Neural network-based risk score for acute kidney injury after liver transplantationABSTRACT
Introduction: Acute kidney injury (AKI) after liver transplantation (LT) has a multifactorial origin, making risk prediction challenging. Given its significant impact on prognosis, a reliable and clinically applicable prediction model is needed. The present study aims to develop a risk score for predicting post-LT AKI using an artificial neural network (ANN) model.
Methods: A total of 145 patients who underwent deceased donor LT (DDLT) were analyzed. Data included recipient demographics and comorbidities, donor and graft characteristics, intraoperative variables, and laboratory findings. The primary outcome was postoperative AKI defined by the International Club of Ascites criteria. Independent predictors were identified using logistic regression and incorporated into the ANN model. Model performance was assessed and compared with logistic regression. A risk score was developed based on the β coefficients of the predictors and stratified into risk groups according to ANN outputs.
Results: AKI incidence was 60.6% (88/145). Independent predictors included MELD ≥25, preexisting kidney dysfunction, extended criteria donor grafts, intraoperative arterial hypotension, intraoperative massive transfusion, and serum lactate ≥2 mmol/L at the end of surgery. The ANN model showed better discrimination than logistic regression (AUROC 0.81 vs. 0.71) and good calibration (Hosmer-Lemeshow χ² = 5.57, p = 0.612). Patients were stratified into low (0-6), moderate (7-15), and high (16-22) risk groups, with significantly different AKI incidences.
Conclusions: This ANN-based risk score is a practical tool for early identification of patients at risk of post-LT AKI, supporting timely preventive and therapeutic strategies.
Keywords:
Acute Kidney Injury; Liver Transplantation; Computer Neural Networks
RESUMO
Introdução: A injúria renal aguda (IRA) após transplante hepático (TH) apresenta origem multifatorial, tornando a predição de risco um desafio. Dado seu impacto significativo no prognóstico, há necessidade de um modelo preditivo confiável e aplicável na prática clínica. O presente estudo tem o objetivo de desenvolver um escore de risco para predizer IRA após TH utilizando um modelo de rede neural artificial (RNA).
Métodos: Foram analisados 145 pacientes submetidos a TH com doador falecido (THDF). Os dados incluíram características demográficas e comorbidades do receptor, aspectos do doador e do enxerto, variáveis intraoperatórias e exames laboratoriais. O desfecho primário foi IRA pós-operatória definida pelos critérios do International Club of Ascites. Os preditores independentes foram identificados por regressão logística e incorporados ao modelo de RNA. O desempenho do modelo foi avaliado e comparado ao da regressão logística. Um escore de risco foi desenvolvido com base nos coeficientes β e estratificado em grupos de risco conforme a RNA.
Resultados: A incidência de IRA foi de 60,6% (88/145). Os preditores independentes foram: escore MELD ≥25, disfunção renal prévia, enxertos de doadores com critérios expandidos, hipotensão arterial intraoperatória, transfusão maciça intra-operatória, e lactato sérico ≥2 mmol/L ao final do transplante. O modelo de RNA apresentou melhor discriminação (AUROC 0,81 vs. 0,71) e boa calibração (Hosmer-Lemeshow χ² = 5,57; p = 0,612). Os pacientes foram estratificados em baixo (0-6), moderado (7-15) e alto risco (16-22), com diferenças significativas na incidência de IRA.
Conclusões: Este escore baseado em RNA é uma ferramenta prática para identificação precoce de pacientes com risco de IRA pós-THDF, auxiliando na adoção de estratégias preventivas e terapêuticas.
Palavras-chave:
Injúria Renal Aguda; Transplante Hepático; Redes Neurais de Computação
INTRODUCTION
Acute kidney injury (AKI) following deceased-donor liver transplantation (DDLT) remains a common and clinically consequential complication, associated with increased morbidity, resource utilization, and reduced survival1,2. Despite advances in perioperative management, individualized risk prediction remains suboptimal, reflecting the complex and multifactorial pathophysiology of post-transplant renal dysfunction3.
Machine learning (ML) approaches have increasingly been applied to improve outcome prediction in liver transplantation, frequently demonstrating performance comparable to or exceeding conventional statistical models4,6. However, existing models are heterogeneous, often lack external validation, and remain poorly integrated into clinical decision-making3. In particular, clinically applicable risk scores specifically targeting AKI after DDLT are scarce.
Artificial neural networks (ANNs) are uniquely suited to model nonlinear, high-order interactions among clinical variables without prespecified assumptions7,8. This capability is particularly relevant in liver transplantation, where outcomes are determined by complex interactions between recipient factors, graft characteristics, ischemia-reperfusion injury, and intraoperative physiology9-13.
We hypothesized that an ANN-based approach could improve risk stratification for AKI after DDLT. Accordingly, this study aimed to develop and internally validate an ANN-derived risk score using routinely available perioperative variables.
METHODS
Study Design and Population
This retrospective cohort study included 145 consecutive adult patients undergoing first DDLT between September 2017 and September 2022 allocated according to Model for End-Stage Liver Disease (MELD) score. Inclusion criteria comprised age >18 years, diagnosis of liver cirrhosis with portal hypertension, with or without hepatocellular carcinoma, no evidence of end-stage kidney disease, and requiring a minimum hospital stay of 7 days after DDLT14.
The study was approved by the institutional review board and conducted in accordance with the Declaration of Helsinki15. Data were obtained from medical records, and the study was approved by the Research Ethics Committee of Assis Gurgacz University (approval No. 4.190.165). The requirement for informed consent was waived due to the retrospective design.
Variables and Definitions
AKI, renal function parameters and hepatorenal syndrome were defined according to ICA criteria14. Estimated glomerular filtration rate was calculated using the Modification of Diet in Renal Disease (MDRD)16. Donor and graft quality were assessed using established extended criteria donor definitions, incorporating ischemia times and graft-related risk factors17.
Intraoperative management was guided by continuous invasive hemodynamic monitoring, including mean arterial pression (MAP). Intraoperative hypotension thresholds associated with organ injury were applied (MAP <60 mmHg for ≥5 minutes or any episode of MAP <50 mmHg)18,19. Transfusion strategies followed balanced transfusion protocols20-22. Postreperfusion syndrome was defined according to established criteria23.
Statistical Analysis
Continuous variables were compared using nonparametric tests, and categorical variables using chi-square analysis. Variables with p<0.05 in univariate analysis were entered into a multivariable logistic regression (LR) model.
Model performance was assessed by discrimination (AUROC), calibration (Hosmer-Lemeshow test and calibration plots), and explained variance (Nagelkerke R²). Internal validation was performed using bootstrap resampling. Model optimism was quantified and corrected using bootstrap-derived shrinkage factors.
Artificial Neural Network Model
An ANN model was developed using a feedforward multilayer perceptron architecture with supervised learning and backpropagation. Data were partitioned into training (70%), testing (20%), and validation (10%) sets, with 10-fold cross-validation to mitigate overfitting. Model performance was evaluated using discrimination, accuracy, and error metrics, and compared directly with logistic regression.
Risk Score Development
A clinically applicable risk score was derived from regression coefficients using linear transformation.The selected predictors on LR were used to construct the risk score. The weighted point assignment for each predictor was obtained by linear transformation of its β coefficient, then dividing the corresponding β coefficient by 0.091 [the lowest β value of individual variable (EDC)], multiplied by a constant (2) and rounded to the nearest integer24. Cutoffs for risk stratification were determined using X-tile® software.
Statistical Software
All analyses were performed under the supervision of the Department of Biostatistics of the Universidade Estadual do Oeste do Paraná (UNIOESTE), using SPSS® (version 25.0; IBM Corp., Armonk, NY), the SPSS Neural Networks® module, and the X-tile® software (Yale University, New Haven, CT). Statistical significance was defined as a two-tailed p-value <0.05.
RESULTS
Between september 2017 and september 2022, 145 cases of DDLT were included in the present study. Postoperative AKI occurred in 88 patients (60.6%) within 7 days: 22 (15.1%) developed stage 1, 36 (24.8%) stage 2, and 30 (20.6%) stage 3 AKI. Renal replacement therapy was required in 12 patients (8.7%). Baseline recipient characteristics, donor and graft variables, intraoperative parameters, and laboratory findings stratified by AKI occurrence are summarized in Table 1.
The multivariable LR model performance was acceptable (Nagelkerke R² = 0.663), with adequate calibration (Hosmer-Lemeshow p = 0.247). Six variables were independently associated with postoperative AKI (Table 2).
The ANN model consisted of a single hidden layer with four nodes using a hyperbolic tangent activation function (Figure 1). Predictive performance was robust, with an AUROC of 0.81 (95% CI, 0.75-0.83) and overall accuracy of 0.68. Calibration analysis demonstrated good agreement between predicted and observed outcomes (Hosmer-Lemeshow χ² = 5.57, p = 0.612) (Figure 2).
ANN structural model diagram for AKI after DDLT. IOAH, intra-operative arterial hypotension; MELD, Model for End-stage Liver Disease; KD, kidney dysfunction; MBT, massive blood transfusion; ECD, extended criteria donor; AKI, Acute kidney injury; ANN, artificial neural network; RMSE, root-mean-square error; MAE, mean absolute error.
Calibration graph demonstrating the relationship between predicted probabilities of AKI based on the ANN model and actual values. ANN, artificial neural network.
A point-based risk score was derived from the LR model by linear transformation of β coefficients. Each coefficient was normalized to the smallest β (0.091), multiplied by 2, and rounded to the nearest integer (Table 3). Based on ANN-derived risk stratification, patients were categorized into low- (0-6), moderate- (7-15), and high-risk (16-22) groups ( Figure 3). Both AKI incidence and severity increased significantly risk strata (all p < 0.05) (Figure 4).
The incidence of AKI according to risk score values in the ANN model (n = 145). AKI, acute kidney injury; ANN, artificial neural network.
Relationship between incidence and grade of AKI based upon risk groups stratification by the score points in the ANN model (n = 145). *p < 0.05 in all comparisons between groups. AKI, acute kidney injury; ANN, artificial neural network.
DISCUSSION
In this study, an ANN-based model demonstrated good discriminative performance and calibration for predicting AKI after DDLT, outperforming conventional LR and enabling clinically meaningful risk stratification.
The current literature on AKI prediction in liver transplantation remains heterogeneous. Although multiple ML-based models have been proposed, most lack external validation and exhibit substantial methodological variability³. Reported AUROC values typically range from 0.61 to 0.92, with many studies at high risk of bias3. More recent comparative studies suggest that ML techniques-including gradient boosting, random forests, and neural networks-offer modest but consistent improvements over traditional regression models, with AUROC values generally between 0.75 and 0.805,7. However, translation into clinically actionable tools remains limited.
The present model aligns with this evolving evidence while advancing the field by providing a clinically interpretable, score-based system derived from ANN-informed predictors. Importantly, the identified predictors: MELD score, baseline renal dysfunction, graft quality, intraoperative hemodynamic instability, transfusion burden, and metabolic stress, are consistent with prior mechanistic and clinical studies5,25-27.
From a pathophysiological perspective, AKI after DDLT reflects the convergence of ischemia-reperfusion injury, systemic inflammation, hemodynamic instability, and nephrotoxic exposure. ANN-based models are particularly suited to capture these nonlinear interactions, which are not adequately represented in conventional linear models9,10.
The proposed risk stratification framework may have direct clinical implications. Identification of intermediate- and high-risk patients may enable targeted perioperative strategies, including optimization of hemodynamic support, minimization of nephrotoxic agents, tighter metabolic control, and early consideration of renal replacement therapy28. This approach is aligned with current trends toward actionable, personalized risk prediction.
Despite these strengths, several limitations should be acknowledged. The relatively small sample size increases susceptibility to overfitting, despite internal validation. The absence of external validation limits generalizability, a common limitation in this field3. Additionally, long-term renal outcomes were not assessed. Finally, interpretability remains a critical challenge for ML models. Future work should incorporate explainability techniques (e.g., SHAP analysis) to enhance transparency and clinical adoption.
In conclusion, AKI remains a frequent and clinically significant complication after DDLT1,2. This ANN-derived risk score demonstrated robust predictive performance and enables clinically meaningful risk stratification using routinely available variables. With external validation and further refinement, this approach has the potential to support individualized perioperative management and improve renal outcomes after liver transplantation. Prospective multicenter validation is warranted.
REFERENCES
-
1 Wang D, Chen X, Tang Y, Xiao W, Xu K, Deng J. Development of a risk prediction model for acute kidney injury in liver transplant recipients. Front Surg. 2025;12:1683424. doi: 10.3389/fsurg.2025.1683424.
» https://doi.org/10.3389/fsurg.2025.1683424 -
2 Thongprayoon C, Kaewput W, Thamcharoen N, Bathini T, Watthanasuntorn K, Lertjitbanjong P, et al. Incidence and Impact of Acute Kidney Injury after Liver Transplantation: A Meta-Analysis. J Clin Med. 2019;8(3):372. doi: 10.3390/jcm8030372.
» https://doi.org/10.3390/jcm8030372 -
3 Huang J, Chen J, Yang J, Han M, Xue Z, Wang Y, et al. Prediction models for acute kidney injury following liver transplantation: A systematic review and critical appraisal. Intensive Crit Care Nurs. 2025;86:103808. doi: 10.1016/j.iccn.2024.103808.
» https://doi.org/10.1016/j.iccn.2024.103808 -
4 Lee HC, Yoon SB, Yang SM, Kim WH, Ryu HG, Jung CW, et al. Prediction of Acute Kidney Injury after Liver Transplantation: Machine Learning Approaches vs. Logistic Regression Model. J Clin Med. 2018;7(11):428. doi: 10.3390/jcm7110428.
» https://doi.org/10.3390/jcm7110428 -
5 Zhang Y, Yang D, Liu Z, Chen C, Ge M, Li X, et al. An explainable supervised machine learning predictor of acute kidney injury after adult deceased donor liver transplantation. J Transl Med. 2021;19(1):321. doi: 10.1186/s12967-021-02990-4.
» https://doi.org/10.1186/s12967-021-02990-4 -
6 Liu HX, Wang X, Xu MM, Wang Y, Lai M, Li GM, et al. A new prediction model for acute kidney injury following liver transplantation using grafts from donors after cardiac death. Front Med (Lausanne). 2024;11:1389695. doi: 10.3389/fmed.2024.1389695.
» https://doi.org/10.3389/fmed.2024.1389695 -
7 Avramidou E, Kougianos N, Chiotis-Miehe G, Vasileiadou S, Katsanos G, Tsoulfas G. ChatGPT in liver transplantation: Current applications, limitations, and future directions. World J Transplant. 2026;16(1):110485. doi: 10.5500/wjt.v16.i1.110485.
» https://doi.org/10.5500/wjt.v16.i1.110485 - 8 Haykin S. Neural Networks: Principles and Practice. 3rd ed. Bookman; 2009:176-191.
-
9 Chen C, Chen X, Gao Y, Deng Y, Li Z. Graded nomograms based on perioperative parameters for predicting New-Onset severe acute kidney injury following liver transplantation in patients with normal preoperative renal function: the SALT scale. Ren Fail. 2025;47(1):2553809. doi: 10.1080/0886022X.2025.2553809.
» https://doi.org/10.1080/0886022X.2025.2553809 -
10 Xin W, Yi W, Liu H, Haixia L, Dongdong L, Ma Y, et al. Early prediction of acute kidney injury after liver transplantation by scoring system and decision tree. Ren Fail. 2021;43(1):1137-1145. doi: 10.1080/0886022X.2021.1945462.
» https://doi.org/10.1080/0886022X.2021.1945462 -
11 Lulic I, Gornik I, Pavicic Saric J, Rogic D, Gallego A, Bozic LK, et al. Artificial Intelligence in Post-Liver Transplantation: A Scoping Review of Comparative Model Performance. J Clin Med. 2026 Feb 13;15(4):1491. doi: 10.3390/jcm15041491.
» https://doi.org/10.3390/jcm15041491 -
12 Guo A, Mazumder NR, Ladner DP, Foraker RE. Predicting mortality among patients with liver cirrhosis in electronic health records with machine learning. PLoS One. 2021 Aug 31;16(8):e0256428. doi: 10.1371/journal.pone.0256428.
» https://doi.org/10.1371/journal.pone.0256428 -
13 Nagai S, Nallabasannagari AR, Moonka D, Reddiboina M, Yeddula S, Kitajima T, et al. Use of neural network models to predict liver transplantation waitlist mortality. Liver Transpl. 2022;28(7):1133-1143. doi:10.1002/lt.26442.
» https://doi.org/10.1002/lt.26442 -
14 Angeli P, Ginès P, Wong F, Bernardi M, Boyer TD, Gerbes A, et al. Diagnosis and management of acute kidney injury in patients with cirrhosis: revised consensus recommendations of the International Club of Ascites. Gut. 2015;64:531-537. doi:10.1136/gutjnl-2014-3088.
» https://doi.org/10.1136/gutjnl-2014-3088 - 15 World Medical Association. World Medical Association Declaration of Helsinki: Ethical principles for medical research involving human subjects. JAMA. 2013;310:2191-2194.
-
16 Levey AS, Bosch JP, Lewis JB, Greene T, Rogers N, Roth D. A more accurate method to estimate glomerular filtration rate from serum creatinine. Ann Intern Med. 1999;130:461-470. doi:10.7326/0003-4819-130-6-199903160-00002.
» https://doi.org/10.7326/0003-4819-130-6-199903160-00002 -
17 Neves DB, Rusi MB, Diaz LGG, Salvalaggio P. Primary graft dysfunction of the liver: definitions, diagnostic criteria and risk factors. Einstein (Sao Paulo). 2016;14(4):567-572. doi:10.1590/S1679-45082016RW3585.
» https://doi.org/10.1590/S1679-45082016RW3585 -
18 Salmasi V, Maheshwari K, Yang D, Mascha EJ, Singh A, Sessler DI, et al. Relationship between intraoperative hypotension and acute kidney and myocardial injury after noncardiac surgery. Anesthesiology. 2017;126:47-65. doi:10.1097/ALN.0000000000001432.
» https://doi.org/10.1097/ALN.0000000000001432 -
19 Joosten A, Lucidi V, Ickx B, et al. Intraoperative hypotension during liver transplantation is associated with postoperative acute kidney injury: a historical cohort study. BMC Anesthesiol. 2021;21:12. doi:10.1186/s12871-020-01228-y.
» https://doi.org/10.1186/s12871-020-01228-y -
20 American Society of Anesthesiologists Task Force. Practice guidelines for perioperative blood management. Anesthesiology. 2015;122:241-275. doi:10.1097/ALN.0000000000000463.
» https://doi.org/10.1097/ALN.0000000000000463 -
21 Pacheco LD, Saade GR, Costantine MM, Clark SL, Hankins GD. Massive transfusion protocols in obstetrics. Am J Obstet Gynecol. 2016;214:340-344. doi:10.1016/j.ajog.2015.08.068.
» https://doi.org/10.1016/j.ajog.2015.08.068 -
22 Holcomb JB, Tilley BC, Baraniuk S, et al. Transfusion strategies in trauma: the PROPPR trial. JAMA. 2015;313:471-482. doi:10.1001/jama.2015.12.
» https://doi.org/10.1001/jama.2015.12 -
23 Croome KP, Lee DD, Croome S, et al. The impact of postreperfusion syndrome during liver transplantation using steatotic grafts. Am J Transplant. 2019;19(9):2550-2559. doi:10.1111/ajt.15330.
» https://doi.org/10.1111/ajt.15330 -
24 Rassi A, Little WC, Xavier SS, Rassi SG, Rassi AG, et al. Development and validation of a risk score for predicting death in Chagas' heart disease. N Engl J Med. 2006;355:799-808. doi:10.1056/NEJMoa053241.
» https://doi.org/10.1056/NEJMoa053241 -
25 Khosravi B, Pourahmad S, Bahreini A, Nikeghbalian S, Mehrdad G. Five-year survival of patients after liver transplantation and its effective factors by neural network and Cox proportional hazard regression models. Hepat Mon. 2015;15(9):e25164. doi:10.5812/hepatmon.25164.
» https://doi.org/10.5812/hepatmon.25164 -
26 Ershoff BD, Lee CK, Wray CL, Agopian VG, Urban G, Baldi P, et al. Training and validation of deep neural networks for prediction of 90-day post-liver transplant mortality using UNOS registry data. Transplant Proc. 2020;52(1):246-258. doi:10.1016/j.transproceed.2019.10.019.
» https://doi.org/10.1016/j.transproceed.2019.10.019 - 27 Zhou J, Zhang X, Lyu L, et al. Modifiable risk factors of AKI after liver transplantation: systematic review and meta-analysis. BMC Nephrol. 2021;22:149.
-
28 O'Connor ME, Kirwan CJ, Pearse RM, Prowle JR. Renoprotective strategies in liver transplantation: current evidence and future directions. Transplantation. 2023;107(5):1012-1021. doi:10.1097/TP.0000000000004578
» https://doi.org/10.1097/TP.0000000000004578
