Abstract
Background: The primary objective of this study is to employ machine learning (ML) algorithms to develop predictive models for renal function recovery in critically ill patients undergoing continuous renal replacement therapy (CRRT) due to acute kidney injury (AKI).
Methods: Data were retrospectively collected from patients with AKI who underwent CRRT during their first Intensive Care Unit (ICU) admission from the Medical Information Mart for Intensive Care-IV database and at Huzhou Central Hospital. The evaluation of model performance was conducted utilizing metrics including the area under the receiver operating characteristic curve (AUROC), calibration curves and decision curve analysis (DCA).
Results: A total of 1078 AKI patients undergoing CRRT were included, with a renal function recovery rate of 18.18% (n = 196). Six features that significantly affect renal function recover rate were identified through least absolute shrinkage and selection operator (LASSO) regression with cross-validation. The Lasso-LR model was chosen as the definitive clinical risk prediction model for this study and served as the basis for constructing the nomogram. Its AUROCs in the training and external validation cohorts were 0.774 (95% CI: 0.735 ∼ 0.814) and 0.748 (95% CI: 0.685 ∼ 0.812), respectively. The calibration curve and ideal curve fit were generally found to be suboptimal. Simultaneously, the DCA curve suggested that the nomogram has clinical value.
Conclusions: Among the ML models developed using data from the two datasets to predict renal function recovery in critically ill AKI patients undergoing CRRT, the Lasso-LR model demonstrated the best performance. It can serve as a valuable tool for more efficient clinical disease management and prognostic evaluation.
Keywords: Acute kidney injury, continuous renal replacement therapy, renal function recovery, intensive care unit, machine learning, predictive models
GRAPHICAL ABSTRACT
Introduction
Acute kidney injury (AKI) is a prevalent complication in critically ill patients, with an incidence rate that can reach as high as 30–60% [1,2]. Over the past few decades, there has been a rising trend in AKI incidence rate, which is associated with elevated mortality rates [3,4]. AKI has long been recognized as a global public health issue, exerting a substantial burden on healthcare systems worldwide [5,6].
Following the occurrence of AKI, approximately 5–15% of critically ill patients require renal replacement therapy (RRT). These patients exhibit more severe conditions and experience elevated mortality rates, occasionally surpassing 50% [7,8]. In the ICU, critically ill patients often experience hemodynamic instability, thereby rendering continuous renal replacement therapy (CRRT) the preferred method for RRT in this population [9]. Research has revealed that the renal function recovery in some AKI patients is suboptimal, with about one-quarter of survivors developing acute kidney disease and one-third developing chronic kidney disease (CKD) within the next 5 years [10,11]. These short-term and long-term complications significantly impact patients’ prognosis and quality of life. Therefore, it is important to enhance the investigation of renal function recovery among critically ill AKI patients.
In the domain of severe AKI, previous research has primarily concentrated on constructing risk prediction models aimed at the early prediction of AKI and mortality in AKI patients. However, there has been comparatively limited emphasis on investigating the recovery of renal function [12]. Moreover, there is a significant dearth of research on the recovery of renal function in critically ill AKI patients undergoing CRRT, which hinders early assessment of renal function recovery for these patients. In recent years, the rapid advancement of electronic health records and the growing application of artificial intelligence (AI) in medicine has enabled the construction of various clinical predictive models using extensive data [13]. The integration of “AI+Healthcare” holds the promise of serving as an auxiliary tool to predict renal function recovery in critically ill AKI patients undergoing CRRT. To date, a limited number of studies have developed clinical risk prediction models for renal function recovery in severe AKI patients undergoing RRT [14,15]. Moreover, these studies primarily rely on traditional statistical methods such as Logistic regression and Cox analysis to construct predictive models. Compared to ML techniques, such approaches have limited predictability, hence, might fail to provide effective early assessment of renal function recovery in such patients. This study was conducted to address the lack of current research on the recovery of renal function in patients with AKI undergoing CRRT.
Therefore, we conducted a retrospective study utilizing real-world clinical data of critically ill AKI patients undergoing CRRT in the Intensive Care Unit (ICU) from the Medical Information Mart for Intensive Care-IV (MIMIC IV) database and Huzhou Central Hospital. Furthermore, the principal aim of this study is to utilize ten machine learning (ML) algorithms to develop predictive models for renal function recovery in critically AKI patients who undergo CRRT.
Materials and methods
Data source and study population
The data for this study were obtained from two healthcare institutions: (1) The MIMIC IV database (version V2.0); and (2) Huzhou Central Hospital. The MIMIC-IV database, developed by the Massachusetts Institute of Technology (MIT) Computational Physiology Laboratory, is an open and freely accessible repository that records clinical data of patients admitted to the ICU at Beth Israel Deaconess Medical Center (BIDMC) from 2008 to 2019 [16]. In the MIMIC database, all personal and sensitive patient information has been de-identified, thus effectively safeguarding patient privacy, and hence, no additional informed consent was required. The release of this database received approval from both BIDMC and the Institutional Ethics Review Board affiliated with MIT. The first author of this paper (ZL) successfully completed the “protecting human subjects” examination, thereby gaining the qualification and authorization to utilize this database. Additionally, Huzhou Central Hospital, an urban teaching hospital with 1500 beds, admits approximately 1000 ICU patients annually. This study adhered to the principles outlined in the Helsinki Declaration and obtained approval from the Ethics Committee of Huzhou Central Hospital (approval numbers: 202203021-01, 202212010-01).
This study follows a retrospective clinical research design and TRIPOD statement (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) for model development and validation [17]. The cohort utilized for model development comprised critically ill AKI patients undergoing CRRT in the MIMIC IV database spanning from 2008 to 2019 (n = 758). The external validation cohort, used to evaluate the model, comprised critically ill AKI patients undergoing CRRT who were admitted to the ICU of Huzhou Central Hospital during the period from June 2019 to December 2022 (n = 320).
Selection of participants
In this study, the inclusion criteria were established as follows: (1) Adult critically ill AKI patients who were admitted to the ICU for the first time; (2) Patients who met the diagnostic criteria outlined by the Kidney Disease: Improving Global Outcomes (KDIGO) in 2012; and (3) Patients who received CRRT treatment. Conversely, the exclusion criteria encompassed the following: (1) Individuals aged below 18 years; (2) Patients with multiple admissions to the ICU; (3) Individuals with end-stage renal disease; (4) Patients discharged within 24 h after ICU admission; and (5) Patients with missing essential data or insufficient information.
Definitions
While the MIMIC IV database offered pre-existing baseline creatinine values, the dataset from Huzhou Central Hospital lacked explicit baseline creatinine values. Based on established literature [12,18], we defined the baseline creatinine as follows in this study: (1) the minimum creatinine value from 7 days to 365 days prior to ICU admission; or (2) the minimum creatinine value up to 7 days before ICU admission. In cases where none of the aforementioned creatinine values were available, baseline creatinine was reverse-calculated using the Modification of Diet in Renal Disease (MDRD) equation [19].
According to clinical practice and previous literature [20], oliguria is defined as a 24-h urine output of less than 400 ml in critically ill AKI patients undergoing CRRT. Currently, there is no unified standard for defining renal function recovery in critically ill AKI patients undergoing CRRT. Therefore, in accordance with the Acute Dialysis Quality Initiative Group (ADQI) [21] and previously published literature [22], this study defines renal function recovery based on the following criteria: patients who are discharged alive and no longer require dialysis, and who have not needed RRT for at least 14 days prior to discharge.
Variable extraction
For the training cohort data, we accessed the MIMIC-IV database using PostgreSQL software (version V14.5-1) and Navicat Premium 15 software. We used SQL language to extract the necessary research data. For the external validation cohort data, we manually collected data from the electronic medical record system of Huzhou Central Hospital. The collected information included patients’ demographic characteristics, history of chronic diseases, laboratory results, medication records, disease severity scores, and outcomes, as outlined in Table 1. The timing for collecting laboratory test results and treatment measures was as follows: data were collected prior to the initiation of CRRT treatment (the closest available data point within 48 h before initiation) and before CRRT withdrawal (the closest available data point within 48 h before withdrawal).
Table 1.
Characteristics of patients in the training cohort and external validation cohort (data were collected prior to the initiation of CRRT treatment).
| Characteristic | Total cohort (n = 1078) | Training cohort (n = 758) | External validation cohort (n = 320) | t/Z/χ2 | P value |
|---|---|---|---|---|---|
| Age (years) | 63.79 ± 16.03 | 62.78 ± 15.47 | 66.19 ± 17.06 | −3.200 | 0.001 |
| Gender (%) | 5.348 | 0.021 | |||
| Female | 407 (37.76) | 303 (39.97) | 104 (32.50) | ||
| Male | 671 (62.24) | 455 (60.03) | 216 (67.50) | ||
| Marital status (%) | 90.431 | <0.001 | |||
| Other* | 499 (46.29) | 422 (55.67) | 77 (24.06) | ||
| Married | 579 (53.71) | 336 (44.33) | 243 (75.94) | ||
| KDIGO stage (%) | 4.866 | 0.027 | |||
| Stage 1 + 2 | 310 (28.76) | 203 (26.78) | 107 (33.44) | ||
| Stage 3 | 768 (71.24) | 555 (73.22) | 213 (66.56) | ||
| Baseline creatinine (umol/L) | 88.40 (70.40,98.54) | 88.40 (61.88,98.62) | 90.93 (72.85,98.54) | −2.636 | 0.008 |
| SOFA score (score) | 11.68 ± 4.25 | 12.37 ± 4.05 | 10.03 ± 4.26 | 8.567 | <0.001 |
| Anion gap (mmol/L) | 20.42 ± 6.95 | 20.79 ± 6.24 | 19.55 ± 8.36 | 2.686 | 0.007 |
| WBC (×109/L) | 13.20 (8.40,19.30) | 13.15 (8.80,19.50) | 13.20 (7.75,18.70) | 0.963 | 0.336 |
| RBC (×1012/L) | 3.26 ± 0.82 | 3.16 ± 0.67 | 3.49 ± 1.05 | −6.241 | <0.001 |
| Hemoglobin (g/L) | 98.88 ± 24.28 | 95.48 ± 19.04 | 106.94 ± 32.21 | −7.252 | <0.001 |
| Platelet (×109/L) | 123.50 (72.00,196.00) | 124.00 (71.00,198.00) | 123.00 (75.50,188.50) | 0.903 | 0.367 |
| RDW (%) | 16.19 ± 2.87 | 16.76 ± 2.78 | 14.82 ± 2.63 | 10.665 | <0.001 |
| MCV (fL) | 93.07 ± 8.08 | 92.46 ± 7.97 | 94.51 ± 8.17 | −3.831 | <0.001 |
| MCH (pg) | 30.53 ± 2.82 | 30.44 ± 2.77 | 30.77 ± 2.93 | −1.752 | 0.082 |
| Haematocrit (%) | 30.09 ± 7.27 | 28.96 ± 5.68 | 32.77 ± 9.57 | −8.100 | <0.001 |
| ALT (U/L) | 53.45 (22.30,245.50) | 60.00 (25.00,291.00) | 44.25 (18.95,142.85) | 3.607 | <0.001 |
| AST (U/L) | 104.00 (42.80,444.00) | 116.00 (47.00,531.00) | 77.70 (32.20,265.95) | 4.427 | <0.001 |
| Total bilirubin (umol/L) | 22.60 (10.30,61.56) | 28.22 (10.26,80.37) | 18.20 (10.30,34.55) | 4.809 | <0.001 |
| Glucose (mmol/L) | 7.50 (5.78,10.41) | 7.28 (5.89,9.67) | 8.68 (5.47,12.66) | −2.942 | 0.003 |
| Creatinine (umol/L) | 326.99 (221.00,442.00) | 335.92 (229.84,450.84) | 284.70 (193.10,426.35) | 3.884 | <0.001 |
| BUN (mmol/L) | 20.29 (13.53,31.21) | 20.65 (13.88,30.62) | 19.31 (12.78,31.38) | 1.054 | 0.292 |
| Prothrombin time (s) | 16.30 (13.70,21.90) | 17.00 (14.10,23.40) | 15.20 (13.30,18.50) | 5.472 | <0.001 |
| Total calcium (mmol/L) | 2.00 ± 0.26 | 2.01 ± 0.27 | 1.97 ± 0.22 | 2.115 | 0.035 |
| Sodium (mmol/L) | 137.71 ± 7.29 | 136.73 ± 6.21 | 140.03 ± 8.94 | −6.950 | <0.001 |
| Chlorine (mmol/L) | 102.97 ± 8.57 | 101.03 ± 7.22 | 107.57 ± 9.69 | −12.216 | <0.001 |
| Potassium (mmol/L) | 4.68 ± 0.98 | 4.69 ± 0.93 | 4.65 ± 1.08 | 0.625 | 0.532 |
| Albumin infusion (%) | 456 (42.30) | 254 (33.51) | 202 (63.13) | 80.859 | <0.001 |
| Norepinephrine (%) | 628 (58.26) | 473 (62.40) | 155 (48.44) | 18.041 | <0.001 |
| Furosemide (%) | 492 (45.64) | 358 (47.23) | 134 (41.88) | 2.600 | 0.107 |
| Mechanical ventilation (%) | 876 (81.26) | 676 (89.18) | 200 (62.50) | 105.201 | <0.001 |
AKI: acute kidney injury, CRRT: continuous renal replacement treatment, KDIGO: Kidney Disease Improving Global Outcomes, WBC: white blood cell, RBC: red blood cell, SOFA: score Sequential Organ Failure Assessment score, RDW: red cell distribution width, MCV: mean corpuscular volume, MCH: mean corpuscular hemoglobin, ALT: alanine aminotransferase, AST: aspartate transaminase, BUN: blood urea nitrogen, *unmarried, widowed, divorced status or unknown. data were collected prior to the initiation of CRRT treatment.
Statistical analysis
The study data were processed and analyzed using Stata version 14.0 and R software (version 4.2.3). Variables with missing values exceeding 20% were excluded from the analysis. For variables with missing values below 20%, multiple imputation was performed using the R software mice package (Appendix, Figures 1 and 2). Quantitative data were presented as mean ± standard deviation and analyzed using t-tests if they followed a normal distribution. If the data did not follow a normal distribution, median (interquartile range) was used, and the rank-sum test was employed. A heatmap of the feature correlation matrix was generated for correlation analysis. Qualitative data were expressed as percentages (%) and analyzed using the χ2 test or Fisher’s exact probability method. A detailed description of building and validating the machine-learning model can be found in the Appendix. Statistical significance was set at p < 0.05.
Figure 1.
Patient flow in the study. AKI: acute kidney injury, CRRT: continuous renal replacement treatment, KDIGO: Kidney Disease Improving Global Outcomes, ICU: intensive care unit.
Figure 2.
Comparisons of renal function recovery prediction models in the external validation cohort. AUROC: area under the receiver operating curve, Lasso-LR: LASSO-logistic regression, DT: decision tree, RR: ridge regression, KNN: K-Nearest Neighbor, LightGBM: Light Gradient Boosting Machine, RF: random forest, XGBoost: Extreme Gradient Boosting, SVM: support vector machines, NN: neural network, Ensemble: Ensemble Learning of logistic regression and neural network.
Results
Study population
In the preliminary screening within the MIMIC IV database, a total of 1,393 critically ill AKI patients who underwent CRRT were identified. Ultimately, 758 patients meeting the criteria were included in the training cohort for model development. Within the electronic medical record system of Huzhou Central Hospital, an initial screening identified 698 patients who underwent CRRT for AKI. Among these, 320 patients who met the specified criteria were selected as the external validation cohort for model evaluation. The detailed workflow is illustrated in Figure 1.
The overall population had an age of (63.79 ± 16.03) years, with 407 (37.76%) being female. As shown in Table 1, there were statistically significant differences (all p < 0.05) between the training cohort and the external validation cohort for all parameters except WBC, platelet count, mean corpuscular hemoglobin (MCH), urea nitrogen, potassium, and furosemide, for which no statistically significant differences were observed (all p > 0.05). In Table 2, the differences between the training cohort and the external validation cohort were statistically significant (all p < 0.05), except for hemoglobin, glucose, potassium, renal function recovery, congestive heart failure, chronic obstructive pulmonary disease, acute pancreatitis, cardiac arrest, and norepinephrine (all p > 0.05). For specific details, please refer to Tables 1 and 2.
Table 2.
Characteristics of patients in the training cohort and external validation cohort (data were collected before CRRT withdrawal).
| Characteristic | Total cohort (n = 1078) | Training cohort (n = 758) | External validation cohort (n = 320) | t/Z/χ2 | P value |
|---|---|---|---|---|---|
| anion gap (mmol/L) | 17.85 ± 7.02 | 18.86 ± 6.96 | 15.44 ± 6.57 | 7.502 | <0.001 |
| WBC (×109/L) | 12.50 (8.20,18.30) | 12.90 (8.60,19.00) | 10.95 (7.35,16.40) | 3.776 | <0.001 |
| RBC (×1012/L) | 2.95 ± 0.60 | 2.97 ± 0.52 | 2.88 ± 0.75 | 2.318 | 0.021 |
| Hemoglobin (g/L) | 89.13 ± 17.32 | 89.52 ± 14.62 | 88.23 ± 22.46 | 1.115 | 0.265 |
| platelet (×109/L) | 104.05 (54.00,188.00) | 122.00 (64.00,217.00) | 69.50 (28.00,118.00) | 10.373 | <0.001 |
| RDW (%) | 16.95 ± 3.09 | 17.61 ± 2.98 | 15.39 ± 2.76 | 11.416 | <0.001 |
| MCV (fL) | 93.11 ± 7.02 | 92.49 ± 7.01 | 94.60 ± 6.82 | −4.554 | <0.001 |
| ALT (U/L) | 65.00 (25.00,191.00) | 76.00 (29.00,216.00) | 42.60 (17.50,145.70) | 5.032 | <0.001 |
| AST (U/L) | 74.00 (39.00,242.00) | 81.50 (42.00,254.00) | 60.70 (31.80,177.90) | 4.219 | <0.001 |
| TBil (umol/L) | 29.07 (13.68,73.53) | 32.49 (13.68,92.34) | 23.40 (13.45,46.75) | 3.312 | 0.001 |
| glucose (mmol/L) | 6.94 (5.50,9.39) | 6.86 (5.56,8.89) | 7.19 (5.30,10.18) | −0.990 | 0.322 |
| Creatinine (umol/L) | 176.80 (106.08,282.88) | 203.32 (132.60,309.40) | 124.85 (76.00,205.05) | 9.589 | <0.001 |
| BUN (mmol/L) | 12.46 (7.12,20.29) | 14.60 (8.54,22.07) | 8.50 (5.83,14.75) | 8.068 | <0.001 |
| prothrombin time (s) | 14.90 (12.90,20.20) | 15.60 (12.90,22.50) | 14.10 (12.50,17.00) | 5.202 | <0.001 |
| total calcium (mmol/L) | 2.13 ± 0.24 | 2.15 ± 0.25 | 2.09 ± 0.22 | 3.962 | <0.001 |
| Sodium (mmol/L) | 138.56 ± 5.64 | 137.30 ± 5.30 | 141.53 ± 5.30 | −11.950 | <0.001 |
| Chlorine (mmol/L) | 101.68 ± 6.59 | 99.97 ± 6.38 | 105.72 ± 5.15 | −14.274 | <0.001 |
| Potassium (mmol/L) | 4.31 ± 0.74 | 4.30 ± 0.78 | 4.33 ± 0.64 | −0.519 | 0.604 |
| Magnesium (mmol/L) | 0.88 ± 0.13 | 0.88 ± 0.13 | 0.90 ± 0.15 | −2.189 | 0.029 |
| Renal function recovery (%) | 196 (18.35) | 135 (17.81) | 61 (19.06) | 0.237 | 0.626 |
| Complication or comorbidity[n (%)] | |||||
| hypertension | 372 (34.51) | 240 (31.66) | 132 (41.25) | 9.152 | 0.002 |
| diabetes | 340 (31.54) | 253 (33.38) | 87 (27.19) | 3.993 | 0.046 |
| cerebral infarction | 93 (8.63) | 55 (7.26) | 38 (11.88) | 6.090 | 0.014 |
| Liver cirrhosis | 181 (16.79) | 168 (22.16) | 13 (4.06) | 52.769 | <0.001 |
| Atrial fibrillation | 348 (32.28) | 288 (38.00) | 60 (18.75) | 38.121 | <0.001 |
| CHF | 414 (38.40) | 279 (36.81) | 135 (42.19) | 2.753 | 0.097 |
| COPD | 137 (12.71) | 101 (13.32) | 36 (11.25) | 0.873 | 0.350 |
| Malignant tumor | 139 (12.89) | 115 (15.17) | 24 (7.5) | 11.790 | 0.001 |
| ARF | 655 (60.76) | 487 (64.25) | 168 (52.50) | 13.025 | <0.001 |
| Acute pancreatitis | 88 (8.16) | 61 (8.05) | 27 (8.44) | 0.046 | 0.831 |
| Septic shock | 429 (39.80) | 327 (43.14) | 102 (31.88) | 11.918 | 0.001 |
| AMI | 102 (9.46) | 87 (11.48) | 15 (4.69) | 12.110 | 0.001 |
| Cardiac arrest | 110 (10.20) | 70 (9.23) | 40 (12.50) | 2.618 | 0.106 |
| Cardiogenic shock | 129 (11.97) | 109 (14.38) | 20 (6.25) | 14.118 | <0.001 |
| Mechanical ventilation (%) | 778 (72.17) | 574 (75.73) | 204 (63.75) | 16.067 | <0.001 |
| albumin infusion (%) | 379 (35.16) | 163 (21.50) | 216 (67.50) | 208.815 | <0.001 |
| Norepinephrine (%) | 461 (42.76) | 315 (41.56) | 146 (45.63) | 1.522 | 0.217 |
| Furosemide (%) | 316 (29.31) | 180 (23.75) | 136 (42.50) | 38.190 | <0.001 |
| Oliguria (%) | 513 (47.59) | 337 (44.46) | 176 (55.00) | 10.024 | 0.002 |
WBC: white blood cell, RBC: red blood cell, RDW: red cell distribution width, MCV: mean corpuscular volume, ALT: alanine aminotransferase, AST: aspartate transaminase, BUN: blood urea nitrogen, CHF: congestive heart failure, COPD: chronic obstructive pulmonary disease, ARF: Acute respiratory failure, AMI: acute myocardial infarction, data were collected before CRRT withdrawal.
Correlation between study features
The R corrplot package facilitated the generation of a heat map representing the feature correlation matrix. As revealed in Appendix (Figure 3), there exist several correlations between features, for example, a positive correlation between urea nitrogen and creatinine, and a negative correlation between red blood cell count and mean red blood cell volume.
Figure 3.
Calibration curves of ten machine learning model in the external validation cohort. Lasso-LR: LASSO-logistic regression, DT: decision tree, RR: ridge regression, KNN: K-Nearest Neighbor, LightGBM: Light Gradient Boosting Machine, RF: random forest, XGBoost: Extreme Gradient Boosting, SVM: support vector machines, NN: neural network, Ensemble: Ensemble Learning of logistic regression and neural network, BS: brier score.
Feature extraction
Factors predictive of renal function recovery
We set the renal function recovery status in critically ill AKI patients undergoing CRRT as the dependent variable and the 68 features of the study subjects as independent variables. Using the R glmnet and MLmetrics packages, we conducted feature selection through (1) Least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation analysis and (2) Light Gradient Boosting Machine (LightGBM) feature importance analysis. After validation with logistic regression, and LightGBM, it was found that the features selected through LASSO regression analysis resulted in higher AUROC outcomes. Therefore, LASSO regression was chosen as the feature selection method for this study. With λ = 0.040 as calculated, the following six features were screened, as detailed in Appendix (Figure 4): baseline creatinine, atrial fibrillation, anion gap within 48 h before CRRT withdrawal, oliguria status within 24 h before CRRT discontinuation, and whether mechanical ventilation and norepinephrine treatment were administered within 48 h before discontinuation. Taking into account clinical realities and expert opinions, these six features were ultimately included to construct the corresponding prediction models.
Figure 4.
Nomogram for predicting renal recovery of AKI patients undergoing CRRT. AKI: acute kidney injury, CRRT: continuous renal replacement treatment.
Development of renal recovery prediction model
The above six features were individually introduced into ten ML models: LASSO-Logistic Regression (Lasso-LR), Decision Tree (DT), Ridge Regression (RR), K-Nearest Neighbor (KNN), LightGBM, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), neural network (NN), and Ensemble models. We performed parameter tuning using 5-fold cross-validation, conducted multiple rounds of model training to find the optimal model. The performance metrics of these ten ML models in both the training and external validation cohorts are presented in Appendix (Table 1) and Figures 2 and 3.
In the training cohort, all models achieved an AUROC exceeding 0.75. Among them, the Ensemble model exhibited the highest AUROC, reaching 0.915. The RR model had the lowest AUROC at 0.763. In the external validation cohort, the AUROC of the ten ML algorithm models ranked from highest to lowest as follows (Figure 2): NN model (0.754), Lasso-LR model (0.748), RR model (0.747), LightGBM model (0.735), XGBoost model (0.724), Ensemble model (0.719), RF model (0.682), SVM model (0.667), KNN model (0.667), and DT model (0.608). Calibration curves for different algorithm models, as shown in Figure 3, indicated that the Lasso-LR model and LightGBM model achieved good calibration (Brier Score: 0.139).
Neural network model
In the external validation cohort, the NN model had the highest AUROC of 0.754, indicating a good fit. However, it’s important to note that the calibration curve of the NN model showed suboptimal results (Brier Score: 0.171). As shown in Appendix (Figure 5), baseline creatinine, concomitant atrial fibrillation, anion gap in the 48h before CRRT withdrawal, mechanical ventilation and norepinephrine treatment in the 48h before withdrawal were negatively correlated with the outcome. This means that as baseline creatinine and anion gap increase, as well as the presence of atrial fibrillation, receiving mechanical ventilation, and norepinephrine treatment in the 48h before withdrawal, the likelihood of renal function recovery in critically ill AKI patients undergoing CRRT decreases. On the other hand, urine output in the 24h before CRRT withdrawal was positively correlated with the outcome. In other words, as urine output in the 24h before withdrawal increases (≥400ml), the likelihood of renal function recovery in these patients increases.
Figure 5.
Impact of relevant features on prediction outcome (Lasso-LR: Partial dependence plot). Lasso-LR: LASSO-logistic regression.
The SHAP method was employed to visualize and interpret the NN model, providing insights into the impact of each feature on the outcome. Within the NN model, the relative importance of features affecting renal function recovery in critically ill AKI patients undergoing CRRT was ranked as follows, as shown in Appendix (Figure 6): whether norepinephrine treatment was received in the 48h before withdrawal, atrial fibrillation, urine output in the 24h before CRRT withdrawal, whether mechanical ventilation was received in the 48h before withdrawal, anion gap in the 48h before CRRT withdrawal, and baseline creatinine levels.
Figure 6.
Evaluation of nomogram model performance. (a) Receiver operating characteristic curve of the nomogram; (b) Decision curve analysis of the nomogram; (c) Calibration curves of nomogram in the training cohort; (d) Calibration curves of nomogram in the External validation cohort.
Lasso-LR model
From the results, the Lasso-LR model achieved a relatively high AUROC in the external validation cohort, reaching 0.748 (slightly lower than the NN model, 0.754). Looking at the calibration curve, the Lasso-LR model and LightGBM model performed the best in the external validation cohort (with the smallest Brier score, 0.139). Therefore, the Lasso-LR model was selected as the final clinical risk prediction model in this study. Next, we will visualize this model as a nomogram and evaluate its performance.
Model construction and visualization
The Lasso-LR model, selected after LASSO regression with 10-fold cross-validation, identified a total of 6 features, including baseline creatinine, atrial fibrillation, anion gap in the 48h before CRRT withdrawal, oliguria in the 24h before CRRT discontinuation, and whether or not to receive mechanical ventilation and norepinephrine treatment in the 48h before discontinuation. We conducted both univariate and multivariate logistic regression analyses on these 6 features, and the specific results are presented in Appendix (Table 2). Based on the data from the multivariate logistic regression analysis, we used the R software rms package to construct a nomogram model for predicting renal function recovery in critically ill AKI patients undergoing CRRT (Figure 4). Simultaneously, in order to better apply the model, a dynamic nomogram was developed via the “Shiny” package in R to help clinicians assess renal function recovery in critically ill AKI patients undergoing CRRT (Appendix, Figure 7; https://perfecticudoctor.shinyapps.io/renalfunctionrecovery/).
The visualization and interpretation the Lasso-LR model was assessed by using the SHAP method. As shown in Appendix (Figure 8), the bar chart illustrated the ranking of feature importance for the Lasso-LR model. As shown in Figure 5, the likelihood of recovery of renal function was reduced in patients with comorbid atrial fibrillation, as well as those who were oliguric and treatment with mechanical ventilation and norepinephrine within 48 h before withdrawal. Furthermore, as baseline creatinine and anion gap increase in these patients, the likelihood of renal function recovery in critically ill AKI patients who receive CRRT decreases.
Evaluation of nomogram model performance
ROC curves for training cohort and external validation cohort
In the training cohort, the nomogram exhibited strong discriminatory power in predicting renal function recovery among critically ill AKI patients undergoing CRRT, with an AUROC of 0.774 (95% CI: 0.735 to 0.814). Similarly, in the external validation cohort, it demonstrated excellent discriminative performance, yielding an AUROC of 0.748 (95% CI: 0.685 to 0.812), as illustrated in Figure 6a.
Clinical use of the nomogram
As demonstrated in Figure 6b, DCA reveals that both the training cohort and the external validation cohort consistently outperform the two extreme curves across a broad range of threshold probabilities (0.05–0.50). This suggests that the model holds significant clinical utility.
Calibration curves for the training cohort and external validation cohort
The calibration curves for both the training cohort and the external validation cohort are depicted in Figure 6c and d. The calibration slopes and Brier scores are 1.000 and 0.126 for the training cohort, and 0.882 and 0.139 for the external validation cohort, respectively. The calibration curve indicated that the predicted probabilities matched the actual probabilities.
Discussion
In daily clinical practice, AKI is a frequently encountered clinical syndrome [23]. For AKI patients who survive, the restoration of renal function is a crucial clinical endpoint. Research indicates that even with RRT treatment, the rate of renal function recovery in AKI patients remains suboptimal, with 30% of survivors progressing to CKD within the subsequent 5 years [24]. Consequently, forecasting the recovery of renal function in critically ill AKI patients undergoing CRRT is of utmost clinical and public health importance. However, accurately predicting renal function recovery in critically ill patients with AKI undergoing CRRT presents a significant challenge for clinicians. At present, with the help of sophisticated ML algorithms, clinical risk prediction models are developed that can be used to make predictions about renal function recovery of these patients.
In this study, we developed a set of ten risk prediction models designed to forecast renal function recovery in critically ill AKI patients receiving CRRT. Ultimately, we selected the Lasso-LR model as our visualized clinical risk prediction model, which facilitates medical decision-making through the use of a dynamic nomogram. From the constructed model, it is evident that patients with critical AKI undergoing CRRT are more likely to experience renal function recovery if they exhibit the following characteristics: lower baseline blood creatinine levels, absence of comorbid atrial fibrillation, lower anion gap values before CRRT withdrawal, absence of comorbid oliguria before CRRT withdrawal, and no requirement for mechanical ventilation or norepinephrine treatment before CRRT withdrawal. In clinical practice, attention should be directed towards patients who do not exhibit these characteristics. Significantly, regarding the modifiable risk factors (anion gap, urine output, mechanical ventilation and norepinephrine) identified in the model, early identification of patients whose renal function may be challenging to recover is essential, and efforts should be made to address these risk factors as much as possible, including correction of acidosis, optimizing patients’ hemodynamics, and closely monitoring changes in urine output. Additionally, non-modifiable risk factors (baseline creatinine and atrial fibrillation) should also be addressed in clinical settings, especially for high-risk populations, so that their AKI treatment can be optimized for better prognosis [25]. Furthermore, it is imperative to enhance post-discharge monitoring of high-risk patients who survive severe AKI in order to prevent recurrence of AKI as well as progression to CKD.
A systematic review regarding the long-term prognosis of AKI patients has indicated that higher baseline blood creatinine levels serve as a predictor for CKD progression in these patients [26]. From a clinical perspective, the recovery of urine output is considered a significant indicator of renal function recovery. Several previous studies have likewise demonstrated that the recovery or increase in urine output may signify renal function recovery in severe AKI patients undergoing CRRT [27,28]. More recently, a guideline has suggested that spontaneous improvement in urine output, including transitioning from an oliguric state to a non-oliguric state (>400 ml/day) or achieving a urinary creatinine clearance exceeding 15–20 mL/minute, can be serve as reliable indicators of RRT weaning [29].
The level of anion gap before CRRT withdrawal was found to be one of determining factors in the recovery of renal function in AKI patients treated with CRRT. Acid-base imbalances are commonly observed in the ICU, with metabolic acidosis being the most prevalent [30,31]. High anion gap metabolic acidosis is a significant subtype of metabolic acidosis, which is commonly caused by sepsis, diabetes, alcohol consumption, and drug use [32]. Metabolic acidosis may have a role in the development of AKI. Several animal studies have demonstrated that metabolic acidosis can induce or exacerbate AKI [33,34]. Also, comorbid atrial fibrillation stands out as a crucial factor influencing the recovery of renal function in critically ill AKI patients undergoing CRRT. A recent extensive prospective cohort study has established atrial fibrillation as a causal risk factor for renal impairment [35]. Additionally, receiving mechanical ventilation and norepinephrine therapy prior to CRRT withdrawal also plays a significant role in determining renal function recovery in these patients. Indeed, previous research results indicate that receiving mechanical ventilation [15,36]and the administration of vasoactive drugs [37,38] independently contribute to the failure of renal function recovery in such patients. The precise pathophysiological mechanisms remain elusive, but it is plausible that CRRT patients requiring mechanical ventilation and norepinephrine treatment are inherently more critically ill, making renal function recovery challenging. The correlation between these factors warrants further exploration in future studies. Different studies have reported different rates of renal function recovery, possibly due to differences in sample size, different causes of AKI and different definitions of renal function recovery.
This study holds several strengths. Firstly, in our previous study [39], the study suggested that a higher albumin corrected anion gap level (> 20 mmol/L) at the initiation of CRRT treatment is associated with ICU all-cause mortality among critically AKI patients requiring CRRT. The present study builds upon and extends the scope of previous research. There is a notable scarcity of research on modeling renal function recovery in critically ill AKI patients undergoing CRRT. Our research employs ten ML algorithms to construct and assess predictive models for renal function recovery in severe AKI patients undergoing CRRT. Secondly, the SHAP method was employed to interpret and visualize the corresponding model in order to solve the concerns about “black box” in some degree. Thirdly, several of the risk factors identified in our study are modifiable. Future research regarding whether early recognition and correction of these risk factors can enhance renal function recovery in these patients need to be carried out. Finally, the developed model is subjected to external validation. The outcomes of spatial external validation indicate that the model exhibits robust discriminatory performance and practical utility in forecasting renal function recovery in these patients, which also suggests that the model possesses a degree of transferability and generalizability.
Certainly, this study comes with certain limitations. Firstly, as a retrospective study, some inherent biases are hard to avoid. In cases where baseline creatinine values were not available, the baseline creatinine was reverse-calculated using the MDRD equation in our study. It is worth noting that this approach may introduce bias into this study. Secondly, there are numerous factors influencing whether renal function can recover after AKI, such as fluid overload, CRRT treatment mode, treatment duration, and the occurrence of hypotension during treatment, among others. Unfortunately, data on these factors were not available in our datasets. Thirdly, novel biological markers of kidney injury (Neutrophil Gelatinase-Associated Lipocalin and cystatin C) and critical ultrasound parameters like the renal artery resistance index were missing. In the future, as detection of these biomarkers and other techniques get advanced and gain broader clinical acceptance, further optimization of the model parameters will be necessary to enhance the predictive performance of the models. Fourthly, different studies have reported different rates of renal function recovery, possibly due to differences in sample size, different causes of AKI and different definitions of renal function recovery. Therefore, the results of our study may need to be confirmed by further large studies, particularly long-term follow-up studies. Furthermore, although the AUROC of the model is high in our study, the Kappa values of most models are relatively low, especially in the external validation set. Besides, the calibration effect of the Lasso-LR model is not satisfactory, and there may be some unmeasured confounding biases. The possible reason for this discrepancy is that the data came from two independent hospitals, and belonged to different ethnic groups. In the MIMIC-IV database, the racial/ethnic composition of the participants is predominantly White, followed by African American, Asian, and other ethnicities. Whereas, all of the participants in the Huzhou Central Hospital are Asian. Therefore, other validation cohorts would encourage to validate this nomogram in the future. Finally, due to the retrospective nature of the study, it only constructed a predictive model for short-term renal function recovery in patients undergoing CRRT for severe AKI. Long-term renal function recovery in these patients was not predicted. In future research, there should be a stronger focus on predicting and assessing long-term renal function prognostic outcomes for these patients.
Conclusion
In conclusion, utilizing the Lasso-LR model, six features were identified, and subsequently employed in the construction of a nomogram. Through a series of comprehensive evaluations, this discriminative model has exhibited excellent performance, high predictive value, and applicability in clinical practice, suggesting this model can be used to assist healthcare professionals in early prognosis assessment for severe AKI patients undergoing CRRT.
Supplementary Material
Acknowledgements
The Graphical Abstract of our study is drawn by Figdraw.
Funding Statement
This study was funded by the Key Research and Development Plan of Zhejiang province (2019C03024), Provincial and ministerial joint construction of key projects of Zhejiang Medical and Health Science and Technology Plan (WKJ-ZJ-1811) and the Medical and Health Science and Technology Project of Zhejiang Province (2023KY314).
Ethical statement and consent to participate
The database was approved by the institutional review boards of the BIDMC and the Massachusetts Institute of Technology. The study was approved by the Institutional Review Board of Huzhou Central Hospital (approval number: 202203021-01, 202212010-01) with a waiver of informed consent because of the anonymous nature of this study.
Consent for publication
Not applicable.
Disclosure statement
The authors declare that they have no competing interests.
Data availability statement
The data are available in the physionet (https://physionet.org/content/mimiciv/2.0/). On reasonable requests, the datasets used and/or analysed during the current investigation were available from the corresponding (first) author (YXH and ZL).
References
- 1.Klingele M, Baerens L.. Impact of renal replacement therapy on mortality in critically ill patients-the nephrologist’s view within an interdisciplinary intensive care team. J Clin Med. 2021;10(15):3379. doi: 10.3390/jcm10153379. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Pickkers P, Darmon M, Hoste E, et al. Acute kidney injury in the critically ill: an updated review on pathophysiology and management. Intensive Care Med. 2021;47(8):835–850. doi: 10.1007/s00134-021-06454-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Vijayan A, Abdel-Rahman EM, Liu KD, et al. Recovery after critical illness and acute kidney injury. Clin J Am Soc Nephrol. 2021;16(10):1601–1609. doi: 10.2215/CJN.19601220. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Schulman IH, Chan K, Der JS, et al. Readmission and mortality after hospitalization with acute kidney injury. Am J Kidney Dis. 2023;82(1):63–74.e1. e61. doi: 10.1053/j.ajkd.2022.12.008. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Khbouz B, Gu S, Pinto Coelho T, et al. Radiotherapy advances in renal disease-focus on renal ischemic preconditioning. Bioengineering (Basel). 2023;10(1):68. doi: 10.3390/bioengineering10010068. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Ying Y, Tang Q, Han D, et al. Nucleic acid nanotechnology for diagnostics and therapeutics in acute kidney injury. Int J Mol Sci. 2022;23(6):3093. doi: 10.3390/ijms23063093. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Vásquez Jiménez E, Anumudu SJ, Neyra JA.. Dose of continuous renal replacement therapy in critically ill patients: a bona fide quality indicator. Nephron. 2021;145(2):91–98. doi: 10.1159/000512846. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Teixeira JP, Griffin BR, Pal CA, et al. Critical illness myopathy and trajectory of recovery in acute kidney injury requiring continuous renal replacement therapy: a prospective observational trial protocol. BMJ Open. 2023;13(5):e072448. doi: 10.1136/bmjopen-2023-072448. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Mishra RC, Sinha S, Govil D, et al. Renal replacement therapy in adult intensive care unit: an ISCCM expert panel practice recommendation. Indian J Crit Care Med. 2022;26:S3–S6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Neyra JA, Chawla LS.. Acute kidney disease to chronic kidney disease. Crit Care Clin. 2021;37(2):453–474. doi: 10.1016/j.ccc.2020.11.013. [DOI] [PubMed] [Google Scholar]
- 11.Chaïbi K, Ehooman F, Pons B, et al. Long-term outcomes after severe acute kidney injury in critically ill patients: the SALTO study. Ann Intensive Care. 2023;13(1):18. doi: 10.1186/s13613-023-01108-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Huang C-Y, Güiza F, De Vlieger G, et al. Development and validation of clinical prediction models for acute kidney injury recovery at hospital discharge in critically ill adults. J Clin Monit Comput. 2023;37(1):113–125. doi: 10.1007/s10877-022-00865-7. [DOI] [PubMed] [Google Scholar]
- 13.Akazawa M, Hashimoto K.. Artificial intelligence in gynecologic cancers: current status and future challenges—a systematic review. Artif Intell Med. 2021;120:102164. doi: 10.1016/j.artmed.2021.102164. [DOI] [PubMed] [Google Scholar]
- 14.Lee BJ, Hsu CY, Parikh R, et al. Predicting renal recovery after dialysis-requiring acute kidney injury. Kidney Int Rep. 2019;4(4):571–581. doi: 10.1016/j.ekir.2019.01.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Hu P, Song L, Liang H, et al. Prospective model for predicting renal recovery in cardiac surgery patients with acute kidney injury requiring renal replacement therapy. Nephrology (Carlton). 2021;26(7):586–593. doi: 10.1111/nep.13878. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Johnson A, Bulgarelli L, Pollard T, et al. 2022., MIMIC-IV (version 2.0). PhysioNet. doi: 10.13026/7vcr-e114. [DOI] [Google Scholar]
- 17.Collins GS, Reitsma JB, Altman DG, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD). Ann Intern Med. 2015;162(10):735–736. doi: 10.7326/L15-5093-2. [DOI] [PubMed] [Google Scholar]
- 18.Duarte TTP, Magro MCS.. Recovery of renal function in clinical patients with acute kidney injury: impact on mortality. Life (Basel). 2022;12(6):852. doi: 10.3390/life12060852. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.De Rosa S, Samoni S, Ronco C.. Creatinine-based definitions: from baseline creatinine to serum creatinine adjustment in intensive care. Crit Care. 2016;20(1):69. doi: 10.1186/s13054-016-1218-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Ricci Z, Romagnoli S, Ronco C.. The 10 false beliefs in adult critical care nephrology. Intensive Care Med. 2018;44(8):1302–1305. doi: 10.1007/s00134-017-5011-x. [DOI] [PubMed] [Google Scholar]
- 21.Chawla LS, Bellomo R, Bihorac A, et al. Acute disease quality initiative, acute kidney disease and renal recovery: consensus report of the acute disease quality initiative (ADQI) 16 Workgroup. Nat Rev Nephrol. 2017;13(4):241–257. doi: 10.1038/nrneph.2017.2. [DOI] [PubMed] [Google Scholar]
- 22.Kim CS, Bae EH, Ma SK, et al. A Prospective observational study on the predictive value of serum cystatin C for successful weaning from continuous renal replacement therapy. Kidney Blood Press Res. 2018;43(3):872–881. doi: 10.1159/000490335. [DOI] [PubMed] [Google Scholar]
- 23.Rossiter A, La A, Koyner JL, et al. New biomarkers in acute kidney injury. Crit Rev Clin Lab Sci. 2024;61(1):23–44. doi: 10.1080/10408363.2023.2242481. [DOI] [PubMed] [Google Scholar]
- 24.Orieux A, Prezelin-Reydit M, Prevel R, et al. Clinical trajectories and impact of acute kidney disease after acute kidney injury in the intensive care unit: a 5-year single-centre cohort study. Nephrol Dial Transplant. 2023;38(1):167–176. doi: 10.1093/ndt/gfac054. [DOI] [PubMed] [Google Scholar]
- 25.See CY, Pan HC, Chen JY, et al. Improvement of composite kidney outcomes by AKI care bundles: a systematic review and meta-analysis. Crit Care. 2023;27(1):390. doi: 10.1186/s13054-023-04641-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.James MT, Bhatt M, Pannu N, et al. Long-term outcomes of acute kidney injury and strategies for improved care. Nat Rev Nephrol. 2020;16(4):193–205. doi: 10.1038/s41581-019-0247-z. [DOI] [PubMed] [Google Scholar]
- 27.Hansrivijit P, Yarlagadda K, Puthenpura MM, et al. A meta-analysis of clinical predictors for renal recovery and overall mortality in acute kidney injury requiring continuous renal replacement therapy. J Crit Care. 2020;60:13–22. doi: 10.1016/j.jcrc.2020.07.012. [DOI] [PubMed] [Google Scholar]
- 28.Thomsen J, Sprogøe U, Toft P.. Urine neutrophil gelatinase-associated lipocalin and urine output as predictors of the successful discontinuation of continuous renal replacement therapy in critically ill patients with acute kidney injury. BMC Nephrol. 2020;21(1):375. doi: 10.1186/s12882-020-02035-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Mishra RC, Sodhi K, Prakash KC, et al. ISCCM guidelines on acute kidney injury and renal replacement therapy. Indian J Crit Care Med. 2022;26(Suppl 2):S13–S42. doi: 10.5005/jp-journals-10071-24109. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Schricker S, Schanz M, Alscher MD[, et al. Metabolic acidosis: diagnosis and treatment]. Med Klin Intensivmed Notfmed. 2020;115(4):275–280. doi: 10.1007/s00063-019-0538-y. [DOI] [PubMed] [Google Scholar]
- 31.Phetrittikun R, Suvirat K, Horsiritham K, et al. Prediction of acid-base and potassium imbalances in intensive care patients using machine learning techniques. Diagnostics (Basel). 2023;13(6):1171. doi: 10.3390/diagnostics13061171. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Achanti A, Szerlip HM.. Acid-base disorders in the critically ill patient. Clin J Am Soc Nephrol. 2023;18(1):102–112. doi: 10.2215/CJN.04500422. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Magalhães PAdF, de Brito TS, Freire RS, et al. Metabolic acidosis aggravates experimental acute kidney injury. Life Sci. 2016;146:58–65. doi: 10.1016/j.lfs.2016.01.007. [DOI] [PubMed] [Google Scholar]
- 34.Bugarski M, Ghazi S, Polesel M, et al. Changes in NAD and lipid metabolism drive acidosis-induced acute kidney injury. J Am Soc Nephrol. 2021;32(2):342–356. doi: 10.1681/ASN.2020071003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Park S, Lee S, Kim Y, et al. Atrial fibrillation and kidney function: a bidirectional Mendelian randomization study. Eur Heart J. 2021;42(29):2816–2823. doi: 10.1093/eurheartj/ehab291. [DOI] [PubMed] [Google Scholar]
- 36.Ihara K, Ishigami J, Inoshita S.. Predictors of withdrawal from renal replacement therapy among patients with acute kidney injury requiring renal replacement therapy. Clin Exp Nephrol. 2019;23(6):814–824. doi: 10.1007/s10157-019-01711-y. [DOI] [PubMed] [Google Scholar]
- 37.Stads S, Kant KM, de Jong MFC, et al. Predictors of short-term successful discontinuation of continuous renal replacement therapy: results from a prospective multicentre study. BMC Nephrol. 2019;20(1):129. doi: 10.1186/s12882-019-1327-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Yu Y, Xu H, Jiang S, et al. Predictors for mortality and recovery in patients with acute renal injury receiving continuous renal replacement therapy. Int J Artif Organs. 2022;45(5):455–461. doi: 10.1177/03913988221086301. [DOI] [PubMed] [Google Scholar]
- 39.Zhong L, Xie B, Ji XW, et al. The association between albumin corrected anion gap and ICU mortality in acute kidney injury patients requiring continuous renal replacement therapy. Intern Emerg Med. 2022;17(8):2315–2322. doi: 10.1007/s11739-022-03093-8. [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 data are available in the physionet (https://physionet.org/content/mimiciv/2.0/). On reasonable requests, the datasets used and/or analysed during the current investigation were available from the corresponding (first) author (YXH and ZL).