Machine learning model predicts acute kidney injury in pediatric patients after cardiac surgery: a systematic review and meta-analysis

Abstract

Background

Acute kidney injury (AKI) is a common complication following pediatric cardiac surgery, frequently leading to poor outcomes and even death in severe cases. Early prevention remains the primary intervention strategy. Studies have developed prediction models to identify at-risk children at an early stage. This study systematically evaluate existing AKI prediction models to support their clinical utility and future refinement.

Methods

PubMed, Embase, Web of Science, Cochrane Library, China National Knowledge Infrastructure, Wanfang and SinoMed were searched from inception to 31 December, 2024. The search of references from included studies, as well as the manual search, extended until November 30, 2025. Literature searching, screening, and data extraction were done by two authors. Quality evaluation according to prediction model risk of bias assessment tool (PROBAST). Area under the receiver operating characteristic curve (AUROC) was pooled using a random-effects model to summarize the overall performance of existing models, exploring sources of heterogeneity of performance through subgroup analysis and meta-regression. Sensitivity analysis and Egger’s method were used to analyze the stability of the included studies and to identify publication bias. This study was registered with PROSPERO (CRD42024593112) and reported following the Transparent Reporting of Multivariable Prediction Models for Individual Prognosis or Diagnosis: Checklist for Systematic Reviews and Meta-Analysis (TRIPOD-SRMA).

Results

A total of 2189 studies were screened which represented the total number of studies retrieved from the database search, the search of references from included studies, and the manual search. Nineteen studies were included in this review. Included studies differed in study design, AKI definition, predictor screening, model development and validation and model performance. The overall pooled AUROC was 0.850 (95% CI, 0.810–0.890), but all studies were evaluated as high risk of bias using the PROBAST. Heterogeneity in model performance was high, and study design and development methods were identified as possible sources of heterogeneity in pooled AUROC. Included studies were stable and free of publication bias.

Conclusions

This systematic review suggested that machine learning models for predicting postoperative AKI in pediatric cardiac surgery indicated good discriminative ability. However, the high risk of bias across all included studies and the significant heterogeneity in model performance indicated that the reported performance may be overestimated. The high heterogeneity observed highlights the substantial variability in model performance, which is likely driven by differences in study design and development methods. The clinical utility of these models was currently limited due to the lack of external validation in most studies and the methodological limitations identified. Future research must incorporate rigorous study design, transparent reporting based on the TRIPOD guidelines, and external validation to develop prediction models with clinical utility.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12882-026-05037-2.

Background

Acute kidney injury (AKI) is a sudden loss of kidney function that frequently occurs in children. It is associated with increased morbidity and mortality [1, 2]. AKI is a common postoperative complication after pediatric cardiac surgery, and the prevalence of AKI ranges from 15% to 64% [3], prolonging the postoperative hospital stay and seriously affecting the patient’s postoperative recovery [4, 5].

Kidney function in children is usually less affected by chronic disease than in adults, but at least until two years of age, children’s kidneys are not fully developed, and cardiac surgery during this time may leave patients vulnerable to functional kidney injury [6]. Most children require cardiopulmonary bypass (CPB) during cardiac surgery. During CPB, factors such as the contact between blood and CPB equipment, endotoxemia, ischemia-reperfusion injury, and surgical trauma act as stimuli [7]. These stimuli trigger the activation of multiple pro-inflammatory and anti-inflammatory mediators, including leukocytes, interleukins, and cytokines [8]. This process induces fibrin deposition in the glomeruli, activates coagulation pathways, and is accompanied by cellular infiltration and vasoconstriction [8, 9]. Consequently, kidney blood flow and glomerular filtration rate decrease, ultimately leading to the development of AKI. Although this pathophysiological mechanism based on the decreased kidney perfusion pressure leading to a transitory ischemic event is not immediately fatal in pediatric patients undergoing cardiac surgery [10], there is no effective treatment for AKI, and early identification and prevention can reduce the likelihood of transition to a severe stage of AKI and improve the patient’s prognosis.

Currently, serum creatinine level and urine output are used clinically to determine whether a pediatric patient has AKI [2]. The changes in serum creatinine and urine output are used to assess changes in glomerular filtration rate (GFR) that evaluate kidney function [11, 12]. The three AKI definitions of Kidney Diseases Improving Global Outcomes (KDIGO), Acute Kidney Injury Network (AKIN), and pediatric RIFLE (pRIFLE) are based on changes in the serum creatinine level and estimated glomerular filtration rate (eGFR) [13]. However, they have limited sensitivity and specificity, especially serum creatinine, with significant changes occurring only when the kidneys experience AKI, limitations that restrict the usefulness of early detection of AKI [14, 15]. Thus, new tools are needed for early identification of AKI and determination of preventive interventions [16].

Renal angina index (RAI) is superior in accuracy to serum creatinine alone for prediction as an early identifier of developing AKI in critical pediatric patients [17]. However, this tool still had limitations for predicting AKI because it requires the use of serum creatinine to calculate the score [18]. Multiple biomarkers for predicting AKI after pediatric cardiac surgery have been studied, such as Cystatin C (Cys-C), Neutrophil gelatinase-associated lipocalin (NGAL), Interleukin-18, Kidney injury molecule-1 (KIM-1), and fatty acid–binding protein (FABP) [19]. Prediction models combining multiple biomarkers also have good performance for predicting AKI and AKI progression after pediatric cardiac surgery [20–22]. However, prediction models incorporating biomarkers still lack external validation, and their clinical utility requires further verification. With the development of algorithms, more and more studies are using machine learning to develop models of predicting AKI in pediatric patients after cardiac surgery [23]. Utilizing machine learning’s ability to process complex data, internally validated prediction models were developed using a variety of indicators such as demographic characteristics, laboratory tests, surgery-related indicators, and biomarkers as predictors [24, 25]. However, the complexity of the pathophysiological mechanisms of postoperative cardiac AKI and the need for external validation of these models limits their clinical application.

This study systematically reviewed and critically evaluated the prediction models of postoperative cardiac complication AKI in children based on machine learning. This study aims to evaluate the predictive performance of these models and identify the relevant factors influencing it, thereby establishing a valuable reference for future model development and validation. It helps healthcare staff identify the strengths and limitations of current AKI prediction models for pediatric cardiac surgery patients, refer to existing models, and optimize them to develop models with clinical utility.

Materials and methods

Search strategy

We performed a systematic review and meta-analysis of the literature according to the Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analysis (TRIPOD-SRMA) [26] (The checklist details are shown in Appendix C). The protocol has been registered on PROSPERO (CRD42024593112). The PubMed, Embase, Web of Science, Cochrane Library, SinoMed, Wanfang and China National Knowledge Infrastructure (CNKI) databases were searched (The specific search strategy shown in Appendix A). The search was limited to December 31, 2024 and the reference lists of included studies were also searched (The studies included in this section were limited to those published before December 31, 2024). We simultaneously continued manual searches using the search strategy outlined in Appendix A to ensure the inclusion of the most recently published studies, with the search date extended until November 30, 2025.

Selection criteria

The inclusion criteria were as follows: (1) The population must consist of pediatric patients after cardiac surgery; (2) Models that were developed in the study must have been validated either internally or externally. Internal validation refers to the process in which a study used a dataset collected within a single time period to develop a model, and then used methods such as random splitting, cross-validation, or bootstrap to validate model performance. External validation refers to the process in which a study used a previously developed model and performed validation through temporal validation (using data from different time periods within the same geographic region), geographic validation (using data from different geographic regions or different healthcare institutions), or domain validation (using data from a domain different from that of the original model, such as validating a hospital-derived model with community data). All of these external validation approaches require the use of independent datasets that were not used in the original model development [27]; (3) Study designs included cohort studies, case-control studies, and cross-sectional studies; (4) The primary outcome of the studies was AKI.

The exclusion criteria were: (1) Studies that did not develop a prediction model; (2) Studies in which the predictive model contained < 2 predictors; (3) Study types were systematic review/meta-analysis (4) Full text not available.

Study screening and data extraction

The entire screening process was independently conducted by two authors (XF and JZ). The screening of references from included studies, as well as the manual search, were also conducted by XF and JZ. When two authors had disagreements, consulting and discussing with the third author (TL) to reach a consensus.

Data were extracted by two reviewers (ZX and ZC). Any disagreements in extracted data were resolved by consensus. We utilized the PICOTS system recommended by the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist to extract the specific item of included studies [28]. If data were missing, we attempted to obtain them by contacting the authors or through calculation.

The PICOTS are described below:

P (Population): The population of pediatric patients (age < 18 years) who received cardiac surgery.

I (Intervention model): Studies focused on prediction models (predictors > 1) that were internally or externally validated.

C (Comparator): None.

O (Outcome): The outcome was AKI (including severe AKI or ARF, the specific AKI definition shown in Appendix B Table S1).

T (Timing): Demographic characteristics, clinical variables, and biomarkers in the preoperative, intraoperative, and postoperative time periods were used as the main predictive information. The follow up duration was seven days after cardiac surgery generally.

S (Setting): Risk of AKI in pediatric patients undergoing cardiac surgery, in the intensive care unit or after transfer to the general ward. Heart transplantation and interventional procedures were considered as other clinical settings.

In this review, area under the receiver operating characteristic curve (AUROC), accuracy and area under the precision recall curve (AUPRC) were used as evaluation indicators for model performance. The data of the best model recommended in the included studies were extracted, and for studies that developed multiple models using the same predictors with different modeling methods and did not propose an optimal model, the optimal model was selected based on the AUROC value (The closer the AUROC value is to 1, the better the model’s classification performance. If models have the same AUROC, further screening was conducted based on accuracy, AUPRC, calibration, and other metrics).

Quality assessment

The risk of bias of each study was assessed using prediction model risk of bias assessment tool (PROBAST), which involved four domains: participants, predictors, outcomes, and analysis, with a total of 20 signaling questions. The risk of bias (ROB) in the research was evaluated through the four domains, while the applicability was evaluated in the first three domains [29]. Within each domain, two or more signaling questions required evaluation. Each signaling question could be rated as low risk, high risk, or unclear. A domain was considered to have low risk only when all its signaling questions were rated as low risk. If any signaling question was rated as high risk, the domain was classified as high risk. If any signaling question was rated as unclear while all others were rated as low risk, the domain was classified as unclear. This rule for assessing overall ROB and applicability was applied consistently across all domains. The quality assessment process was similarly carried out by two authors (XF and JZ). Any disagreements were resolved by team discussion.

Statistical analysis

In this study, we conducted the meta-analysis of AUROC using values and 95% confidence intervals (CIs). Because the AUROC could be chosen to represent the discrimination result. Some studies reported only the AUROC without 95% CIs; we derived the CIs using the inverse-variance method [30]. Detailed instructions for completing the calculations using code are provided in Table S2 of Appendix B. Although calibration data was also a measure of model performance, analysis of calibration data was not performed due to the paucity of calibration data in the included studies. We also conducted the accuracy, which was calculated using the Wilson interval method for models that did not report a 95% CI of accuracy [31], but this method required the study to report the sensitivity and specificity of the model. For studies that did not report these two indicators and only reported the AUROC as a graph, the corresponding sensitivity and specificity were extracted using GetData Graph Digitizer (version 2.26). We also conducted the meta-analysis of AUPRC using values and 95%CIs.

Heterogeneity between studies was assessed using I², I² ≤ 25%, 25% < I² < 75%, and I² ≥ 75% were considered as having low, moderate, and high heterogeneity, respectively [32]. If the Q test of the included studies with P > 0.05 and I² ≤ 50%, we used a fixed effect model to calculate effect values. Otherwise, a random effect model was used. Subgroup analysis of AUROC was conducted by development methods, study design, region, predictor numbers, AKI definition, AKI stage, age, outcome, clinical settings and model category that in order to explore the source of heterogeneity. In this study, we defined age groups based on the time of maturation of renal function in children using the mean and median as follows: age < 2 and ≥ 2 years (Within the pediatric population of the included studies, the maximum age range was < 21 years) [33, 34]. Meta-regression analysis was used for subgroups where the source of heterogeneity could not be determined. We conducted the sensitivity analysis which was leave-one-out meta-analysis [35] to evaluate stability of included studies and used the Egger’s method [36] to assess publication bias, simultaneously plotting the funnel that adjusted by trim and fill method [37].

A summary receiver operating characteristic (sROC) curve with a 95% CI was also generated using the hierarchical summary receiver operating characteristics (HSROC) model to evaluate the pooled discrimination ability from the studies where the requisite data were available [38].

The Microsoft Excel was used to input data, Statistics and Data Analysis (Stata) 17 and R studio version 4.5.0 were used to analyze data (Specific code shown in Appendix B Table S2). Statistical significance was set at P < 0.05, and all tests were 2-tailed.

Results

Studies included

A total of 2189 studies were screened. A total of 1579 records were retrieved in the database, with 1127 remaining after removing duplicate records. Ninety-three studies met the eligibility criteria that screened by titles and abstracts. These studies were further evaluated by reading the full text. Nineteen studies were excluded due to absence of prediction model. While these studies conducted logistic regression analyses, their primary focus was on identifying influencing factors, and they not calculated model performance metrics [39, 40]. Thirty-four studies developed prediction models, but these models contained < two predictors [41, 42]. Twenty-two studies utilized > two predictors to develop models, but these models not underwent validation [43]. Reasons for exclusion at this stage are shown in Fig. 1. Sixteen studies were included in this review preliminarily. A total of 610 records were retrieved from references list and manual searched. The reasons for study exclusion were consistent with those mentioned above. Three studies met the inclusion criteria. Ultimately, nineteen studies were included in this review.

Fig. 1.

PRISMA2020 flowchart of studies search and selection

Study characteristics

These models were developed to predict AKI in pediatric patients undergoing cardiac surgery. All studies included pediatric patients as participants, 73.7% of the studies had an age range of < 18 years. Some studies did not describe the age range. Among the studies included in this study, the maximum age range for pediatric patients was < 21 years. Details are shown in Table 1.

Table 1.

Basic characteristic of included studies

Study	Country	Study design	Participants	Prediction outcome*	Outcome definition
Bucholz 2015	US	Prospective study	Pediatric patients (age between 1 month and 18 years) who received cardiac surgery	AKI	AKIN
Bianchi 2013	Italy	Retrospective study	Pediatric patients (age ≤ 12 years) who received cardiac surgery and an angiographic study up to 14 days before surgery	ARF	pRIFLE
Basu 2014	US	Retrospective study	Pediatric patients (age < 18 years) who underwent CPB	Persistent AKI	KDIGO
Luo 2023	China	Retrospective study	Pediatric patients (age between 1 month and 18 years) who underwent cardiac surgery with CPB	AKI	KDIGO
Flechet 2019	Belgium	Prospective blinded observational study	Pediatric patients (age < 12 years) underwent cardiac surgery who were expected to stay in PICU for at least 24 h	Severe AKI	KDIGO
Parikh 2011	US	Prospective study	Pediatric patients (age between 1 month and 18 years) who underwent surgery for congenital heart disease	Severe AKI	pRIFLE, AKIN
Parikh 2013	US	Prospective study	Pediatric patients (age between 1 month and 18 years) who underwent surgery for congenital heart disease	AKI	pRIFLE, AKIN
Jia 2021	China	Retrospective study	Pediatric patients (age < 18 years) who underwent TCPC surgery	Persistent AKI	pRIFLE
Dasgupta 2021	US	Retrospective study	Pediatric patients (aged < 21 years) who underwent orthotopic heart transplantation	AKI	KDIGO
Zeng 2022	China	Retrospective cohort study	Pediatric patients who underwent congenital heart surgery	AKI	KDIGO
Cardoso 2016	Portugal	Retrospective study	Pediatric patients (aged < 18 years) who underwent cardiac surgery	AKI	pRIFLE
Kong 2023	China	Retrospective study	Pediatric patients who underwent CPB surgery	AKI	KDIGO, pRIFLE
Shi 2024	China	Retrospective cohort study	Pediatric patients (age 1 month to 14 years) who underwent cardiac surgery with CPB	AKI	KDIGO
Tong 2024	China	Retrospective cohort study	pediatric patients who received congenital heart surgery	ARF	KDIGO
Fragasso 2023	Italy	Retrospective study	Patients from birth to 17 years old who required elective and urgent admission to the PCICU for surgical	Severe AKI	KDIGO
Nagy 2024	US	Retrospective cohort study	Pediatric patients (age < 21 years) who underwent cardiac surgery	Severe AKI	KDIGO
Krawczeski 2010	US	Prospective study	Pediatric patients (age < 18 years) who underwent CPB	AKI	AKIN
Baloglu 2025	US	Retrospective cohort study	Pediatric patients (7 day ≤ age < 18 years) who underwent cardiac surgery	AKI	KDIGO
Deal 2025	UK	Retrospective study	Pediatric patients (age < 4 years) who underwent CPB	AKI	KDIGO

Note: *: In pRIFLE, AKIN and KDIGO, severe AKI is defined as stage 2–3 AKI, ARF is defined as stage 3 AKI, persistent AKI is defined as the duration of severe AKI > 2 days; AKI: acute kidney injury; ARF: acute renal failure; AKIN: Acute Kidney Injury Network; CPB: cardiopulmonary bypass; KDIGO: Kidney Disease Improving Global Outcomes; pRIFLE: pediatric Risk, Injury, Failure and Loss, and End-Stage; PCICU: pediatric cardiac intensive care unit; PICU: pediatric intensive care unit; TCPC: total cavopulmonary connection

The 19 included studies were mainly from the United States and China. Five studies applied prospective study design [20–22, 44, 45], 14 studies were retrospective study design [3, 23–25, 46–55], and three studies used registered database, eleven studies used patients’ medical records from a single center. Fourteen studies reported follow-up (Time range for predicting AKI) within seven days of cardiac surgery. The sample sizes ranged from 76 to 23,000 participants across the studies and the prevalence of AKI ranged from 2.0% to 74.9%. Specific information shown in Table 2. We summarized the clinical settings, model categories, timeframe for determining predictors, and outcome follow-up duration of all models. The most common clinical setting was routine postoperation which patients who have undergone cardiac surgery and have been transferred to a general ward. We categorized the models into three types: preoperative, intraoperative, and early postoperative, with the early postoperative model accounting for 57.9% of the total. The time periods used to determine the predictors in the two models overlap with the follow-up periods for the outcomes [22, 48]. This means that the predictors and the outcomes may be measured within the same time frame (Details are shown in Appendix B Table S3).

Table 2.

Methods and performance of the included models

Study	Cases/sample size (%)		Missing data handling	Predictors screening
Bucholz 2015	Derivation cohort: 55/106 (51.9%)		Only samples with complete data were included	Univariate analysis, multivariable analysis
Bianchi 2013	Derivation cohort: 55/277 (19.9%)		Deleted missing data	Univariate analysis, stepwise forward multivariable analysis
Basu 2014	Derivation cohort: 29/345 (8.4%) (Persistent AKI)		-	CART
Luo 2023	Derivation cohort: 564/3278 (17.2%) Validation cohort: 51/585 (8.7%)		Imputed by the random forest method	LASSO, Boruta algorithm, random forest-recursive feature elimination, random forest-filtering
Flechet 2019	Derivation cohort: 55/156 (35.3%)		-	Univariate analysis, stepwise backward multivariate analysis
Parikh 2011	Derivation cohort: 53/311 (17.0%)		-	-
Parikh 2013	Derivation cohort: 53/311 (17.0%)		-	-
Jia 2021	Derivation cohort: 72/465 (15.5%)		Imputed by the random forest algorithm	Univariate analysis, stepwise backward multivariate analysis
Dasgupta 2021	Derivation cohort: 131/175 (74.9%)		Only samples with complete data were included	Univariate analysis, multivariate analysis
Zeng 2022	Derivation cohort: 331/3386 (9.8%)		Sample-and-hold imputation	Incorporation of all variables collected in the database
Cardoso 2016	Derivation cohort: 40/325 (12.3%)		Deleted missing data	Univariate analysis, multivariate analysis
Kong 2023	Derivation cohort: 67/134 (50.0%)		Only samples with complete data were included	LASSO regression, univariate analysis
Shi 2024	Derivation cohort: 203/1137 (17.9%)		Only samples with complete data were included	LASSO regression
Tong 2024	Derivation cohort: 458/23,000 (2.0%)		-	Incorporation of all variables collected in the dataset
Fragasso 2023	Derivation cohort: 223/419 (53.2%) (Severe AKI)		A nonparametric missing value imputation algorithm	RF
Nagy 2024	Derivation cohort: 55/402 (13.7%)		The Light GBM algorithm considers missing data as an observation	Incorporation of all variables collected in the dataset
Krawczeski 2010	Derivation cohort: 119/374 (31.8%)		Deleted missing data	Univariate analysis, multivariate analysis
Baloglu 2025	Derivation cohort: 402/841 (47.8%) (AKI) Validation cohort: 138/259 (53.3%) (AKI)		The selected machine learning method for handling missing data was reliable.	Selection based on the availability of the PC4 database and its established correlation with AKI (as determined by prior research)
Deal 2025	Derivation cohort: 43/76 (56.6%) (AKI)		Multiple imputation	Univariate analysis, LASSO regression
Study		Development methods	Validation methods	Calibration methods
Bucholz 2015		LR	Internal validation (bootstrap)	-
Bianchi 2013		LR	Internal validation (A maximum number of 1 independent risk factor per each 10 events was admitted to the model)	The Hosmer-Lemeshow test (p = 0.705)
Basu 2014		CART	Internal validation	-
Luo 2023		K-nearest neighbor, NB, SVM, RF, XGBoost, NN	Internal validation (5-fold cross-validation) External validation (geographical validation)	Calibration curve, Brier score = 0.085 (internal validation) and 0.060 (external validation)
Flechet 2019		LR	Internal validation (1000 bootstrap)	Calibration curve
Parikh 2011		LR	Internal validation (3-fold cross validation)	-
Parikh 2013		LR	Internal validation (3-fold cross validation)	-
Jia 2021		LR	Internal validation (random split)	The Hosmer-Lemeshow test (p = 0.620), calibration curve
Dasgupta 2021		LR	Internal validation (bootstrap)	Calibration slope = 0.930
Zeng 2022		SVM, LR, LSTM, GRU, Dipole, RETAIN, a time-aware attention RNN	Internal validation (random split training set: testing set = 6:4, 5-fold cross validation)	Brier score = 0.061 or 0.062
Cardoso 2016		LR	Internal validation	The Hosmer-Lemeshow test (p = 0.834)
Kong 2023		XGBoost, LR, LGBM GNB, MLP, SVM	Internal validation (random split training set: testing set = 4:1, 5-fold cross validation)	The Hosmer–Lemeshow test, Brier score = 0.129, calibration curve
Shi 2024		LR	Internal validation (random split training set: testing set = 7:3)	Calibration curve
Tong 2024		LR, SVM, RF, LGBM, CatBoost	Internal validation (random split training set: testing set = 6:4, 10-flod cross validation)	-
Fragasso 2023		RF	Internal validation (random split training set: testing set = 7:3, 10-fold cross validation)	-
Nagy 2024		LGBM	Internal validation (random split training set: testing set = 8:2, 10-fold cross validation)	Brier score = 0.090
Krawczeski 2010		LR	Internal validation (random split training set: testing set = 1:1)	-
Baloglu 2025		LGBM, XGBoost, CatBoost, histogram gradient boosting	Internal validation (random split training set: testing set = 8:2, 10-fold cross validation)	Brier score = 0.200 (internal validation) Brier score = 0.200 (external validation)
Deal 2025		LR	Internal validation (1000 bootstrap)	Brier score = 0.130
Study		Model predictors	Model performance	Model presentation
Bucholz 2015		Preoperative h-FABP, age, site, preoperative eGFR, RACHS-1 ≥ 3	LR: AUROC = 0.800 (95%CI 0.71–0.89), Se = 0.680, Sp = 0.688, Accuracy = 0.684 (95%CI 0.590–0.765)	-
Bianchi 2013		Age, iodine dose on angiography, low cardiac output state	LR: AUROC = 0.848 (95%CI 0.778–0.918)	Formula
Basu 2014		Urine NGAL, plasma CysC	CART: AUROC = 0.950 (95%CI 0.895-1), Se = 0.930, Sp = 0.920, Accuracy = 0.921 (95%CI 0.887–0.945)	-
Luo 2023		Age, body length, weight, cyanotic heart disease, previous cardiac surgery, preoperative LOS, ASA physical status, baseline creatinine, baseline eGFR, hemoglobin, platelets, total bilirubin, ALT, AST, albumin, potassium, sodium, chloride, calcium, diuretics, operation time, perfusion time, cross clamp time, lowest MAP, lowest core temperature, intraoperative blood loss, RACHS-1 score	XGBoost internal validation: AUROC = 0.912 (95%CI 0.899–0.924), Se = 0.603, Sp = 0.950, Accuracy = 0.890 (95%CI 0.879–0.901), AUPRC = 0.747 XGBoost external validation: AUROC = 0.899 (95%CI 0.844–0.920), Se = 0.275, Sp = 0.987, Accuracy = 0.925 (95%CI 0.901–0.944), AUPRC = 0.476	SHAP
Flechet 2019		Baseline serum creatinine, cyanotic heart defect pre-surgery, heart frequency, blood pressure, NIRS predictors RMSSD	LR: AUROC = 0.790 (95%CI 0.790–0.800), Se = 0.770, Sp = 0.700, Accuracy = 0.725 (95%CI 0.650–0.789), DCA (clinical usefulness in ranges 16–97%)	-
Parikh 2011		Urine IL-18, urine NGAL, plasma NGAL (within three days after surgery)	LR: AUROC = 0.720 (95%CI 0.638–0.802)	-
Parikh 2013		Urine IL-18 (0–6 h postoperatively), urine L-FABP (second postoperative day), urine NGAL (0–6 h postoperatively)	LR: AUROC = 0.780 (95%CI 0.756–0.804)	-
Jia 2021		SpO2 upon admission the hospital, intraoperative fluid balance, POD0 maximal lactic acid, RPP, POD0 eGFR, POD0 total bilirubin	LR: AUROC = 0.750 (95%CI 0.660–0.820), Se = 0.753, Sp = 0.620, Accuracy = 0.641 (95%CI 0.596–0.683)	-
Dasgupta 2021		Sex, kGFR1	LR: AUROC = 0.720 (95%CI 0.679–0.761), Se = 0.939, Sp = 0.480, Accuracy = 0.824 (95%CI 0.760–0.873)	Formula
Zeng 2022		Perspective (24 h immediately after surgery): Autologous blood transfusion, mechanical ventilation time, FFP transfusion, pressure of oxygen, aortic cross-clamp time, c-reactive protein, platelet count, CPB time, pulse, preoperative LOS, heart rate, hemoglobin, height, lymphocytes, operation time, oxygen saturation, age, platelet distribution width, anion gap	A time-aware attention RNN, AUROC = 0.908 (95%CI 0.907–0.909), Se = 0.872, Sp = 0.816, AUPRC = 0.898 (95%CI 0.896–0.899), Accuracy = 0.832 (95%CI 0.830–0.833)
Cardoso 2016		Serum creatinine, BUN, serum lactate, age	LR: AUROC = 0.909 (95%CI 0.866–0.951), Se = 0.821, Sp = 0.754, Accuracy = 0.762 (95%CI 0.713–0.805)	Formula
Kong 2023		Weight, eGFR, duration of renal ischemia, opening diameter of the PDA, cyanosis, newborn birth	LR: AUROC = 0.840 (95%CI 0.675–0.996), Se = 0.792, Sp = 0.795, Accuracy = 0.547 (95%CI 0.463–0.629), DCA (not reported threshold range)	Nomogram
Shi 2024		Age, preoperative cyanosis, CPB time, albumin, creatinine, cross clamp time	LR: AUROC = 0.755 (95%CI 0.694–0.816), Se = 0.672, Sp = 0.730, Accuracy = 0.720 (95%CI 0.693–0.745), DCA (not reported threshold range)	Nomogram
Tong 2024		Machine ventilation time, STAT score, operative time, mechanical ventilation time classification, anesthesia time, oxygen support, RACHS-1, ABC score, preoperative hematocrit, age, albumin, height, total bilirubin, hemoglobin, creatinine, globulin, weight, red blood cell, urea, white blood cell	LGBM: AUROC = 0.963 (95%CI 0.947–0.979), Se = 0.870, Sp = 0.958, Accuracy = 0.871 (85%CI 0.867–0.875)	SHAP
Fragasso 2023		Serum creatinine, CPB duration, basal creatinine, platelets, and lactate dehydrogenase	RF: AUROC = 0.990 (95%CI 0.98-1), Se = 0.740, Sp = 0.990, Accuracy = 0.857 (95%CI 0.820–0.887)	-
Nagy 2024		Preoperative serum creatinine, surgery duration, POD0 serum pH, POD0 lactate, CPB duration, POD0 vasoactive inotropic score, sex, POD0 hematocrit, Preoperative weight, POD0 serum creatinine, POD0 first 8 h urine output, aorta clamping duration, age, POD0 albumin, gestation weeks, POD0 CVP, cyanotic defect, STAT category, term delivery, single ventricle	LGBM: AUROC = 0.880 (95%CI 0.720-1.00), Se = 0.380, Sp = 0.994, Accuracy = 0.910 (95%CI 0.820-1.00), AUPRC = 0.810 (95%CI 0.610-1.00)	SHAP
Krawczeski 2010		Age, CysC (within postoperative 12 h)	LR: AUROC = 0.810 (95%CI 0.740–0.880), Se = 0.630, Sp = 0.910, Accuracy = 0.821 (95%CI 0.779–0.857)	-
Baloglu 2025		Preoperative serum creatinine, CPB time, Clamp time, Weight, Heart transplantation, Open sternum, gender, Age, STAT, Preterm, Preoperative feeding, chromosomal abnormality, Postoperative extubation in operating room, Single ventricle	LGBM internal validation: AUROC = 0.830 (95%CI 0.807–0.851), Se = 0.840, Sp = 0.580, AUPRC = 0.760, Accuracy = 0.680 (95%CI 0.673–0.735) LGBM external validation: AUROC = 0.750 (95%CI 0.640–0.860), Se = 0.590, Sp = 0.800, AUPRC = 0.750 (95%CI 0.640–0.890), Accuracy = 0.700 (95%CI 0.600–0.820)	SHAP
Deal 2025		IL-18, uromodulin	LR: AUROC = 0.899 (95%CI 0.816–0.981), Se = 0.815, Sp = 0.783, Accuracy = 0.803 (95%CI 0.700-0.877), DCA (not reported threshold range)	-

Note: LASSO: least absolute shrinkage and selection operator

Note: GNB: Gaussian Naive Bayes; GRU: gate recurrent units; LSTM: long short-term memory; MLP: multilayer perceptron; NB: naive bayes; NN: neural networks; SVM: support vector machines; XGBoost: extreme gradient boosting

Note: ABC: Aristotle Basis Complexity; ALT: alanine aminotransferase; ASA: American Society of Anesthesiologists; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CART: classification and regression tree; CK-MB: creatine kinase-MB; CPB: cardiopulmonary bypass; CVP: Central Venous Pressure; CysC: cystatin C; eGFR: estimated glomerular filtration rates; FFP: fresh frozen plasma; FABP: fatty acid binding protein; IL-18: interleukin-18; kGFR: kinetic estimated glomerular filtration rates; LGBM: light gradient boosting machine; LOS: length of stay; LR: logistic regression; MAP: mean arterial pressure; NGAL: neutrophil gelatinase-associated lipocalin; NIRS: near-infrared spectroscopy; PDA: patent ductus arteriosus; POD: postoperative period; RACHS-1: risk adjustment of congenital heart surgery-1; RF: random forest; RMSSD, root mean square of successive differences; RNN: recurrent neural networks; RPP: renal perfusion pressure; Se: sensitivity; Sp: specificity; SHAP: SHapley Additive exPlanations; STAT: Society of Thoracic Surgeons-European Association for Cardiothoracic Surgery congenital heart surgery mortality

Model assessment results

The included prediction models differed in missing data handling, screening of predictors, modeling algorithms, calibration methods, model validation, and model presentation (Table 2). Imputation (5/19[26.3%]) was the most common method for handling missing data. Univariate and multivariate analysis remained the most common methods to screen predictors. Predictors in each model also varied, including demographic characteristics, clinical variables (e.g., laboratory values and vital signs related to surgery), and biomarkers. Most studies (14/19[73.7%]) still used logistic regression to develop models. The most frequently reported measure in performance assessment was the AUROC for discrimination. Seven studies (7/19[36.8%]) not reported model calibration, and the lack of calibration directly impacted the clinical utility of the models. In studies reporting calibration, calibration curves and brier scores are the most commonly used measures, only one study reported the calibration slope. Decision curve analysis (DCA) was used in four studies to evaluate the clinically practical benefit of the models developed [45, 52, 54, 55]. However, only one study reported the threshold range. Most models still lacked a quantitative description of clinical net benefit. All studies were internally validated, but only two were externally validated. Cross-validation was the most commonly used internal validation method, each accounting for 47.4%. The most commonly used external validation method was geographic validation.

Pooled model performance

Nineteen studies included 21 models, of which 19 were internally validated and two were externally validated. Reported AUROC, accuracy and AUPRC in the internal validation set ranged from 0.720 to 0.990, 0.547–0.921 and 0.747–0.898 respectively (Table 2). The pooled AUROC across all models by internal validation was 0.850 (95% CI, 0.810–0.890). Detail shown in forest plot Fig. 2. The pooled AUROC across all models by external validation was 0.830 (95% CI, 0.690–0.980). Detail shown in Appendix B Figure S1. Given that only two external validation models were available, the results of the sensitivity and subgroup analysis lacked sufficient statistical reliability and were therefore omitted from the report. For models with imbalanced outcomes reporting AUPRC values and 95% CI, the pooled AUPRC was 0.900 (95% CI 0.900–0.900). Detail shown in Appendix B Figure S2. We conducted sROC curve for all studies (Appendix B Figure S3), illustrating the change of sensitivity and specificity of models. The pooled sensitivity and specificity were 0.860 (95%CI, 0.750–0.920) and 0.860 (95%CI, 0.750–0.920) respectively. The range of true sensitivity and specificity of these models was reflected by the broad 95% prediction contours.

Fig. 2.

The pooled AUROC for all internal validation models. Note: The red line indicates 95% prediction interval, which is wider than 95% CI

Sensitivity analysis and source of heterogeneity

Pooled AUROC showed high heterogeneity (P < 0.05 and I² = 99%), thus, we explored the stability of the included studies through sensitivity analysis. After using leave-one-out meta-analysis, we found that the pooled results did not show large changes. Given that all included studies were assessed as having a high risk of bias, the stability of the pooled estimates was limited. (Appendix B Table S4).

To further explore potential sources of the high heterogeneity, we conducted ten subgroup analyses of pooled AUROC. Subgroup analysis of pooled AUROC determined the study design as sources of heterogeneity possibly. In the subgroups of AKI definition, AKI stage, outcome and model category, there was significant heterogeneity. However, the test for differences between these subgroups was not significant, suggesting that the observed heterogeneity is likely contributed to by random variation rather than by differences in the true effect sizes across subgroups (Appendix B Table S5 and Figure S4).

Meta-regression analysis was performed for subgroups in pooled AUROC for which the source of heterogeneity could not be identified, and found development methods were identified as possible sources of heterogeneity (Appendix B Table S6 and Figure S5).

Publication bias assessment

We conducted Egger’s test to assess publication bias, plotted funnel and found that there was no publication bias (Appendix B Figure S6). In addition, through using trim and fill method to adjust funnel plot, included studies showed a symmetric distribution (Appendix B Figure S7). In summary, these results suggested that publication bias was not a source of heterogeneity among the studies.

Risk of bias and applicability of studies

Based on the PROBAST tool to assess the ROB and applicability of included studies, indicating all studies had high ROB. Eleven studies were rated as high ROB in the domain of participant selection and twelve studies in the domain of outcome. Seventeen studies were rated as high ROB in the domain of analysis because deficiencies in missing data handling, predictor screening methods, and model validation. In domain of overall applicability, most studies were rated as low concern (Table 3; Fig. 3A-B). The detailed results of the risk of bias and applicability assessments for each study, along with the overall summaries, are reported in Appendix C.

Table 3.

Risk of bias and applicability of included studies

Study	Type	Risk of bias				Applicability	Predictors	Outcome	Overall	Applicability
		Participants	Predictors	Outcome	Analysis	Participants			Risk of bias
Bucholz 2015	A	+	+	+	-	+	+	+	-	+
Bianchi 2013	A	-	+	?	-	+	+	?	-	?
Basu 2014	A	+	+	+	-	+	+	+	-	+
Luo 2023	B	-	+	-	?	+	+	?	-	?
Flechet 2019	A	+	+	-	-	+	+	?	-	?
Parikh 2011	A	+	+	-	-	+	+	-	-	-
Parikh 2013	A	+	+	+	-	+	+	+	-	+
Jia 2021	A	-	+	-	-	+	+	+	-	+
Dasgupta 2021	A	-	+	-	-	+	+	?	-	?
Zeng 2022	A	+	+	+	-	+	+	+	-	+
Cardoso 2016	A	-	+	-	-	+	+	-	-	-
Kong 2023	A	-	+	-	-	+	+	+	-	+
Shi 2023	A	-	+	-	-	+	+	+	-	+
Tong 2024	A	-	+	-	?	+	+	?	-	?
Fragasso 2023	A	-	+	-	-	+	+	?	-	?
Nagy 2024	A	-	+	-	-	+	+	?	-	?
Krawczeski 2010	A	+	+	+	-	+	+	+	-	+
Baloglu 2025	B	+	+	-	-	+	+	+	-	+
Deal 2025	A	-	+	+	-	+	+	+	-	+

A: Study only included internal validation; B: Study included internal and external validation; “+”: Domain was rated as high risk of bias; “-”: Domain was rated as low risk of bias; “?”: Domain was rated as unclear

Fig. 3.

A-B ROB and applicability assessment results. In terms of overall risk of bias, all studies were rated as high risk. Although the predictor domain of all studies was assessed as low risk of bias, over 50% of the studies were rated as high risk in the domains of participants, outcome, and analysis. Regarding overall applicability, more than 50% of the studies were judged to have low concerns regarding applicability (where “low” indicates low concern regarding applicability, “high” indicates high concern, and “unclear” indicates unclear concern). For all studies, the domains of participants and predictors were rated as low concern regarding applicability

Discussion

With the development of machine learning, it is gradually being used in the clinical practice research. Our study found that since 2022, a variety of machine learning methods have been used to develop models in addition to traditional logistic regression. Models developed using machine learning indicated superior performance compared to those developed with traditional logistic regression. But only two of the included studies were externally validated. The lack of external validation indicates the insufficient clinical applicability of models [23, 24, 56]. Thus, external validation of currently existing internally validated models needs to be explored in future studies. Models developed through machine learning need to be presented using a suitable approach. This study found that nomogram and SHAP were the more commonly used approaches. Although nomogram and SHAP have been shown to be suitable presentation formats [57, 58], there is still a need to further develop their clinical utilities, such as through simple computing tools or AI apps [59, 60].

Despite the high consistency of population and outcome definitions across the included studies, there was still large heterogeneity between models. Therefore, we designed subgroup analysis based on Feng’s study [61]. We found that study design was the source of heterogeneity of pooled AUROC possibly. Because prospective studies standardized the definition of the study population and outcome definition, and the rigorous design may lead to a reduction in the sample size included, which in turn led to a narrower distribution of predictors (a smaller range of values or a limited degree of variability) and a decrease in the performance of the developed model [62]. In contrast, retrospective studies developed model based on existing data, which collected a larger sample size, exhibited a wider distribution of predictors, and demonstrated better model performance than prospective studies. In addition, prospective studies had a longer horizon, and during the course of the study, the external environment and medical technology may change, introducing new confounders or altering the mechanism of action of the original factors. However, this did not mean that prediction models developed in retrospective studies were superior to prospective studies, rigorous study design was the key.

We performed meta-regression analysis on subgroups that could not be identified as sources of heterogeneity. Development methods were the sources of heterogeneity in pooled AUROC possibly. Logistic regression was the most traditional development method, and despite its advantages such as interpretability and transparency, other machine learning methods had greater advantages when dealing with large amounts of complex data [63]. Song’s study found that while logistic regression and other machine learning methods had similar performance in developing predictive AKI models, other machine learning methods exhibited greater advantage when dealing with complex data, especially gradient boosting (XGBoost) and Random Forest (RF) [64]. One of the studies included in our research achieved an AUROC of 0.990 using an RF model, approaching the perfect value of 1. Thus, we considered that this difference in the ability to process the data led to differences in the number of predictors included, the distribution of predictors, and the period in which the predictors were included, thus creating heterogeneity among model performance. Although machine learning methods are used to develop models exhibiting favorable AUROC, we must also evaluate model performance comprehensively by integrating predictors screening, internal and external validation methods, calibration, and DCA. Consequently, more studies have constructed multiple models using various machine learning approaches, using logistic regression as the baseline model for comparative analysis to identify the model with the most optimal overall performance.

This study also analyzed the impact of these subgroups on the pooled results. The four subgroups—region, predictor numbers, age and clinical settings —all explained relatively low levels of heterogeneity. In terms of region, the included criteria for participants exhibited a high degree of consistency, despite the participants coming from regions such as the US and China. Consequently, the pooled AUROC of the models constructed across different regions were relatively similar. In terms of predictor numbers, the classification of the number of predictors in this study was based on Feng’s study [61]. The latest guidelines for prediction model development did not specify the number of predictors, indicating that the number of predictors contributed little to explaining heterogeneity [65]. It was essential to focus on the selection process of predictors—identifying those with strong predictive capability or using appropriate screening methods to prevent overfitting when numerous factors were included—thereby improving the quality of the selected predictors. In terms of age, there was a difference in the incidence of AKI when cardiac surgery was performed during the period of mature or immature kidney function, as the time of maturation of renal function was usually 2 years after birth [6]. In cardiac surgery, CPB caused hemodynamic instability, which reduced renal perfusion and led to renal tissue hypoxia [34]. Kidneys were very sensitive to ischemia, and prolonged hypoperfusion caused renal tubular cell injury, which impaired renal filtration and reabsorption, leading to AKI [2]. In cardiac interventional procedures, accumulation of contrast media or other nephrotoxic drugs in the kidneys led to cellular damage, resulting in AKI [66]. Surgery at an age when the kidneys were immature, the kidneys had poorer reserve function and compensatory capacity, and were less tolerant to surgical shock, making them more susceptible to AKI after cardiac surgery. Most of the included studies did not classify the age of the study population as < 2 years or ≥ 2 years, and we reclassified the studies according to their age (median or mean) and found that the pooled AUROC of the model was higher in the subgroup of < 2 years of age, possibly models focused on immature renal development and poor tolerance to surgical stress in capturing and analyzing the association of predictors with AKI, leading to the development of model performance differences between different age groups. This study grouped participants based on the median and mean reported in the included studies. However, the determined groupings may not correspond to all participants. This might have been the reason why age explained a low level of heterogeneity. Therefore, further delineation of age stages is necessary in subsequent prediction model development or validation studies. In terms of clinical settings, differences in clinical settings led to variations in the subsequent treatment patients received after cardiac surgery. We adopted the classification method from Feng’s study to categorize the clinical settings [61]. The setting described in most models typically involved patients transitioning from the ICU back to the general ward after surgery. Some models defined the clinical setting as the period from preoperative preparation through postoperative ICU stay (only data collected preoperatively and during the postoperative ICU stay were included). Undergoing cardiac transplantation or interventional procedures constituted another distinct clinical setting. For instance, specialized treatments received by patients while staying in the ICU postoperatively could influence the determination of predictors, and there were often significant variations in the measurements of the same predictors among patients during their ICU stay. The subsequent supportive care following heart transplantation or interventional surgery also differed from that after cardiac surgery to some extent [67]. Therefore, the clinical environment in which patients were situated could lead to variations in model performance. Ensuring that patients were in similar clinical environments postoperatively could enhance the model’s application value for a specific clinical setting.

FABP, cyanotic heart disease, age, serum creatinine, eGFR, CPB time and surgical risk assessment score were commonly used predictors and were strongly associated with the development of AKI. We found that FABP levels were higher in pediatric patients with AKI compared to pediatric patients without AKI after cardiac surgery. When renal perfusion is reduced, hypoxic stress upregulates the L-FABP gene, causing increased levels of FABP to bind to lipid peroxidation products accumulated in the kidneys and excrete them in the urine, thereby inhibiting renal damage from oxidation products [68]. Similarly, when the heart is hypoperfused during cardiac surgery, resulting in myocardial damage, cardiomyocytes release h-FABP to bind to lipid peroxides accumulated in the heart for transport into the circulation [44]. However, we need to further explore which type of FABP is better at predicting AKI after cardiac surgery in children. One study found that children with cyanotic heart disease were more likely to develop AKI after cardiac surgery than children without cyanotic heart disease [34]. In children with cyanotic heart disease, structural and functional cardiac abnormalities impaired pumping, reduced cardiac output, and disrupted systemic and pulmonary circulation. This, in turn, reduced renal blood flow. The consequent renal hypoxia contributed to the development of AKI [69, 70]. Existing evidence suggested that younger age and longer CPB duration could serve as predictors for AKI onset [34, 71]. We also found that although more models of this review incorporated serum creatinine and eGFR as predictors (creatinine also served as a core predictor in the models that presented SHAP results), their use as key indicators defining AKI may have led to an overestimation of the association with AKI, thereby exaggerating the models’ performance [29]. Models included in this study, over 50% incorporated age as a predictive factor. Given its close association with the maturation of pediatric kidney function, age served as a key demographic predictor [72]. When kidney function is immature and undergoes surgical injury during specific developmental periods, it significantly increases susceptibility to AKI. In this study, CPB duration emerged as a common predictor in both intraoperative and early postoperative models. The length of CPB duration showed a strong association with the occurrence of AKI. Simultaneously, it served as a specific indicator of the cardiac surgical process, thereby possessing high predictive value. Surgical risk assessment was used to predict outcomes in children after cardiac surgery. The Society of Thoracic Surgeons-European Association for Cardiothoracic Surgery congenital heart surgery mortality (STAT) score, STAT category, and Risk adjustment of congenital heart surgery-1 (RACHS-1) category were commonly used risk assessment tools. Categorical tools stratified patients based on the type of surgery, and building models based on this stratification allowed for greater specificity. Scoring systems, on the other hand, performed individualized risk assessments for patients, directly predicting patient outcomes by combining surgery type with patient factors [25, 73, 74]. We also found that biomarkers such as FABP were frequently included in the developed models, which correlates with the recommendation from the 23rd consensus conference of the ADQI to incorporate biomarkers to improve the accuracy of AKI prediction [16], demonstrating the significant value of biomarkers in predicting postoperative AKI, but the clinical utility of biomarkers requires further validation.

Perioperative hemodynamic management, fluid load management, and nephrotoxic drug management serve as common AKI prevention strategies. Meanwhile, models for predicting AKI in patients undergoing cardiac surgery have matured significantly in recent years and have gradually become an important component of AKI prevention strategies [2]. The synergistic integration of controllable strategies and predictive models is key to preventing postoperative AKI. Modifiable factors such as hemodynamics, fluid load, and nephrotoxic drug use also serve as predictors, further contributing to the development of effective prediction models. Although there is currently no single preventive intervention specifically for cardiac surgery-associated AKI, studies have confirmed that intravenous amino acid infusion improves renal hypoperfusion and increases renal plasma flow, effectively reducing the incidence of AKI, though further validation in pediatric cardiac surgery patients is still needed [75, 76]. Therefore, integrating existing prevention strategies, predictive models, and effective interventions to form a comprehensive prevention strategy is a key focus for future research.

Limitations

There are several limitations in this review. First, although the model performance metric of this study (pooled AUROC = 0.860) was good, all included studies were assessed as high risk of bias using the PROBAST tool. This indicated that the model performance may be overly optimistic and failed to reflect its real-world applicability. Sensitivity analysis only indicated that regardless of which study was excluded, the pooled results did not change substantially, it could not indicate that the results were high stability. Secondly, Among the included models, only two models underwent external validation. Most models still lacked external validation, resulting in an inability to assess their generalizability and clinical utility. Third, the high statistical heterogeneity in the meta-analysis suggested that the pooled estimates were an average of widely varying results, which made a single pooled value less informative for any specific clinical setting. This may have been related to various factors during the model development and validation process, and further work was considered necessary to more comprehensively identify the factors contributing to heterogeneity. Finally, most models failed to report calibration comprehensively, which made it difficult to conduct meta-analyses of model calibration and was insufficient for assessing their overall performance. Therefore, future research needs reporting model calibration accurately and comprehensively. Additionally, there was no quality grading tool specifically for predictive modeling research category, this review not used the GRADE guidelines to rate the quality of the literature in terms of level of evidence.

Conclusions

All 19 models included in this systematic review, although included models possess good discriminative ability, the performance was severely overestimated. Thus, this finding must be interpreted with significant caution. The most critical finding of this review was the universal high risk of bias across all included studies, as assessed by PROBAST. This methodological limitation strongly suggested that the reported performance metrics were likely overly optimistic and might not be reproducible in clinical practice. Furthermore, the high statistical heterogeneity observed (I² > 75%) highlighted the substantial variability in model performance, which was likely driven by differences in study design and development methods. Although the included studies all validated the models internally, they still lacked external validation. The current evidence base is insufficient to recommend any of the reviewed models for clinical implementation. Future studies should conduct rigorous study design and produce more transparent research reports in accordance with the TRIPOD or TRIPOD + AI guidelines, enhancing model generalization and clinical utility through external validation. At the same time, through effective model presentation formats, healthcare staff can validate the practical value of these models in clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AKI

Acute kidney injury

AKIN

Acute Kidney Injury Network

AUPRC

Area under the precision recall curve

AUROC

Area under the receiver operating characteristic curve

CPB

Cardiopulmonary bypass

CHARMS

Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies

Cys-C

Cystatin C

DCA

Decision curve analysis

eGFR

Estimated glomerular filtration rate

FABP

Fatty acid–binding protein

GFR

Glomerular filtration rate

HSROC

Hierarchical summary receiver operating characteristics

KDIGO

Kidney Diseases Improving Global Outcomes

KIM-1

Kidney injury molecule-1

NGAL

Neutrophil gelatinase-associated lipocalin

pRIFLE

Pediatric RIFLE

PROBAST

Quality evaluation according to prediction model risk of bias assessment tool

RACHS-1

Risk adjustment of congenital heart surgery-1

RAI

Renal angina index

RF

Random Forest

ROB

Risk of bias

sROC

Summary receiver operating characteristic

STAT

The Society of Thoracic Surgeons-European Association for Cardiothoracic Surgery congenital heart surgery mortality

TRIPOD

Transparent reporting of multivariable prediction models for individual prognosis or diagnosis

TRIPOD + AI

Transparent reporting of multivariable prediction models for individual prognosis or diagnosis+ Artificial intelligence

TRIPOD-SRMA

Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analysis

XGBoost

Gradient boosting

Author contributions

Xuanhao Fan: Writing - original draft, Conceptualization, Methodology, Visualization, Formal analysis, Data Curation, Software. Jiehao Zhuang: Writing - original draft, Conceptualization, Methodology, Visualization, Formal analysis. Ziyi Xiong: Methodology, Visualization, Formal analysis, Data Curation. Zhongqing Chen: Methodology, Visualization, Formal analysis. Niu Yang: Methodology, Visualization, Data Curation. Tungshing Li: Writing - Review & Editing, Supervision, Funding acquisition, Project administration.

Funding

This study is supported entirely by the TL. The whole study was finished without any external financial support.

Data availability

All data used for analysis in this study can be obtained from the original studies. The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethical approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.