Abstract
Acute kidney injury (AKI) after liver transplantation (LT) is a frequent complication associated with adverse outcomes, yet practical early postoperative prediction tools remain limited. This retrospective study developed and internally validated an interpretable model for AKI after LT using perioperative variables available by the time of the first urgent postoperative blood test, generally obtained within 1 h after surgery. Adult LT recipients treated from June 2023 to January 2026 were screened. AKI was defined according to Kidney Disease: Improving Global Outcomes criteria. Among 146 screened recipients, 127 were included and randomly divided into training and validation cohorts at a 3:1 ratio. Least absolute shrinkage and selection operator regression selected estimated glomerular filtration rate calculated using the CKD-EPI equation, anhepatic phase time, postoperative natural logarithm-transformed D-dimer, and postoperative alanine aminotransferase. Multivariable logistic regression showed that lower estimated glomerular filtration rate and longer anhepatic phase time were independently associated with AKI. The nomogram achieved area under the receiver operating characteristic curve values of 0.861 in the training cohort and 0.761 in the validation cohort. Five-fold cross-validation and bootstrap internal validation were performed to assess model stability and optimism. Among five evaluated models, support vector machine showed the numerically highest validation AUC of 0.846. Shapley additive explanations identified estimated glomerular filtration rate as the dominant predictor. This early postoperative model may help identify high-risk recipients for intensified renal monitoring and individualized management. External multicenter validation is required before routine clinical implementation.
Keywords: Acute kidney injury, liver transplantation, machine learning, nomogram, renal function, D-dimer
Introduction
Acute kidney injury (AKI) is one of the most frequent complications after liver transplantation (LT). Its reported incidence varies according to diagnostic criteria, perioperative management, and recipient characteristics. A meta-analysis reported that post-LT AKI occurred in approximately 40.8% of recipients and that severe AKI requiring renal replacement therapy occurred in approximately 7.0% [1]. Recent reviews and multicenter perioperative data confirm that AKI after LT remains a major postoperative complication [2,3]. Beyond the early postoperative period, post-LT AKI has been linked to chronic kidney disease, major adverse kidney events, and poorer long-term outcomes [4,5].
The pathogenesis of AKI after LT is multifactorial. Preoperative vulnerability may result from advanced liver disease, systemic inflammation, malnutrition, infection, cirrhotic cardiomyopathy, baseline renal dysfunction, and altered effective circulating volume [2,6]. Intraoperatively, prolonged anhepatic phase, vascular clamping, blood loss, transfusion, hypotension, venous congestion, and ischemia-reperfusion injury may reduce renal perfusion and intensify tubular and endothelial injury [3,7]. A systematic review also identified potentially modifiable perioperative risk factors, supporting the importance of intraoperative management in renal protection [8]. Additional cohort evidence has linked AKI incidence and severity to clinical and transplant-related factors [9]. Early after transplantation, graft injury, systemic inflammatory activation, and coagulation-fibrinolytic disturbance may further aggravate renal stress. Molecular profiling of the postreperfusion milieu has suggested a link between reperfusion-related inflammatory activity and AKI [10]. Postoperative coagulation parameters within 24 h after LT have also been associated with severe AKI [11]. D-dimer, fibrinogen, and the D-dimer-to-fibrinogen ratio have been reported as renal-risk markers after living donor LT [12–14]. Therefore, accurate early risk assessment requires a model that can integrate clinically available variables from several perioperative domains.
Conventional logistic regression remains useful for clinical prediction because it provides transparent effect estimates and can be converted into a nomogram. However, post-LT AKI may involve nonlinear relationships and interactions that are not fully captured by conventional regression. Several score-based and nomogram-based tools have been developed, including the AKI Prediction Score [15], nomogram models [16,17], and an online calculator [18]. Machine learning approaches have also been explored for post-LT AKI prediction [19,20], including a recent model for recipients receiving grafts from donors after cardiac death [21]. Nevertheless, a recent systematic review and critical appraisal emphasized that many existing models remain limited by risk of bias, insufficient external validation, small validation cohorts, and restricted clinical translation [22]. Recently reported prediction models further indicate ongoing interest in this field but still require broader external validation before routine use [23,24]. Transparent reporting under TRIPOD and TRIPOD+AI guidance is therefore essential for prediction model studies that use regression or machine learning methods [25,26].
Explainable machine learning provides a practical way to improve model transparency. Shapley additive explanations (SHAP) analysis estimates the contribution of each variable to a model output and allows both global and individual-level interpretation [27]. A combined strategy using least absolute shrinkage and selection operator (LASSO)-based feature selection, logistic regression, nomogram construction, machine learning comparison, and SHAP explanation may therefore provide a clinically understandable framework for early postoperative AKI risk stratification after LT. This study aimed to develop and internally validate an interpretable early postoperative risk-stratification model for AKI after LT, focusing on simple perioperative variables available by the time of the first urgent postoperative blood test, generally obtained within 1 h after transplantation.
Materials and methods
Study design and population
This single-center retrospective cohort study was conducted at the Second Affiliated Hospital of Chongqing Medical University. Adult patients who underwent LT between June 2023 and January 2026 were screened. Eligible patients were 18 years or older and underwent LT during the study period.
The exclusion criteria were preoperative hepatorenal syndrome, preoperative AKI or other renal insufficiency, combined liver-kidney transplantation, and active postoperative bleeding after LT. A total of 146 recipients were assessed for eligibility. Nineteen patients were excluded, including 15 with preoperative renal insufficiency, one who underwent combined liver-kidney transplantation, and three with active postoperative bleeding. Finally, 127 patients were included and randomly divided into training and validation cohorts at a 3:1 ratio. The study flowchart is provided as Supplementary Figure S1.
Because this was a retrospective single-center study, the sample size was determined by the number of eligible LT recipients available during the study period rather than by an a priori sample-size calculation. Patients with preoperative hepatorenal syndrome, preoperative AKI, or other renal insufficiency were excluded because the present study focused on new-onset postoperative AKI in recipients without established preoperative renal dysfunction. This approach reduced baseline renal heterogeneity and helped distinguish newly developed postoperative renal injury from persistent or worsening preoperative dysfunction; however, it also limits the generalizability of the model to recipients with preserved preoperative renal function.
The study was approved by the Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University, with Scientific Ethics Pre-Review No. [2025] 361. The study was conducted in accordance with the Declaration of Helsinki, national brain-death donation guidelines, and applicable Chinese laws. Organs were procured from controlled brain-dead donors and allocated through the China Organ Transplant Response System. No organs from executed prisoners were used. Liver transplantations were performed by an experienced transplant team using classic or piggyback techniques according to recipient condition and surgical judgment.
Outcome definition
Post-transplant AKI was defined according to the 2012 Kidney Disease: Improving Global Outcomes (KDIGO) clinical practice guidelines [28]. AKI was diagnosed when at least one of the following criteria was met: an increase in serum creatinine of ≥26.5 μmol/L (0.3 mg/dL) within 48 h after LT; an increase in serum creatinine to ≥1.5 times baseline within 7 days after LT; or urine output <0.5 mL/kg/h for 6 consecutive hours. Baseline serum creatinine was defined as the most recent stable value within 24 h before LT. Unless otherwise stated, all postoperative laboratory indicators in the text, tables, and figures refer to results from the first urgent postoperative blood test after LT, generally obtained within 1 h after surgery.
Postoperative eGFR was calculated using the CKD-EPI equation from serum creatinine measured in the first urgent postoperative blood test after LT, generally within 1 h after surgery. Because serum creatinine is involved in both eGFR calculation and KDIGO-based AKI diagnosis, postoperative eGFR was interpreted as an early postoperative marker for risk stratification rather than as a preoperative or fully independent predictor.
Candidate predictors
Candidate variables included demographic and clinical characteristics, preoperative laboratory variables, intraoperative variables, and early postoperative laboratory variables. Demographic and clinical variables included sex, age group, body mass index (BMI) group, height, weight, smoking history, diabetes mellitus, previous abdominal surgery, hypertension, coronary heart disease, Child-Pugh grade, and Model for End-stage Liver Disease (MELD) score. Preoperative variables included renal function, blood cell counts, liver function, coagulation parameters, inflammatory markers, and composite inflammatory indices. Intraoperative variables included surgical technique, operation time, anhepatic phase time, infusion, urine output, blood loss, plasma transfusion, and red blood cell transfusion. Early postoperative laboratory variables included coagulation parameters, blood cell counts, inflammatory markers, natural logarithm-transformed D-dimer [ln(D-dimer)], alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin, estimated glomerular filtration rate (eGFR), and composite inflammatory indices. ICU stay was summarized as a postoperative clinical outcome and was not used as a candidate predictor for model development. All postoperative laboratory variables were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery.
Postoperative eGFR was calculated using the CKD-EPI equation. In this study, postoperative eGFR referred specifically to the value derived from serum creatinine in the first urgent postoperative blood test after LT, which was generally available within 1 h after surgery.
Statistical analysis and model development
Continuous variables were summarized as mean ± standard deviation or median (25th percentile, 75th percentile), according to distribution. Categorical variables were summarized as number and percentage. Between-group comparisons were performed using Student’s t-test or Welch’s t-test for normally distributed continuous variables, the Wilcoxon rank-sum test for non-normally distributed variables, and the chi-square test or Fisher’s exact test for categorical variables, as appropriate.
In the training cohort, candidate variables associated with AKI and clinically plausible variables were screened using LASSO regression. The optimal penalty parameter was selected using cross-validation, and variables with nonzero coefficients were retained [29]. LASSO-selected variables were entered into a multivariable logistic regression model. Regression coefficients, standard errors, odds ratios (ORs), 95% confidence intervals (CIs), and P values were reported. A nomogram was built from the logistic regression model to provide individualized risk estimation. Model development and reporting were guided by TRIPOD and TRIPOD+AI principles [25,26].
For the final logistic model, four predictors were retained. In the training cohort, 45 AKI events were available, corresponding to an events-per-variable value of 11.25. This value meets the commonly used minimum rule of approximately 10 events per variable, but the small validation cohort still limited the reliability of calibration assessment and model comparison. Therefore, validation results were interpreted cautiously.
Five models were developed and compared: logistic regression, light gradient-boosting machine (LightGBM), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost). Tree-based machine learning models included RF [30], XGBoost [31], and LightGBM [32], while SVM was used as a kernel-based classifier [33]. Model discrimination was assessed using receiver operating characteristic (ROC) curves and area under the curve (AUC) values with 95% confidence intervals (CIs) [34]. Calibration was evaluated using calibration curves as recommended for prediction model assessment [35], and clinical utility was assessed using decision curve analysis (DCA) [36]. Performance was evaluated separately in the training and validation cohorts. To reduce the dependence on a single small validation split, 5-fold cross-validation and bootstrap internal validation were additionally performed on the whole available dataset. The 5-fold cross-validation generated cross-validated ROC curves, calibration curves, and DCA curves. Bootstrap resampling was used to estimate AUC distributions, apparent and optimism-corrected AUCs, and optimism in AUC. Pairwise comparisons of AUCs between models were performed using the DeLong test, and P values were adjusted using the Benjamini-Hochberg procedure for multiple comparisons. These additional analyses were used to strengthen internal assessment, but they were interpreted as exploratory and not as substitutes for independent external validation. Because SVM yielded the numerically highest validation AUC in the original validation cohort, SHAP analysis was applied to explain the SVM model [27]; this model choice was interpreted cautiously because DeLong testing did not demonstrate statistically significant superiority. All analyses were performed using R software. A two-sided p value <0.05 was considered statistically significant.
Results
Study population and baseline characteristics
Among 127 included recipients, 94 were assigned to the training cohort and 33 to the validation cohort. In the training cohort, 45 patients developed AKI and 49 did not. In the validation cohort, 16 patients developed AKI and 17 did not. The incidence of AKI was 47.9% in the training cohort and 48.5% in the validation cohort. Because of the limited validation sample size, this similar event rate was regarded as descriptive and was not interpreted as evidence of external generalizability.
In the training cohort, demographic characteristics and most preoperative clinical variables were comparable between AKI and non-AKI groups. Sex, age group, body mass index group, smoking history, diabetes mellitus, previous abdominal surgery, hypertension, coronary heart disease, Child-Pugh grade, Model for End-stage Liver Disease score, preoperative creatinine, and preoperative eGFR did not differ significantly between groups.
Among intraoperative variables, anhepatic phase time was significantly longer in patients with AKI than in those without AKI [70.00 (58.00, 84.00) vs 59.00 (48.00, 69.00) min; p = 0.004]. Other intraoperative variables, including surgical technique, operation time, infusion, urine output, blood loss, plasma transfusion, and red blood cell transfusion, were not significantly different between groups.
Among early postoperative variables from the first urgent postoperative blood test after LT, generally obtained within 1 h after surgery, patients with AKI in the training cohort had higher postoperative ln(D-dimer) [8.05 (7.05, 8.94) vs 7.54 (6.69, 8.23); p = 0.024], higher postoperative ALT [834.00 (453.00, 1747.00) vs 682.00 (393.00, 914.00) U/L; p = 0.044], and lower postoperative eGFR (79.09 ± 25.48 vs 110.27 ± 22.70 mL/min; p < 0.001). In the validation cohort, lower postoperative eGFR was also observed in the AKI group. Additional differences were observed in inflammatory and coagulation-related markers, including higher preoperative neutrophil count, preoperative monocyte count, preoperative systemic inflammation response index (SIRI), preoperative SIRI-related platelet index (SIRI-P), and postoperative activated partial thromboplastin time (APTT). Detailed univariate results are shown in Table 1.
Table 1.
Univariate analysis of factors associated with acute kidney injury in the training and validation cohorts.
| Training cohort |
Validation cohort |
|||||||
|---|---|---|---|---|---|---|---|---|
| Variables | Overall (n = 94) | Non-AKI (n = 49) | AKI (n = 45) | p value | Overall (n = 33) | Non-AKI (n = 17) | AKI (n = 16) | p value |
| Demographic and clinical characteristics | ||||||||
| Sex | 0.125 | 0.221 | ||||||
| Female | 17 (18.1%) | 6 (12.2%) | 11 (24.4%) | 16 (48.5%) | 10 (58.8%) | 6 (37.5%) | ||
| Male | 77 (81.9%) | 43 (87.8%) | 34 (75.6%) | 17 (51.5%) | 7 (41.2%) | 10 (62.5%) | ||
| Age group | 0.864 | 0.485 | ||||||
| ≤60 years | 83 (88.3%) | 43 (87.8%) | 40 (88.9%) | 31 (93.9%) | 15 (88.2%) | 16 (100.0%) | ||
| >60 years | 11 (11.7%) | 6 (12.2%) | 5 (11.1%) | 2 (6.1%) | 2 (11.8%) | 0 (0.0%) | ||
| BMI group | 0.903 | 0.622 | ||||||
| ≤24 kg/m² | 57 (60.6%) | 30 (61.2%) | 27 (60.0%) | 22 (66.7%) | 12 (70.6%) | 10 (62.5%) | ||
| >24 kg/m² | 37 (39.4%) | 19 (38.8%) | 18 (40.0%) | 11 (33.3%) | 5 (29.4%) | 6 (37.5%) | ||
| Height, cm | 168.00 (161.00, 172.00) | 168.00 (162.00, 172.00) | 168.00 (160.00, 171.00) | 0.417 | 163.27 ± 8.20 | 163.59 ± 8.59 | 162.94 ± 8.03 | 0.824 |
| Weight, kg | 64.04 ± 12.23 | 63.67 ± 11.23 | 64.44 ± 13.34 | 0.762 | 60.00 (53.00, 70.00) | 60.00 (53.00, 65.00) | 61.50 (50.75, 70.00) | 0.745 |
| Smoking history | 0.979 | 1.000 | ||||||
| No | 50 (53.2%) | 26 (53.1%) | 24 (53.3%) | 26 (78.8%) | 13 (76.5%) | 13 (81.2%) | ||
| Yes | 44 (46.8%) | 23 (46.9%) | 21 (46.7%) | 7 (21.2%) | 4 (23.5%) | 3 (18.8%) | ||
| Diabetes mellitus | 0.644 | 1.000 | ||||||
| No | 77 (81.9%) | 41 (83.7%) | 36 (80.0%) | 31 (93.9%) | 16 (94.1%) | 15 (93.8%) | ||
| Yes | 17 (18.1%) | 8 (16.3%) | 9 (20.0%) | 2 (6.1%) | 1 (5.9%) | 1 (6.2%) | ||
| Previous abdominal surgery | 0.905 | 1.000 | ||||||
| No | 34 (36.2%) | 18 (36.7%) | 16 (35.6%) | 9 (27.3%) | 5 (29.4%) | 4 (25.0%) | ||
| Yes | 60 (63.8%) | 31 (63.3%) | 29 (64.4%) | 24 (72.7%) | 12 (70.6%) | 12 (75.0%) | ||
| Hypertension | 0.073 | — | ||||||
| No | 79 (84.0%) | 38 (77.6%) | 41 (91.1%) | 33 (100.0%) | 17 (100.0%) | 16 (100.0%) | ||
| Yes | 15 (16.0%) | 11 (22.4%) | 4 (8.9%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | ||
| Coronary heart disease | 0.496 | 1.000 | ||||||
| No | 92 (97.9%) | 47 (95.9%) | 45 (100.0%) | 32 (97.0%) | 16 (94.1%) | 16 (100.0%) | ||
| Yes | 2 (2.1%) | 2 (4.1%) | 0 (0.0%) | 1 (3.0%) | 1 (5.9%) | 0 (0.0%) | ||
| Child-Pugh grade | 0.372 | 0.735 | ||||||
| A | 22 (23.4%) | 13 (26.5%) | 9 (20.0%) | 7 (21.2%) | 4 (23.5%) | 3 (18.8%) | ||
| B | 31 (33.0%) | 13 (26.5%) | 18 (40.0%) | 16 (48.5%) | 9 (52.9%) | 7 (43.8%) | ||
| C | 41 (43.6%) | 23 (46.9%) | 18 (40.0%) | 10 (30.3%) | 4 (23.5%) | 6 (37.5%) | ||
| MELD score | 15.50 (9.25, 23.75) | 14.00 (9.00, 23.00) | 16.00 (10.00, 25.00) | 0.500 | 12.00 (9.00, 17.00) | 12.00 (9.00, 16.00) | 11.50 (9.75, 20.50) | 0.800 |
| Preoperative laboratory variables | ||||||||
| Preoperative creatinine, μmol/L | 62.51 ± 16.38 | 64.19 ± 16.98 | 60.68 ± 15.69 | 0.302 | 57.20 ± 15.60 | 56.38 ± 17.47 | 58.08 ± 13.87 | 0.761 |
| Preoperative eGFR, mL/min/1.73 m2 | 110.25 (99.17, 119.60) | 110.50 (100.00, 120.80) | 110.00 (96.80, 117.60) | 0.394 | 108.10 (101.20, 117.80) | 107.20 (101.20, 114.10) | 112.05 (101.28, 118.25) | 0.815 |
| Preoperative RBC, 10¹²/L | 3.42 (2.71, 4.12) | 3.44 (2.90, 4.28) | 3.37 (2.64, 3.85) | 0.093 | 3.52 ± 0.76 | 3.43 ± 0.73 | 3.62 ± 0.81 | 0.476 |
| Preoperative hemoglobin, g/L | 104.06 ± 23.93 | 107.65 ± 23.62 | 100.16 ± 23.91 | 0.130 | 108.85 ± 22.37 | 105.35 ± 18.54 | 112.56 ± 25.93 | 0.363 |
| Preoperative platelet count, 10⁹/L | 65.00 (42.00, 95.00) | 69.00 (42.00, 100.00) | 63.00 (46.00, 83.00) | 0.440 | 71.00 (45.00, 107.00) | 71.00 (34.00, 103.00) | 75.50 (47.25, 118.00) | 0.407 |
| Preoperative neutrophil count, 10⁹/L | 2.62 (1.49, 4.34) | 2.71 (1.62, 4.48) | 2.58 (1.46, 3.69) | 0.465 | 2.23 (1.83, 3.58) | 1.87 (1.50, 2.28) | 2.93 (2.14, 3.67) | 0.005 |
| Preoperative lymphocyte count, 10⁹/L | 0.59 (0.39, 0.88) | 0.58 (0.34, 0.98) | 0.59 (0.40, 0.83) | 0.832 | 0.78 (0.52, 1.28) | 0.57 (0.50, 0.90) | 1.16 (0.72, 1.28) | 0.144 |
| Preoperative monocyte count, 10⁹/L | 0.34 (0.22, 0.51) | 0.34 (0.22, 0.51) | 0.34 (0.24, 0.47) | 0.633 | 0.29 (0.24, 0.46) | 0.25 (0.20, 0.34) | 0.33 (0.28, 0.62) | 0.049 |
| Preoperative albumin, g/L | 35.93 ± 5.76 | 36.65 ± 5.45 | 35.14 ± 6.04 | 0.205 | 34.47 ± 6.00 | 35.72 ± 5.30 | 33.13 ± 6.56 | 0.220 |
| Preoperative C-reactive protein, mg/L | 8.79 (5.05, 23.41) | 9.03 (5.00, 23.44) | 7.68 (5.42, 23.31) | 0.709 | 5.76 (5.00, 12.52) | 7.93 (5.00, 12.52) | 5.57 (5.00, 13.51) | 0.779 |
| Preoperative total bilirubin, μmol/L | 47.95 (21.55, 180.65) | 44.00 (17.40, 177.50) | 50.80 (23.90, 181.70) | 0.465 | 37.20 (22.70, 92.90) | 37.20 (22.70, 92.90) | 40.30 (23.15, 127.08) | 0.914 |
| Preoperative prothrombin time, s | 17.75 (14.62, 21.98) | 17.50 (14.30, 22.10) | 18.20 (14.90, 21.90) | 0.711 | 16.20 (14.30, 18.60) | 16.20 (14.30, 17.80) | 16.55 (14.35, 20.90) | 0.564 |
| Preoperative SII | 322.95 (148.77, 577.04) | 308.97 (162.04, 618.72) | 336.93 (146.23, 495.96) | 0.677 | 243.36 (122.22, 386.56) | 206.00 (76.86, 323.70) | 281.55 (195.15, 394.91) | 0.177 |
| Preoperative CLR | 17.54 (9.12, 53.18) | 17.39 (9.80, 48.83) | 22.12 (8.08, 54.03) | 0.886 | 9.77 (5.10, 21.96) | 10.00 (7.14, 26.61) | 7.67 (4.87, 15.67) | 0.368 |
| Preoperative NLR | 4.04 (2.42, 7.42) | 4.04 (2.66, 7.52) | 3.83 (2.40, 6.74) | 0.838 | 3.33 (2.00, 4.41) | 3.33 (1.68, 4.18) | 3.29 (2.28, 4.53) | 0.482 |
| Preoperative PLR | 102.90 (71.67, 177.86) | 93.10 (71.64, 183.67) | 118.37 (72.41, 157.50) | 0.820 | 84.21 (48.98, 112.35) | 90.42 (42.00, 125.37) | 81.26 (58.16, 112.11) | 0.986 |
| Preoperative SIRI | 1.32 (0.57, 3.66) | 1.35 (0.52, 4.17) | 1.12 (0.58, 2.62) | 0.607 | 1.04 (0.55, 1.61) | 0.73 (0.51, 1.04) | 1.37 (1.00, 1.82) | 0.022 |
| Preoperative PNI | 39.05 (34.23, 43.45) | 40.10 (35.05, 43.45) | 38.20 (33.40, 42.85) | 0.223 | 39.50 (34.50, 42.80) | 40.35 (36.45, 42.80) | 37.48 (33.77, 42.42) | 0.407 |
| Preoperative CRP-to-albumin ratio | 0.28 (0.15, 0.60) | 0.28 (0.14, 0.60) | 0.30 (0.16, 0.59) | 0.650 | 0.21 (0.14, 0.43) | 0.21 (0.14, 0.42) | 0.20 (0.15, 0.45) | 0.843 |
| Preoperative SIRI-P | 112.12 (26.89, 295.77) | 122.16 (34.79, 350.14) | 100.07 (25.74, 208.30) | 0.418 | 75.54 (41.33, 136.88) | 48.59 (19.21, 92.77) | 94.49 (71.93, 213.72) | 0.035 |
| Intraoperative characteristics | ||||||||
| Surgical technique | 0.117 | 1.000 | ||||||
| Classic | 80 (85.1%) | 39 (79.6%) | 41 (91.1%) | 28 (84.8%) | 14 (82.4%) | 14 (87.5%) | ||
| Piggyback | 14 (14.9%) | 10 (20.4%) | 4 (8.9%) | 5 (15.2%) | 3 (17.6%) | 2 (12.5%) | ||
| Operation time, min | 432.50 (390.25, 494.25) | 420.00 (380.00, 488.00) | 445.00 (402.00, 500.00) | 0.255 | 465.70 ± 106.81 | 445.59 ± 91.57 | 487.06 ± 120.21 | 0.272 |
| Anhepatic phase time, min | 61.00 (53.00, 78.00) | 59.00 (48.00, 69.00) | 70.00 (58.00, 84.00) | 0.004 | 67.00 (56.00, 72.00) | 67.00 (59.00, 72.00) | 66.50 (55.25, 72.75) | 0.829 |
| Intraoperative infusion, mL | 3900.00 (3300.00, 4700.00) | 4000.00 (3350.00, 4700.00) | 3750.00 (3125.00, 4850.00) | 0.362 | 3800.00 (3400.00, 4875.00) | 3800.00 (3600.00, 4750.00) | 3887.50 (3250.00, 4956.25) | 0.986 |
| Intraoperative urine output, mL | 2150.00 (1600.00, 2787.50) | 2300.00 (1600.00, 2800.00) | 2000.00 (1500.00, 2600.00) | 0.609 | 2261.21 ± 1222.37 | 2518.82 ± 1268.72 | 1987.50 ± 1146.95 | 0.217 |
| Blood loss, mL | 800.00 (600.00, 1000.00) | 800.00 (500.00, 1000.00) | 800.00 (600.00, 1400.00) | 0.153 | 800.00 (600.00, 1000.00) | 800.00 (600.00, 1000.00) | 800.00 (600.00, 1000.00) | 0.956 |
| Intraoperative plasma transfusion, U | 1.00 (0.00, 2.00) | 1.00 (0.00, 2.25) | 1.50 (0.00, 2.00) | 0.936 | 2.00 (0.00, 2.00) | 2.00 (0.00, 2.00) | 0.00 (0.00, 2.25) | 0.371 |
| Intraoperative red blood cell transfusion, U | 3.00 (0.00, 4.38) | 3.00 (0.00, 4.00) | 3.50 (1.50, 6.00) | 0.176 | 3.00 (0.00, 5.00) | 3.00 (0.00, 5.00) | 2.25 (0.00, 4.25) | 0.447 |
| Postoperative laboratory variables | ||||||||
| Postoperative APTT, s | 64.15 (51.25, 77.92) | 60.20 (50.40, 75.80) | 67.50 (51.40, 81.00) | 0.233 | 65.60 (54.60, 79.50) | 55.80 (46.00, 68.20) | 72.45 (64.78, 100.47) | 0.011 |
| Postoperative PTA, % | 37.47 ± 12.52 | 36.57 ± 11.23 | 38.44 ± 13.85 | 0.472 | 38.00 (33.00, 46.00) | 39.00 (35.00, 44.00) | 31.00 (24.25, 47.25) | 0.071 |
| Postoperative C-reactive protein, mg/L | 44.36 (21.14, 80.29) | 57.50 (27.40, 88.17) | 32.13 (14.58, 76.63) | 0.121 | 42.38 (9.34, 69.08) | 33.18 (7.69, 65.79) | 56.38 (19.23, 94.93) | 0.145 |
| Postoperative hemoglobin, g/L | 80.50 (72.00, 90.75) | 81.00 (74.00, 92.00) | 80.00 (69.00, 89.00) | 0.205 | 85.64 ± 14.08 | 83.53 ± 13.39 | 87.88 ± 14.88 | 0.384 |
| Postoperative albumin, g/L | 38.93 ± 5.05 | 39.03 ± 5.27 | 38.82 ± 4.87 | 0.843 | 38.80 (35.20, 40.40) | 39.40 (35.90, 43.70) | 36.60 (34.72, 39.78) | 0.160 |
| Postoperative neutrophil count, 10⁹/L | 8.23 (5.53, 10.22) | 8.29 (6.32, 9.63) | 8.11 (4.89, 11.47) | 0.791 | 7.11 (4.65, 10.10) | 5.88 (4.65, 7.71) | 8.10 (6.28, 12.12) | 0.109 |
| Postoperative lymphocyte count, 10⁹/L | 0.32 (0.18, 0.45) | 0.32 (0.21, 0.46) | 0.32 (0.17, 0.44) | 0.396 | 0.33 (0.23, 0.45) | 0.36 (0.26, 0.41) | 0.26 (0.17, 0.48) | 0.194 |
| Postoperative monocyte count, 10⁹/L | 0.35 (0.24, 0.60) | 0.37 (0.25, 0.60) | 0.34 (0.22, 0.58) | 0.628 | 0.33 (0.24, 0.53) | 0.33 (0.23, 0.44) | 0.35 (0.24, 0.64) | 0.589 |
| Postoperative platelet count, 10⁹/L | 51.00 (33.50, 80.00) | 52.00 (44.00, 83.00) | 38.00 (30.00, 80.00) | 0.072 | 51.00 (34.00, 62.00) | 56.00 (37.00, 81.00) | 43.00 (32.50, 59.25) | 0.304 |
| Postoperative ln(D-dimer) | 7.68 (6.88, 8.59) | 7.54 (6.69, 8.23) | 8.05 (7.05, 8.94) | 0.024 | 7.76 (6.79, 8.26) | 7.73 (6.35, 8.01) | 7.90 (6.96, 9.20) | 0.109 |
| Postoperative AST, U/L | 1588.00 (950.25, 2661.50) | 1419.00 (899.00, 2103.00) | 1823.00 (1012.00, 3541.00) | 0.147 | 1555.00 (849.00, 2706.00) | 1126.00 (849.00, 2128.00) | 2005.00 (1052.75, 4201.75) | 0.109 |
| Postoperative ALT, U/L | 722.50 (430.25, 1193.75) | 682.00 (393.00, 914.00) | 834.00 (453.00, 1747.00) | 0.044 | 699.00 (498.00, 1601.00) | 674.00 (452.00, 762.00) | 923.00 (524.25, 2190.00) | 0.101 |
| Postoperative total bilirubin, μmol/L | 78.55 (42.33, 134.83) | 75.20 (40.20, 134.90) | 79.70 (43.30, 134.00) | 0.392 | 74.40 (53.10, 113.10) | 74.40 (50.30, 111.70) | 72.65 (57.37, 138.15) | 0.528 |
| Postoperative eGFR, mL/min/1.73 m2 | 95.34 ± 28.61 | 110.27 ± 22.70 | 79.09 ± 25.48 | <0.001 | 98.60 (76.90, 110.70) | 109.00 (98.00, 119.00) | 79.90 (52.95, 99.70) | 0.004 |
| Postoperative SII | 1306.39 (781.72, 2487.28) | 1315.20 (1002.29, 2235.73) | 1130.55 (467.72, 2662.85) | 0.291 | 929.61 (578.41, 2198.18) | 797.88 (453.45, 1740.64) | 1274.51 (696.23, 2521.12) | 0.272 |
| Postoperative CLR | 153.24 (48.43, 336.12) | 172.57 (53.73, 349.86) | 135.87 (47.87, 259.44) | 0.414 | 101.32 (51.06, 192.74) | 85.08 (26.53, 186.56) | 127.42 (80.85, 206.96) | 0.145 |
| Postoperative NLR | 24.87 (16.18, 36.69) | 24.58 (17.30, 32.29) | 25.12 (14.31, 37.68) | 0.952 | 21.49 (15.89, 30.91) | 17.73 (11.34, 28.21) | 24.83 (19.62, 48.20) | 0.094 |
| Postoperative PLR | 181.58 (105.57, 291.26) | 186.21 (117.78, 283.78) | 165.00 (77.08, 300.00) | 0.362 | 141.67 (108.70, 248.65) | 150.00 (108.70, 248.65) | 141.52 (115.25, 257.81) | 0.787 |
| Postoperative SIRI | 7.74 (4.29, 14.32) | 7.70 (4.60, 12.81) | 8.87 (4.11, 18.46) | 0.649 | 7.15 (3.98, 13.64) | 6.78 (2.02, 9.70) | 8.27 (5.10, 22.49) | 0.281 |
| Postoperative PNI | 40.87 ± 4.79 | 41.14 ± 4.92 | 40.57 ± 4.67 | 0.565 | 40.10 ± 6.37 | 41.76 ± 5.85 | 38.35 ± 6.60 | 0.126 |
| Postoperative CRP-to-albumin ratio | 1.20 (0.53, 2.25) | 1.44 (0.67, 2.31) | 0.86 (0.44, 2.12) | 0.128 | 1.02 (0.26, 1.92) | 0.92 (0.23, 1.51) | 1.70 (0.52, 3.12) | 0.145 |
| Postoperative SIRI-P | 461.04 (274.94, 1002.94) | 459.93 (303.98, 930.14) | 465.92 (209.81, 1004.36) | 0.607 | 383.35 (139.11, 1040.51) | 213.81 (139.11, 487.38) | 677.25 (212.82, 1089.18) | 0.242 |
| Postoperative clinical outcome | ||||||||
| ICU stay, days | 4.00 (4.00, 5.00) | 4.00 (4.00, 5.00) | 5.00 (4.00, 6.00) | 0.139 | 4.00 (3.00, 6.00) | 4.00 (3.00, 4.00) | 5.00 (4.00, 8.50) | 0.004 |
Note: Data are presented as mean ± standard deviation, median (P25, P75), or n (%), as appropriate. P values compare the AKI and non-AKI groups within each cohort.
Normally distributed continuous variables were compared using Student’s t-test or Welch’s t-test; non-normally distributed variables were compared using the Wilcoxon rank-sum test. Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate.
All postoperative laboratory variables in this table were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery. ICU stay was summarized as a postoperative clinical outcome and was not used for model development.
Abbreviations: AKI, acute kidney injury; ALT, alanine aminotransferase; APTT, activated partial thromboplastin time; AST, aspartate aminotransferase; BMI, body mass index; CLR, C-reactive protein-to-lymphocyte ratio; CRP, C-reactive protein; eGFR, estimated glomerular filtration rate; ICU, intensive care unit; ln(D-dimer), natural logarithm-transformed D-dimer; LT, liver transplantation; MELD, Model for End-stage Liver Disease; NLR, neutrophil-to-lymphocyte ratio; P25, 25th percentile; P75, 75th percentile; PLR, platelet-to-lymphocyte ratio; PNI, prognostic nutritional index; PTA, prothrombin activity; RBC, red blood cell; SII, systemic immune-inflammation index; SIRI, systemic inflammation response index; SIRI-P, SIRI-related platelet index.
Predictor selection and logistic nomogram
LASSO regression selected four predictors with nonzero coefficients: postoperative eGFR, anhepatic phase time, postoperative ln(D-dimer), and postoperative ALT. The three postoperative laboratory predictors were all obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery. These variables reflected four clinically relevant domains: early postoperative renal function, intraoperative surgical and ischemic burden, coagulation-fibrinolysis activation, and early hepatocellular injury.
The selected variables were entered into a multivariable logistic regression model. Postoperative eGFR was independently associated with reduced AKI risk (odds ratio [OR], 0.9507; 95% CI, 0.9276–0.9745; p < 0.001). Anhepatic phase time was independently associated with increased AKI risk (OR, 1.0372; 95% CI, 1.0027–1.0729; p = 0.034). Postoperative ln(D-dimer) showed a positive but non-significant association with AKI (OR, 1.4933; 95% CI, 0.8878–2.5119; p = 0.131). Postoperative ALT was not independently associated with AKI in the multivariable model (OR, 1.0002; 95% CI, 0.9996–1.0008; p = 0.425). All postoperative laboratory variables in this model were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery. The full regression results are shown in Table 2.
Table 2.
Multivariable logistic regression analysis for postoperative acute kidney injury in the training cohort.
| Variable | β | SE | OR (95% CI) | p value |
|---|---|---|---|---|
| Postoperative ln(D-dimer) | 0.4010 | 0.2653 | 1.4933 (0.8878–2.5119) | 0.131 |
| Postoperative eGFR, mL/min/1.73 m² | −0.0505 | 0.0126 | 0.9507 (0.9276–0.9745) | <0.001 |
| Anhepatic phase time, min | 0.0365 | 0.0172 | 1.0372 (1.0027–1.0729) | 0.034 |
| Postoperative ALT, U/L | 0.0002 | 0.0003 | 1.0002 (0.9996–1.0008) | 0.425 |
Note: The model was constructed using the training cohort. Odds ratios and 95% confidence intervals were estimated from multivariable logistic regression. All postoperative laboratory variables in this model were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery. Odds ratios for continuous variables are expressed per 1-unit increase in the corresponding variable. Abbreviations: AKI, acute kidney injury; ALT, alanine aminotransferase; CI, confidence interval; eGFR, estimated glomerular filtration rate; ln(D-dimer), natural logarithm-transformed D-dimer; LT, liver transplantation; OR, odds ratio; SE, standard error.
A nomogram was constructed based on the logistic regression model. The nomogram assigns point values to postoperative eGFR, anhepatic phase time, postoperative ln(D-dimer), and postoperative ALT, and the total score corresponds to the predicted probability of AKI. Because the LASSO curves and the nomogram were combined into one figure, predictor selection and individualized risk estimation are presented together in Figure 1.
Figure 1.
LASSO regression and logistic nomogram for early postoperative acute kidney injury after liver transplantation. (A) LASSO coefficient profiles of candidate predictors. (B) Cross-validation curve for selecting the optimal penalty parameter. (C) Nomogram based on the multivariable logistic regression model. All postoperative laboratory predictors shown in the nomogram were obtained from the first urgent postoperative blood test after liver transplantation, generally within 1 h after surgery. Abbreviations: AKI, acute kidney injury; ALT, alanine aminotransferase; eGFR, estimated glomerular filtration rate; LASSO, least absolute shrinkage and selection operator; ln(D-dimer), natural logarithm-transformed D-dimer.
Model performance
Five prediction models were evaluated in the training and validation cohorts. In the training cohort, RF showed the highest apparent discrimination, with an AUC of 0.959 (95% CI, 0.926–0.993). LightGBM, XGBoost, logistic regression, and SVM achieved AUC values of 0.886 (95% CI, 0.818–0.954), 0.873 (95% CI, 0.799–0.947), 0.861 (95% CI, 0.783–0.939), and 0.851 (95% CI, 0.769–0.934), respectively.
In the validation cohort, SVM yielded the numerically highest discrimination, with an AUC of 0.846 (95% CI, 0.711–0.980). XGBoost showed similar validation performance, with an AUC of 0.842 (95% CI, 0.700–0.984). RF, LightGBM, and logistic regression achieved AUC values of 0.816 (95% CI, 0.661–0.971), 0.811 (95% CI, 0.652–0.970), and 0.761 (95% CI, 0.581–0.941), respectively. Pairwise DeLong tests with Benjamini-Hochberg correction showed that no model was statistically superior to another in the validation cohort; for example, the difference between SVM and XGBoost was minimal (ΔAUC = 0.004; adjusted p = 0.926). Calibration curves showed acceptable overall agreement in the training cohort but greater fluctuation in the validation cohort, likely because of the limited validation sample size. DCA indicated potential clinical net benefit compared with treat-all and treat-none strategies across a range of threshold probabilities (Figure 2).
Figure 2.
Predictive performance and internal validation of five models. (A-C) Receiver operating characteristic curves, decision curve analysis, and calibration curves in the training cohort. (D-F) Corresponding curves in the validation cohort. (G-I) Five-fold cross-validation receiver operating characteristic curves, decision curve analysis, and calibration curves on the whole dataset. (J) Bootstrap AUC density. (K) Bootstrap apparent and optimism-corrected AUC. (L) Bootstrap optimism in AUC. These additional internal validation analyses were performed to assess model stability and optimism under limited sample size and should not be interpreted as a substitute for external validation. Abbreviations: AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; LightGBM, light gradient-boosting machine; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, extreme gradient boosting.
In the whole-dataset 5-fold cross-validation, the AUCs were 0.703 for logistic regression, 0.788 for LightGBM, 0.807 for RF, 0.704 for SVM, and 0.792 for XGBoost. Bootstrap internal validation showed that optimism-corrected AUCs were lower than apparent AUCs for all models, with optimism in AUC ranging from approximately 0.025 to 0.059 (Figure 2). These findings indicate that apparent model performance, especially for flexible algorithms, should be interpreted cautiously under the limited sample size. Pairwise DeLong test results are summarized in Table 3.
Table 3.
Pairwise comparisons of models in the DeLong test.
| Dataset | Model 1 | AUC 1 | Model 2 | AUC 2 | ΔAUC | Adjusted p value (BH) |
|---|---|---|---|---|---|---|
| Apparent whole dataset | Random Forest | 0.941 | SVM | 0.731 | 0.210 | <0.001 |
| Apparent whole dataset | Logistic | 0.729 | Random Forest | 0.941 | −0.212 | <0.001 |
| Apparent whole dataset | Random Forest | 0.941 | XGBoost | 0.881 | 0.060 | <0.001 |
| Apparent whole dataset | LightGBM | 0.888 | SVM | 0.731 | 0.158 | <0.001 |
| Apparent whole dataset | LightGBM | 0.888 | Random Forest | 0.941 | −0.053 | <0.001 |
| Apparent whole dataset | SVM | 0.731 | XGBoost | 0.881 | −0.151 | <0.001 |
| Apparent whole dataset | Logistic | 0.729 | LightGBM | 0.888 | −0.159 | <0.001 |
| Apparent whole dataset | Logistic | 0.729 | XGBoost | 0.881 | −0.152 | <0.001 |
| Apparent whole dataset | LightGBM | 0.888 | XGBoost | 0.881 | 0.007 | 0.382 |
| Apparent whole dataset | Logistic | 0.729 | SVM | 0.731 | −0.001 | 0.948 |
| Training set | Random Forest | 0.959 | SVM | 0.851 | 0.108 | <0.001 |
| Training set | Random Forest | 0.959 | XGBoost | 0.873 | 0.086 | <0.001 |
| Training set | Logistic | 0.861 | Random Forest | 0.959 | −0.098 | <0.001 |
| Training set | LightGBM | 0.886 | Random Forest | 0.959 | −0.073 | 0.001 |
| Training set | Logistic | 0.861 | SVM | 0.851 | 0.010 | 0.201 |
| Training set | LightGBM | 0.886 | XGBoost | 0.873 | 0.013 | 0.201 |
| Training set | LightGBM | 0.886 | SVM | 0.851 | 0.035 | 0.201 |
| Training set | Logistic | 0.861 | LightGBM | 0.886 | −0.025 | 0.301 |
| Training set | SVM | 0.851 | XGBoost | 0.873 | −0.022 | 0.334 |
| Training set | Logistic | 0.861 | XGBoost | 0.873 | −0.012 | 0.536 |
| Validation set | Logistic | 0.761 | SVM | 0.846 | −0.085 | 0.609 |
| Validation set | Logistic | 0.761 | XGBoost | 0.842 | −0.081 | 0.609 |
| Validation set | LightGBM | 0.811 | XGBoost | 0.842 | −0.031 | 0.609 |
| Validation set | Logistic | 0.761 | Random Forest | 0.816 | −0.055 | 0.609 |
| Validation set | Random Forest | 0.816 | XGBoost | 0.842 | −0.026 | 0.609 |
| Validation set | LightGBM | 0.811 | SVM | 0.846 | −0.035 | 0.609 |
| Validation set | Random Forest | 0.816 | SVM | 0.846 | −0.029 | 0.609 |
| Validation set | Logistic | 0.761 | LightGBM | 0.811 | −0.050 | 0.609 |
| Validation set | LightGBM | 0.811 | Random Forest | 0.816 | −0.006 | 0.926 |
| Validation set | SVM | 0.846 | XGBoost | 0.842 | 0.004 | 0.926 |
Note: Pairwise AUC comparisons were performed using the DeLong test. Delta AUC was calculated as the AUC of Model 1 minus the AUC of Model 2. Adjusted p values were corrected using the Benjamini-Hochberg procedure. Significant and non-significant comparisons are both shown to provide transparent model comparison.
Abbreviations: AUC, area under the receiver operating characteristic curve; BH, Benjamini-Hochberg; LightGBM, Light Gradient Boosting Machine; SVM, support vector machine; XGBoost, extreme gradient boosting; ΔAUC, difference in AUC between Model 1 and Model 2.
SHAP interpretation of the SVM model
Because SVM yielded the numerically highest validation AUC in the original validation cohort, SHAP analysis was used to interpret this model; however, this selection should be interpreted cautiously because DeLong testing did not show statistically significant superiority of SVM in the validation cohort and the cross-validation results showed performance variability. Global SHAP feature importance showed that postoperative eGFR was the most influential predictor, with a mean absolute SHAP value of 0.191. Anhepatic phase time ranked second, with a mean absolute SHAP value of 0.092. Postoperative ln(D-dimer) ranked third, with a mean absolute SHAP value of 0.048. Postoperative ALT showed the smallest contribution among the selected variables, with a mean absolute SHAP value of 0.014. All postoperative laboratory variables used in SHAP analysis were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery.
The SHAP beeswarm, heatmap, and dependence plots showed that lower postoperative eGFR, longer anhepatic phase time, and higher postoperative ln(D-dimer) increased the predicted risk of AKI. Representative waterfall plots further illustrated individualized risk prediction. In a high-risk patient, low postoperative eGFR, prolonged anhepatic phase time, and elevated postoperative ln(D-dimer) jointly increased the predicted probability of AKI. In a low-risk patient, preserved postoperative eGFR and shorter anhepatic phase time reduced the predicted probability of AKI (Figure 3).
Figure 3.
SHAP interpretation of the support vector machine model. SHAP feature importance, beeswarm plot, heatmap, dependence plots, and representative waterfall plots show global and individual-level interpretation of the support vector machine model. All postoperative laboratory predictors shown in this figure were obtained from the first urgent postoperative blood test after liver transplantation, generally within 1 h after surgery. Abbreviations: ALT, alanine aminotransferase; eGFR, estimated glomerular filtration rate; ln(D-dimer), natural logarithm-transformed D-dimer; SHAP, Shapley additive explanations; SVM, support vector machine.
Supplementary Figure 1. Patient selection flowchart. Among 146 liver transplantation recipients assessed for eligibility, 19 were excluded, including 15 with preoperative renal insufficiency, one who underwent combined liver-kidney transplantation, and three with active postoperative bleeding. A total of 127 patients were included in the final analysis and randomly divided into the training cohort (n = 94) and validation cohort (n = 33) at a 3:1 ratio. Acute kidney injury was defined according to the 2012 Kidney Disease: Improving Global Outcomes criteria within 7 days after liver transplantation. Baseline serum creatinine was defined as the most recent stable value within 24 h before liver transplantation. Abbreviations: AKI, acute kidney injury; KDIGO, Kidney Disease: Improving Global Outcomes; LT, liver transplantation; SCr, serum creatinine.
Discussion
This study developed and internally validated an interpretable early postoperative prediction model for AKI after LT. Four predictors were selected by LASSO regression: postoperative eGFR, anhepatic phase time, postoperative ln(D-dimer), and postoperative ALT. All postoperative laboratory predictors were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery, so the model should be understood as an early postoperative risk-stratification tool rather than a preoperative prediction model. In multivariable logistic regression, postoperative eGFR and anhepatic phase time remained independently associated with AKI. The logistic model was converted into a nomogram to support individualized risk estimation, while machine learning models were used as complementary predictive tools. Among the tested algorithms, SVM yielded the numerically highest validation AUC in the original validation cohort, but pairwise DeLong testing did not demonstrate statistically significant superiority of SVM or any other model in the validation cohort. The additional 5-fold cross-validation and bootstrap analyses further supported cautious interpretation of apparent performance. SHAP analysis confirmed that postoperative eGFR and anhepatic phase time were the dominant contributors to model prediction.
The AKI incidence in this cohort was approximately 48%, which is consistent with the high burden reported in a meta-analysis [1] and with recent multicenter perioperative evidence [3]. The clinical importance of post-LT AKI lies not only in its frequency but also in its association with early graft and patient outcomes, subsequent chronic kidney disease, and major adverse kidney events [2,4]. Recent data from living donor LT recipients further support the link between AKI and later kidney-related adverse outcomes [5]. Early identification of high-risk recipients is therefore clinically meaningful. Unlike purely preoperative models, the present model uses variables from the first urgent postoperative blood test after transplantation, generally obtained within 1 h after surgery, and is best interpreted as an early postoperative risk-stratification tool. This positioning is important because postoperative variables may capture the cumulative impact of operative stress, reperfusion injury, and immediate postoperative organ function; however, it also means that the model should not be interpreted as predicting AKI before any postoperative renal signal is available.
Postoperative eGFR, calculated using the CKD-EPI equation from the first urgent postoperative blood test generally obtained within 1 h after LT, was the most important predictor in both the logistic regression model and SHAP analysis. This finding is clinically plausible because early postoperative renal function reflects the integrated effects of preoperative susceptibility, intraoperative renal perfusion, ischemia-reperfusion injury, inflammatory activation, and postoperative hemodynamic stability [2,6]. However, postoperative eGFR is closely linked to serum creatinine, which forms part of the AKI diagnostic framework [28]. We therefore acknowledge the possibility of partial overlap between postoperative eGFR and the AKI outcome definition. Accordingly, the model should not be interpreted as a preoperative prediction tool or as a model based entirely on predictors independent of the outcome definition. Instead, it may help clinicians identify recipients who need intensified renal monitoring, optimization of volume and hemodynamics, avoidance of nephrotoxic agents, adjustment of calcineurin inhibitor exposure, and early nephrology consultation.
Anhepatic phase time was independently associated with increased AKI risk. The anhepatic phase is a critical period during LT, during which major vascular clamping, hepatic inflow interruption, venous congestion, and hemodynamic instability may occur. A prolonged anhepatic phase may reflect greater surgical complexity and a longer period of renal hypoperfusion and systemic stress. Large perioperative cohort data and intraoperative risk-factor studies support the contribution of operative stress and modifiable intraoperative factors to AKI after LT [3,7]. Renal injury during this phase may result from reduced effective circulating volume, venous pressure changes, microcirculatory dysfunction, inflammatory mediator release, and reperfusion-associated oxidative stress. The consistent contribution of anhepatic phase time across regression and SHAP analyses suggests that surgical and anesthetic strategies aimed at shortening ischemic burden and maintaining stable renal perfusion may be relevant to AKI prevention.
Postoperative ln(D-dimer), obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery, was selected by LASSO regression and contributed to SVM prediction, although it was not statistically significant in multivariable logistic regression. D-dimer reflects activation of coagulation and fibrinolysis. After LT, elevated D-dimer may represent endothelial injury, thromboinflammation, microvascular disturbance, and reperfusion-related coagulation imbalance. Previous studies have suggested that perioperative D-dimer is associated with AKI after living donor LT [12]. Fibrinogen and the D-dimer-to-fibrinogen ratio may provide related information on coagulation-fibrinolysis imbalance [13,14]. Postoperative coagulation parameters within 24 h after LT have also been associated with severe AKI [11]. The present findings support the view that coagulation-fibrinolysis activation may provide complementary information for early renal risk stratification, even when its independent linear association is attenuated after adjustment.
Postoperative ALT, obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery, was retained by LASSO but showed a relatively small SHAP contribution and was not independently associated with AKI in the multivariable model. Early postoperative ALT elevation likely reflects graft ischemia-reperfusion injury and perioperative hepatocellular injury. Its inclusion in the selected feature set suggests that early graft injury and renal injury may share common pathways, such as hypoperfusion, systemic inflammation, oxidative stress, and endothelial dysfunction. Nevertheless, the weak independent effect indicates that ALT may function more as a marker of global perioperative injury burden than as a dominant renal-specific predictor.
A major strength of this study is the combination of interpretability and predictive modeling. Logistic regression provides transparent ORs and permits construction of a nomogram that can be understood by clinicians. Several post-LT AKI prediction tools have used score-based formats [15], nomogram-based formats [16,17], or a web-calculator format [18], supporting the clinical value of individualized risk estimation. Machine learning models can complement logistic regression by capturing nonlinear patterns in a compact feature set, as shown in previous post-LT AKI machine learning studies [19–21]. In this study, the validation AUCs of SVM and XGBoost were numerically higher than that of logistic regression, but DeLong testing did not show statistically significant superiority of any model in the validation cohort. The additional 5-fold cross-validation and bootstrap analyses suggested that model performance varied across validation strategies and that apparent performance could be optimistic, especially for flexible models. Therefore, the machine learning results should be considered exploratory rather than evidence that one algorithm is definitively superior. For a single-center cohort with limited sample size, internal validation is necessary but not sufficient for clinical implementation, as emphasized by recent systematic appraisal and TRIPOD+AI reporting guidance [22,26].
SHAP analysis improved the interpretability of the SVM model. Instead of presenting machine learning as a black box, SHAP demonstrated how each predictor contributed to both population-level model behavior and individual-level risk estimates [27]. The agreement between SHAP findings and logistic regression results strengthens the biological credibility of the model. Specifically, both approaches emphasized postoperative renal function and anhepatic phase duration as central determinants of AKI risk. This consistency may make the model more acceptable to clinicians and may help translate prediction results into actionable perioperative management.
This study has several clinical implications. First, the model uses a small number of routinely available variables and can therefore be applied soon after LT without specialized biomarkers. Second, the nomogram may help clinicians communicate individualized AKI risk in an intuitive way. Third, recognition of anhepatic phase time as an important predictor emphasizes the need for close communication among surgeons, anesthesiologists, and intensivists during critical operative phases. Fourth, the contribution of postoperative ln(D-dimer) supports further investigation of coagulation-fibrinolysis imbalance and microcirculatory injury as potentially relevant mechanisms of post-LT AKI.
Several limitations should be acknowledged. First, this was a retrospective single-center study with a relatively small sample size, especially in the validation cohort. Our center has only recently developed its liver transplantation program, which limited the number of eligible recipients. In addition, due to patient privacy and ethical restrictions, we did not have access to liver transplantation data from other centers, and no suitable public liver transplantation clinical database with the required perioperative variables was available for external validation. External validation in larger multicenter cohorts is required before clinical application, and this requirement is consistent with current prediction-model appraisal and reporting recommendations [22,25,26]. Second, postoperative eGFR was calculated from serum creatinine measured in the first urgent postoperative blood test, generally within 1 h after LT, and is closely related to the AKI definition; therefore, the present model should be viewed as an early postoperative risk-stratification model rather than a preoperative prediction model. Third, the validation cohort included a limited number of events, which reduced the reliability of calibration assessment and pairwise model comparisons. Although 5-fold cross-validation, bootstrap internal validation, and DeLong testing were added to strengthen the internal assessment, these methods cannot replace independent external validation. Fourth, some potentially important variables were not fully included, such as donor characteristics, cold ischemia time, detailed intraoperative blood pressure trajectories, vasopressor dose, immunosuppressive regimen, nephrotoxic drug exposure, and kidney injury biomarkers. Finally, the study design does not support causal inference, and prospective clinical impact studies are needed to determine whether model-guided interventions can improve renal outcomes after LT.
Conclusions
This study developed and internally validated an interpretable early postoperative prediction model for AKI after LT. LASSO regression identified postoperative eGFR, anhepatic phase time, postoperative ln(D-dimer), and postoperative ALT as key predictors. All postoperative laboratory predictors were obtained from the first urgent postoperative blood test after LT, generally within 1 h after surgery. The logistic nomogram provides a simple and transparent tool for individualized risk estimation. SVM yielded the numerically highest validation AUC among the tested machine learning models, but DeLong testing did not demonstrate statistically significant superiority in the validation cohort. Additional 5-fold cross-validation and bootstrap internal validation supported the exploratory internal assessment of the modeling framework while reinforcing the need for cautious interpretation under limited sample size. SHAP analysis confirmed postoperative eGFR and anhepatic phase time as the most influential contributors to prediction. These findings suggest the potential value of combining conventional regression, machine learning, and explainable artificial intelligence for early postoperative AKI risk stratification after LT, but external validation is needed before routine clinical use.
Supplementary Material
Acknowledgements
The authors thank the clinical staff involved in the perioperative care of liver transplant recipients at the Second Affiliated Hospital of Chongqing Medical University. Xintao Chen and Wen Luo contributed equally to this work. Xintao Chen and Wen Luo contributed to study design, data collection, statistical analysis, machine learning modeling, and manuscript drafting. Mingxiang Cheng and Lingxiang Xu supervised the study, contributed to interpretation of the results, and critically revised the manuscript. All authors read and approved the final manuscript.
Funding Statement
No funding was received for this study.
Ethical approval and consent to participate
This study was approved by the Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University, Scientific Ethics Pre-Review No. [2025] 361. The study was conducted in accordance with the Declaration of Helsinki, national brain-death donation guidelines, and applicable Chinese laws. Because this was a retrospective study using anonymized data, the requirement for written informed consent was waived by the ethics committee.
Consent for publication
Not applicable.
Disclosure statement
No potential conflict of interest was reported by the authors.
Supplemental material
Supplementary Figure S1 is submitted as a separate file. It shows patient screening, exclusions, final inclusion, random split into training and validation cohorts, and the distribution of acute kidney injury in each cohort.
Data availability statement
The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request, subject to institutional and ethical restrictions.
References
- 1.Thongprayoon C, Kaewput W, Thamcharoen N, 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Dong V, Nadim MK, Karvellas CJ.. Post-liver transplant acute kidney injury. Liver Transpl. 2021;27(11):1653–1664. doi: 10.1002/lt.26094. [DOI] [PubMed] [Google Scholar]
- 3.Fiorelli S, Biancofiore G, Feltracco P, et al. Acute kidney injury after liver transplantation, perioperative risk factors, and outcome: prospective observational study of 1681 patients. Minerva Anestesiol. 2022;88(4):248–258. doi: 10.23736/S0375-9393.21.15860-2. [DOI] [PubMed] [Google Scholar]
- 4.Chiu P-F, Lin P-R, Tsai C-C, et al. The impact of acute kidney injury with or without recovery on long-term kidney outcomes after liver transplantation. Nephrol Dial Transplant. 2023;38(8):1817–1822. doi: 10.1093/ndt/gfad005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Chan Y-C, Yeh C-H, Li L-C, et al. Excess risk of major adverse cardiovascular and kidney events after acute kidney injury following living donor liver transplantation. J Clin Med. 2022;11(11):3100. doi: 10.3390/jcm11113100. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Durand F, Francoz C, Asrani SK, et al. Acute kidney injury after liver transplantation. Transplantation. 2018;102(10):1636–1649. doi: 10.1097/TP.0000000000002305. [DOI] [PubMed] [Google Scholar]
- 7.Berkowitz RJ, Engoren MC, Mentz G, et al. Intraoperative risk factors of acute kidney injury after liver transplantation. Liver Transpl. 2022;28(7):1207–1223. doi: 10.1002/lt.26417. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Zhou J, Zhang X, Lyu L, et al. Modifiable risk factors of acute kidney injury after liver transplantation: a systematic review and meta-analysis. BMC Nephrol. 2021;22(1):149. doi: 10.1186/s12882-021-02360-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Zhang D, Liu Z, Bi JF, et al. Risk factors for the incidence and severity of acute kidney injury after liver transplantation. Turk J Gastroenterol. 2021;32(1):82–87. doi: 10.5152/tjg.2020.19592. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Pulitano C, Ho P, Verran D, et al. Molecular profiling of postreperfusion milieu determines acute kidney injury after liver transplantation: a prospective study. Liver Transpl. 2018;24(7):922–931. doi: 10.1002/lt.25178. [DOI] [PubMed] [Google Scholar]
- 11.Chen C, Chen X, Zhao X, et al. Using postoperative coagulation parameters within 24 h after liver transplantation to predict incidence of stage 3 acute kidney injury. Ann Transplant. 2022;27:e937535. doi: 10.12659/AOT.937535. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Park J, Kim SU, Choi HJ, et al. Predictive role of the D-dimer level in acute kidney injury in living donor liver transplantation: a retrospective observational cohort study. J Clin Med. 2022;11(2):450. doi: 10.3390/jcm11020450. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Park J, Joo MA, Choi HJ, et al. Predictive utility of fibrinogen in acute kidney injury in living donor liver transplantation: a propensity score-matching analysis. PLoS One. 2021;16(6):e0252715. doi: 10.1371/journal.pone.0252715. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Park J, Kim M, Kim JW, et al. Predictive value of the D-dimer-to-fibrinogen ratio for acute kidney injury after living-donor liver transplantation. J Clin Med. 2024;13(18):5499. doi: 10.3390/jcm13185499. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Kalisvaart M, Schlegel A, Umbro I, et al. The AKI Prediction Score: a new prediction model for acute kidney injury after liver transplantation. HPB (Oxford). 2019;21(12):1707–1717. doi: 10.1016/j.hpb.2019.04.008. [DOI] [PubMed] [Google Scholar]
- 16.Guo D, Wang H, Lai X, et al. Development and validation of a nomogram for predicting acute kidney injury after orthotopic liver transplantation. Ren Fail. 2021;43(1):1588–1600. doi: 10.1080/0886022X.2021.2009863. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Wu Z, Wang Y, He L, et al. Development of a nomogram for the prediction of acute kidney injury after liver transplantation: a model based on clinical parameters and postoperative cystatin C level. Ann Med. 2023;55(2):2259410. doi: 10.1080/07853890.2023.2259410. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zeng J, Li Q, Wu Q, et al. A novel online calculator predicting acute kidney injury after liver transplantation: a retrospective study. Transpl Int. 2023;36:10887. doi: 10.3389/ti.2023.10887. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Lee H-C, Yoon SB, Yang S-M, 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhang Y, Yang D, Liu Z, 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Liu H-X, Wang X, Xu M-M, 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Huang J, Chen J, Yang J, 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. [DOI] [PubMed] [Google Scholar]
- 23.Wang D, Chen X, Tang Y, et al. 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Chen C, Deng Y, Li Z, et al. Graded nomograms based on perioperative parameters for predicting new-onset severe acute kidney injury following liver transplantation in patients with normal preoperative renal function. Ren Fail. 2025;47(1):2553809. doi: 10.1080/0886022X.2025.2553809. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Collins GS, Reitsma JB, Altman DG, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis: the TRIPOD statement. Ann Intern Med. 2015;162(1):55–63. doi: 10.7326/M14-0697. [DOI] [PubMed] [Google Scholar]
- 26.Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi: 10.1136/bmj-2023-078378. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Lundberg SM, Lee SI.. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30:4765–4774. [Google Scholar]
- 28.Kidney Disease: improving Global Outcomes Acute Kidney Injury Work Group . KDIGO clinical practice guideline for acute kidney injury. Kidney Int Suppl. 2012;2:1–138. [Google Scholar]
- 29.Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Series B. 1996;58(1):267–288. doi: 10.1111/j.2517-6161.1996.tb02080.x. [DOI] [Google Scholar]
- 30.Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. doi: 10.1023/A:1010933404324. [DOI] [Google Scholar]
- 31.Chen T, Guestrin C.. XGBoost: a scalable tree boosting system proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. ; 2016:p. 785–794. [Google Scholar]
- 32.Ke G, Meng Q, Finley T, et al. LightGBM: a highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst. 2017;30:3146–3154. [Google Scholar]
- 33.Cortes C, Vapnik V.. Support-vector networks. Mach Learn. 1995;20(3):273–297. doi: 10.1023/A:1022627411411. [DOI] [Google Scholar]
- 34.DeLong ER, DeLong DM, Clarke-Pearson DL.. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44(3):837–845. [PubMed] [Google Scholar]
- 35.Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21(1):128–138. doi: 10.1097/EDE.0b013e3181c30fb2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Vickers AJ, Elkin EB.. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–574. doi: 10.1177/0272989X06295361. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request, subject to institutional and ethical restrictions.