths annually (1). A meta-analysis of hospitalized children reported that the overall incidence of AKI was 26% and the incidence of AKI-related mortality was 11% (2). A recent multicenter study in China reported that the incidence of postoperative AKI was 28% in infants and 12% in adolescents (3), the prevalence of AKI among hospitalized children was 20%, and the mortality rate among children with AKI was 4% (3). Compared to adults, neonates are at greater risk of AKI because their low glomerular filtration rate and immature renal function make them more vulnerable to hemodynamic changes during nephrogenesis, and their capacity for self-repair after injury is limited (4). Neonates are also more vulnerable to the triggers of AKI, such as infection, surgery, and nephrotoxic drugs (5). Additionally, early fluid overload is closely correlated with the occurrence and prognosis of AKI in neonates (6,7). The current diagnostic standard, which is mainly based on serum creatinine level and urine output, lacks sensitivity and accuracy for neonates. Recent advances in the diagnosis of neonatal AKI have focused on identifying potential urinary biomarkers, such as neutrophil gelatinase-associated lipocalin, kidney injury molecule-1, liver fatty acid-binding protein, and interleukin-18, which are early predictors of AKI in adults (8). However, due to differences in growth, development, and other variables, these biomarkers cannot be directly applied to the prediction of neonatal AKI (9). Dong et al. established and validated a multi-approach machine learning (ML) model that successfully predicted stage 2/3 AKI up to 48 h before onset in pediatric patients (aged 1 month to 21 years) who were hospitalized in an intensive care unit (10). However, the applicability of this previous model to neonates remains unclear. Given the immature renal function, the physiological stress of surgery, and the risk of exposure to nephrotoxic agents in neonates (11), the early prediction of AKI is crucial to enable timely interventions that improve clinical outcomes and reduce morbidity and mortality (12). To identify potential risk factors before the onset of AKI, we developed a ML model by analyzing and evaluating a cohort of 2,533 neonates. We hypothesized that a ML model could accurately predict the occurrence of postoperative AKI in neonates, enabling clinicians to implement preventive strategies that improve patient outcomes. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/rc). Methods Study population and design A total of 2,533 neonates (aged up to 28 days) who underwent surgery at Guangzhou Women and Children’s Medical Center between January 2015 and October 2022 were selected for inclusion in this retrospective study. Data on demographics, renal function indicators, surgical information, intraoperative medications, and prognostic parameters were collected from the patients’ clinical records. Patients were excluded from the study if they had preoperative AKI, had incomplete information, had serum creatinine measured more than 48 h after surgery, died within 48 h after birth, had a maternal history of renal failure, and/or had undergone unrelated surgeries. Ultimately, 2,025 newborns were included in the cohort for model construction. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Medical Ethics Committee of Guangzhou Women and Children’s Medical Center (IRB No. 2024[064A01]). Written informed consent for participation and the use of clinical data was obtained from the parents or legal guardians of all the neonates included in this study. Definition of AKI AKI was defined according to the Kidney Disease: Improving Global Outcomes criteria (13) as an increase in serum creatinine of 0.3 mg/dL or more within 48 h, or an increase in serum creatinine of at least 1.5-fold from the baseline within the previous 7 days. Data selection Based on previous studies (14,15), data from six categories were retrieved from the hospital and laboratory records: (I) demographics: age (in days), sex, and body weight; (II) perioperative laboratory tests (predictors): all kidney function-related test results (e.g., baseline creatinine) collected within 48 h before surgery; (III) intraoperative medications: injections of dopamine, epinephrine, norepinephrine, deoxyadrenaline, nitroglycerin, sodium bicarbonate, furosemide, and albumin, as well as total colloid and crystalloid volumes (as indicators of hemodynamic instability and the risk of hemorrhage); (IV) intraoperative fluid management: intraoperative fluid administration, urine output, transfusions of all blood products [red blood cells (RBCs), plasma, and total blood products], and total fluid intake, which were classified according to specific criteria; (V) postoperative indicators: duration of surgical procedure and indicators of hemodynamic instability (e.g., hypernatremia, hypokalemia, and metabolic acidosis); and (VI) prognostic information: in-hospital mortality, length of hospitalization, surgical cost, and total hospital expenses. Definitions of outcomes The primary study outcome was defined as the incidence of postoperative AKI before hospital discharge. Postoperative AKI is a common complication of major surgical procedures, and is associated with significant short- and long-term morbidity. The long-term adverse outcomes of AKI include increased costs for surgery and overall care; prolonged hospital stays; higher rates of admission to an intensive care unit; increased rates of readmission; and progression to chronic kidney disease, end-stage renal disease, and even death. The associated complications include hyperkalemia, hyponatremia, and metabolic acidosis (16,17). Variable selection A univariate analysis was first conducted to eliminate irrelevant variables. Quantitative data were analyzed using the t-test, Mann-Whitney U test, Chi-squared test, or Fisher’s exact test as appropriate (16). Statistically significant variables (P<0.05) were selected for further analysis. The least absolute shrinkage and selection operator (LASSO) was also used for variable selection (17,18). Finally, the model was established using the non-zero coefficient features obtained from the LASSO regression. Model construction and interpretation Models were constructed using six different algorithms: logistic regression (LR), multi-layer perceptron (MLP) classifier, support vector machine (SVM), random forest (RF) classifier, eXtreme Gradient-Boosting (XGBoost) with classification trees, and light gradient-boosting machine (LightGBM). The main cohort was randomly divided into a training group (80%) and a validation group (20%), and the same proportions were used for model development and internal validation. The grid search method was used to optimize the parameters for each model (19). After model establishment, the bootstrap method was used to derive the 95% confidence intervals (CIs) for the evaluation metrics of the best-tuned models (20). The models were compared based on the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, accuracy, sensitivity, and F1 score. The SHapley Additive exPlanations (SHAP) method was used to interpret the effect of selected variables on model prediction (21). No variable had more than 1% missing values; the missing values were imputed using the mean of the associated variable. Before applying the ML models, continuous variables were normalized using the mean and standard deviation of the training set. Categorical variables were transformed into binary variables, with 1 indicating the presence of an incident and 0 indicating its absence. Statistical analysis The data processing and analysis were conducted using Python (3.9.13), Pandas (1.4.4), Bumpy (1.23.5), and other related packages. The basic models (LR, MLP, SVM, and RF) were built using scikit-learn (1.0.2), while the XGBoost and LightGBM models were established using the XGBoost and LightGBM packages, respectively. The SHAP package (0.41.0) was used to interpret the model results. The LR model calculates the probability of a neonate developing postoperative AKI (AKI =1) using a sigmoid function. This probability is derived from a linear predictor (Z) that integrates eight key clinical variables. Specifically, the probability of (AKI =1) was calculated as follows: P(AKI =1) = 1/(1 + exp (−Z)). The linear predictor Z itself was determined by the weighted combination of the selected features as follows: Z = −2.305 + 0.150 × (duration of operation) − 0.027 × (urine volume) − 0.120 × (total input) − 0.315 × (age in days) − 0.044 × (weight) + 0.006 × (use of RBC suspension) + 0.012 × (use of platelets) + 0.099 × (receipt of emergency surgery). SHAP values explain how different variables affect a model’s predictions by determining the influence of different characteristics of all samples. In the plot, the horizontal axis represents the SHAP value, and the vertical axis ranks variables by importance, from bottom to top. Each point in the figure represents a sample, and a greater dispersal of points indicates a more significant variable. If the points are concentrated on the middle line, then the variable has a lower effect on the outcome. A blue color indicates a lower value, and a red color indicates a higher value. A negative SHAP value (to the left of the median line) indicates that the variable biases the model toward negative outcomes; a positive SHAP value (to the right of the median line) indicates that the variable biases the model toward positive outcomes. Results Incidence of AKI We assessed the records of 2,533 neonates who underwent surgical procedures at our hospital between January 2015 and December 2022 to determine their eligibility for inclusion in the study. After excluding 397 neonates with preoperative AKI, 22 who underwent kidney, bladder, or ureter surgery, and 89 who lacked sufficient postoperative data, the data of 2,025 neonates were used to develop and evaluate the performance of the ML models that used six different algorithms to predict postoperative AKI. Of the 2,025 neonates, 110 (5.43%) were diagnosed with postoperative AKI, while the other 1,915 neonates (94.6%) did not develop AKI (Figure 1). Disposition and enrollment of neonatal surgery patients. AKI, acute kidney injury; LASSO, least absolute shrinkage and selection operator; LightGBM, light gradient-boosting machine; LR, logistic regression; MLP, multi-layer perceptron; RF, random forest; SVM, support vector machine; XGBoost, eXtreme Gradient-Boosting. Analysis of intraoperative and postoperative factors related to AKI To compare the two groups, we used the rank-sum test to analyze the continuous variables and the Chi-squared test to analyze the categorical variables (Table 1). The results revealed no significant differences (P>0.05) between the two groups in terms of the use of autologous blood, washed RBCs, albumin, epinephrine, norepinephrine, nitroglycerin, sodium bicarbonate, furosemide, blood loss, total colloid volume, and total crystalloid volume. However, operation duration, urine output, total fluid intake, patient age, bodyweight, the use of five different blood products (dopamine, RBC suspension, plasma, platelets, and cryoprecipitate), and receipt of emergency surgery were found to be associated with postoperative AKI (P<0.10). These 11 variables were thus included in the multivariate LR analysis, in which the variables were selected using the stepwise method. The optimal model had an Akaike Information Criterion value of 489.52, and included the following eight statistically significant variables: operation duration, urine volume, total fluid intake, age, bodyweight, use of RBC suspension, use of platelets, and receipt of emergency surgery. Table 1Intraoperative variables in neonatal patients with or without AKIVariablesPatients without AKI (n=1,915)Patients with AKI (n=110)P valueDuration of operation (min)150.00 (95.00, 198.50)192.00 (102.00, 255.00)0.002Urine volume (mL)20.00 (5.00, 50.00)15.00 (3.00, 40.00)0.02Total fluid intake (mL)50.00 (10.00, 100.00)40.00 (0.00, 80.00)0.050Blood loss volume (mL)10.00 (2.00, 30.00)20.00 (2.00, 30.00)0.83Total colloid volume (mL)0.00 (0.00, 0.00)0.00 (0.00, 0.00)0.72Total crystalloid volume (mL)0.00 (0.00, 60.00)0.00 (0.00, 30.00)0.34Age (days)11.00 (4.00, 20.00)2.00 (1.00, 7.00)<0.001Gender0.16 Male1,17569 Female74041Weight (kg)3.00 (2.70, 3.40)3.00 (2.50, 3.20)0.08Use of dopamine646 (63.6)40 (51.9)0.056Use of furosemide66 (6.5)3 (3.9)0.51Use of albumin29 (2.9)1 (1.3)0.66Use of epinephrine380 (37.4)30 (39.0)0.88Use of norepinephrine12 (1.2)1 (1.3)>0.99Use of deoxyadrenaline8 (0.8)0 (0.0)0.93Use of nitroglycerin127 (12.5)12 (15.6)0.55Use of sodium bicarbonate374 (36.8)26 (33.8)0.68Use of RBC suspension148 (14.6)19 (24.7)0.03Use of plasma111 (10.9)14 (18.2)0.08Use of platelets95 (9.4)5 (19.5)0.008Use of autologous blood402 (39.6)37 (48.1)0.18Use of washed RBCs25 (2.5)0 (0.0)0.32Use of cryoprecipitate73 (7.2)11 (14.3)0.042Receipt of emergency surgery364 (35.8)46 (59.7)<0.001Continuous variables are presented as median (interquartile range); categorical variables are presented as number or number (%). AKI, acute kidney injury; RBC, red blood cell. Intraoperative variables in neonatal patients with or without AKI Continuous variables are presented as median (interquartile range); categorical variables are presented as number or number (%). AKI, acute kidney injury; RBC, red blood cell. Feature selection using univariate and recursive feature elimination methods Because partially relevant or less important features may decrease the performance of a ML model, we also performed feature selection and rated feature importance. Feature selection by use of univariate and LASSO regression methods decreased the number of variables from 26 to 8. The selected features were: age, duration of operation, total fluid intake, receipt of emergency surgery, weight, urine volume, the use of RBC suspension, and the use of platelets. A feature importance plot was then generated using a fine-tuned LR model (22). Age (days), duration of operation (min), and urine output (mL) were the most important features. Performance assessment of the ML models for the prediction of postoperative AKI Six ML models were then constructed using the LR, SVM, RF, MLP, XGBoost, and LightGBM algorithms. Model performance in predicting postoperative AKI was assessed based on the AUC (Figure 2). The LR model had the highest AUC value (0.807). The AUC values of the models using the LR, SVM, and MLP algorithms were significantly higher than those using the XGBoost, RF, and LightGBM algorithms. Performance of external validation. (A) ROC curves of six models in the internal validation set. (B) ROC curves of the LR model in the internal validation set. (C) Calibration plot of the external validation set. LightGBM, light gradient-boosting machine; LR, logistic regression; MLP, multi-layer perceptron; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme Gradient-Boosting. In addition to the AUCs, we also calculated the accuracy, sensitivity, specificity, and F1 score of each model using 1,000 bootstrapped test datasets (Table 2). The LR model had the highest AUC (median, 0.807; 95% CI: 0.701–0.897) and sensitivity (median, 0.733; 95% CI: 0.5–0.938), the MLP model had the best accuracy (median, 0.755; 95% CI: 0.701–0.809) and the highest specificity (median, 0.791; 95% CI: 0.735–0.847), and the LightGBM model had the highest F1 score (median, 0.237; 95% CI: 0.113–0.356). Table 2Performance of the six ML models in the testing setML modelsAUCAccuracySensitivitySpecificityF1Internal validation set LR0.807 (0.701, 0897)0.685 (0.626, 0.744)0.733 (0.5, 0.938)0.68 (0.621, 0.739)0.235 (0.12, 0.362) MLP0.638 (0.561, 0.783)0.755 (0.701, 0.809)0.400 (0.215, 0.630)0.791 (0.735, 0.847)0.229 (0.133, 0.342) SVM0.747 (0.607, 0.867)0.658 (0.589, 0.717)0.667 (0.417, 0.917)0.655 (0.589, 0.718)0.211 (0.103, 0.319) RF0.705 (0.568, 0.821)0.658 (0.598, 0.717)0.533 (0.267, 0.778)0.668 (0.608, 0.729)0.174 (0.072, 0.288) LightGBM0.696 (0.543, 0.838)0.74 (0.685, 0.795)0.6 (0.333, 0.857)0.75 (0.689, 0.806)0.237 (0.113, 0.356) XGBoost0.735 (0.613, 0.851)0.667 (0.607, 0.726)0.611 (0.333, 0.846)0.672 (0.607, 0.736)0.2 (0.092, 0.306)External validation set LR0.779 (0.671, 0.842)0.738 (0.507, 0.912)0.576 (0.427, 0.693)0.901 (0.872, 0.943)0.349 (0.218, 0.435) MLP0.597 (0.465, 0.729)0.547 (0.421, 0.677)0.640 (0.530, 0.729)0.538 (0.411, 0.640)0.205 (0.118, 0.323) SVM0.713 (0.603, 0.823)0.737 (0.611, 0.852)0.600 (0.486, 0.784)0.751 (0.665, 0.819)0.294 (0.118, 0.362) RF0.638 (0.518, 0.759)0.755 (0.640, 0.826)0.400 (0.319, 0.507)0.791 (0.637, 0.887)0.230 (0.153, 0.374) LightGBM0.677 (0.557, 0.764)0.650 (0.543, 0.748)0.560 (0.473, 0.664)0.659 (0.547, 0.738)0.226 (0.117, 0.339) XGBoost0.646 (0.522, 0.770)0.667 (0.526, 0.769)0.520 (0.437, 0.631)0.683 (0.542, 0.788)0.222 (0.137, 0.309)AUC, area under the curve; LightGBM, light gradient-boosting machine; LR, logistic regression; ML, machine learning; MLP, multi-layer perceptron; RF, random forest; SVM, support vector machine; XGBoost, eXtreme Gradient-Boosting. Performance of the six ML models in the testing set AUC, area under the curve; LightGBM, light gradient-boosting machine; LR, logistic regression; ML, machine learning; MLP, multi-layer perceptron; RF, random forest; SVM, support vector machine; XGBoost, eXtreme Gradient-Boosting. To further verify the generalizability of the models, we used data from 448 newborns collected between November 2022 and December 2023 for external validation (Table 2). The results confirmed that the LR model had the highest AUC, specificity, and F1 score, consistent with the results from the test set. Interpretation of the model and analysis of influencing factors A SHAP analysis was then performed to determine which variables had the greatest influence on the model (Figure 3). The risk of postoperative AKI increased with younger age, longer operation time, lower total fluid intake, lower bodyweight, decreased urine volume, receipt of emergency surgery, and use of platelets. The use of RBCs had no significant effect. SHAP summary plot and dependence plot. (A) SHAP summary plot demonstrating the general importance of each feature in the LR model. The color bar on the right indicates the relative value of a variable in each patient, with red dots indicating high values, and blue dots indicating low values. The distance between the upper and lower margins of each violin plot represents the number of patients that had the same SHAP values described by this variable. (B) Feature importance ranking of the eight selected variables. (C) SHAP dependence plots demonstrating the distribution of the SHAP output values of individual variables. LR, logistic regression; SHAP, SHapley Additive exPlanations. Based on the distribution of the SHAP values among the continuous variables, an age of less than 10 days, operation duration exceeding 200 min, total fluid intake less than 100 mL, bodyweight below 3 kg, and urine volume less than 30 mL increased the risk of AKI. Among the categorical variables, receipt of emergency surgery and the use of platelets were associated with an increased risk of AKI. The use of RBCs did not significantly affect the risk of AKI. Discussion In this study, we developed and validated a ML-based model for predicting postoperative AKI in neonates. After thorough analysis, we selected eight key variables: age, duration of operation, total fluid intake, receipt of emergency surgery, bodyweight, urine volume, use of a RBC suspension, and use of platelets. We then evaluated the performance of multiple ML algorithms using these variables for predicting postoperative AKI in neonates. The LR algorithm demonstrated the best overall performance, achieving the highest AUC value and sensitivity. While the MLP model attained the best accuracy and the highest specificity. These results suggest that these models could be used to predict postoperative AKI in neonates. Among the eight selected variables, urine output (which is easily acquired) was of paramount importance. Urine output is also widely used as an indicator of renal function in contemporary AKI classification. Notably, urine output may be more useful in detecting shifts in renal hemodynamics than biochemical markers of solute clearance (23). The total water intake refers to the overall quantity of water consumed, typically measured over a specific period, and includes water from drinking, eating (food), and metabolic processes, such as cellular respiration. Crystalloid volume refers to the accumulation of crystals in bodily fluids or tissues due to various metabolic processes or medical conditions. These crystals can form due to the precipitation of substances such as uric acid, calcium oxalate, or calcium phosphate, and can potentially contribute to conditions such as kidney stones or gout. The duration of an operation is another variable that is crucial for surgical planning, resource allocation, and assessing procedural complexity. The transfusion of autologous blood and platelets involves the collection and storage of a patient’s own blood or platelets before a scheduled procedure for re-transfusion during or after the surgery. This approach reduces the need for donated blood, decreases the risk of transfusion reactions and the transmission of infectious diseases, and conserves donated blood for patients with urgent needs. After carefully analyzing our results, we selected the following key variables to compare neonates with and without postoperative AKI: demographic characteristics, surgical indicators, intraoperative medications, use of crystalloids and colloids, and postoperative indicators (Table 1). We found that AKI was more common in younger neonates and in those undergoing longer operations. The increased susceptibility of younger neonates is attributable to their incomplete nephron development and limited capacity for renal repair, which makes it difficult for them to recover from AKI. A prolonged surgical duration exacerbates the effects of reduced kidney blood supply, and can lead to an increased risk of kidney ischemia and subsequent ischemic/reperfusion injury. Statistically significant differences were observed between the groups with and without AKI in terms of urine output, total fluid intake, and autologous blood and platelet transfusion (all P<0.05). A SHAP model from another recent study suggested that off-pump coronary artery bypass grafting was associated with AKI, and identified intraoperative urine volume as a major contributor to this perioperative outcome (18). Another study of children under 12 years of age who were admitted to a pediatric intensive care unit between 2013 and 2014 found that reduced fluid intake was a significant risk factor for AKI (P=0.005) (19). A study performed between 2015 and 2017 evaluated the effect of platelet transfusion in patients with acute Stanford type A aortic dissection, and reported that platelet this transfusion was an independent risk indicator for AKI (20). Our groups with and without AKI also differed significantly in terms of the use of norepinephrine during surgery (P<0.05). Norepinephrine is one of the most commonly used medications to maintain organ perfusion pressure in patients with sepsis-induced AKI (21). The combined use of cryoprecipitate was also significantly greater in the with AKI group than the without AKI group (P<0.05). Finally, our analysis of postoperative indicators indicated that the neonates with AKI had significantly higher total hospitalization and surgical costs than those without AKI (both P<0.001). The incidence of hyponatremia, metabolic acidosis, and mortality was also greater in the patients with AKI than those without AKI (all P<0.001). Multiple studies have found that ML models have significant potential in predicting postoperative AKI (22,24-26). Previous studies have applied these models to predict postoperative AKI in adults, among which, the XGBoost algorithm has shown superior performance (higher AUC values and lower error rates) than traditional LR (22). Tseng et al. developed a ML model that incorporated preoperative and intraoperative variables to predict AKI in adults after cardiac surgery, and found that the ensemble RF and XGBoost algorithms performed best (27). However, models for the prediction of AKI in adults cannot be directly applied to neonates due to their immature development of nephrons and unique kidney physiology (28). Thus, predicting postoperative AKI in neonates remains challenging, highlighting the urgent need for a reliable and accurate predictive model. The early detection of postoperative AKI is critical so that timely interventions can be administered to prevent complications and death. Similar to our model, several previous studies have found that age was a key independent factor for the diagnosis of postoperative AKI in children (5,10,29). Compared to healthy individuals, those with AKI typically have greater fluid intake and reduced urine output. Hypovolemia due to reduced total fluid intake is a recognized risk factor for the development of AKI (30). Renal perfusion and cardiac output are lower in patients with hypovolemia. For example, a study of pediatric patients (aged 1 month to 12 years) suggested that hypovolemia was an independent risk factor for AKI (P=0.005) (19). Urine output is an easily determined measure of kidney function, and decreased urine volume is a common and simple indicator of AKI. A definition of AKI suggested by the Acute Dialysis Quality Initiative is urine output below 0.5 mL/kg/h for more than 6 h (31). The morphological and functional development of the kidneys in neonates continues after birth until the age of 22 years (32). Younger children are more sensitive to AKI following major surgery due to the incomplete repair capacity of their kidneys. Consistent with previous reports, our ML model demonstrated that several intraoperative variables, including longer operation time, excessive transfusion of blood products, and total fluid intake, were also significantly correlated with AKI in neonates following surgery. Neonates are particularly vulnerable to prolonged surgical procedures, which can increase the risk of renal ischemia. Following reperfusion, this is one of the most common causes of AKI. Notably, from 1999 to 2001, ischemia was the leading cause of AKI (21%) among inpatients admitted to tertiary hospitals in the United States (33). Another study between 2015 and 2017 evaluated the effect of platelet transfusion in patients with acute Stanford type A aortic dissection, and showed that transfusion of a platelet concentrate was an independent risk factor for AKI (20). Conversely, we found an association between the use of cryoprecipitate and postoperative AKI in neonates. Crystal deposition can cause AKI due to mechanical obstruction, cytotoxicity, and nephro-inflammation (34). Mechanical obstruction significantly reduces blood supply to the kidneys and can lead to local renal ischemia. Renal tubular cells absorb and digest crystal microparticles through lysosomes. This releases free calcium into the cytosol, which is essential in the calpain-mediated necrosis pathway (34,35). Crystal-induced necrosis generates renal danger-associated molecular patterns (DAMPs), and the recognition of these DAMPs by innate immune cells promotes inflammation (36,37). This study had several limitations. First, the ML models were developed using data from a single center. Thus, a multicenter study is needed for future clinical validation. Second, the retrospective design of this study might have led to data collection and entry bias. Finally, there might have been residual confounding by unidentified variables. Conclusions This study established a ML-based model for the prediction of postoperative AKI in a population of neonates. The LR algorithm, with eight variables, achieved the best performance among all the tested ML models. Our ML model could serve as a powerful clinical tool for predicting postoperative AKI in neonatal patients. Supplementary The article’s supplementary files as Acknowledgments We would like to thank Medjaden Inc. for professional assistance in providing English language editing and proofreading services for this manuscript. Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Medical Ethics Committee of Guangzhou Women and Children’s Medical Center (IRB No. 2024[064A01]). Written informed consent for participation and use of clinical data was obtained from the parents or legal guardians of all neonates included in this study. Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/rc Funding: This work was supported by the Guangzhou Science and Technology Project (No. 2025A03J3709). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/coif). The authors have no conflicts of interest to declare. (English Language Editor: L. Huleatt) Data Sharing Statement Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/dss References Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/prf Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-428/dss