Machine Learning-Based Prediction Model for ICU Mortality After Continuous Renal Replacement Therapy Initiation in Children

Abstract

BACKGROUND:

Continuous renal replacement therapy (CRRT) is the favored renal replacement therapy in critically ill patients. Predicting clinical outcomes for CRRT patients is difficult due to population heterogeneity, varying clinical practices, and limited sample sizes.

OBJECTIVE:

We aimed to predict survival to ICUs and hospital discharge in children and young adults receiving CRRT using machine learning (ML) techniques.

DERIVATION COHORT:

Patients less than 25 years of age receiving CRRT for acute kidney injury and/or volume overload from 2015 to 2021 (80%).

VALIDATION COHORT:

Internal validation occurred in a testing group of patients from the dataset (20%).

PREDICTION MODEL:

Retrospective international multicenter study utilizing an 80/20 training and testing cohort split, and logistic regression with L2 regularization (LR), decision tree, random forest (RF), gradient boosting machine, and support vector machine with linear kernel to predict ICU and hospital survival. Model performance was determined by the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) due to the imbalance in the dataset.

RESULTS:

Of the 933 patients included in this study, 538 (54%) were male with a median age of 8.97 years and interquartile range (1.81–15.0 yr). The ICU mortality was 35% and hospital mortality was 37%. The RF had the best performance for predicting ICU mortality (AUROC, 0.791 and AUPRC, 0.878) and LR for hospital mortality (AUROC, 0.777 and AUPRC, 0.859). The top two predictors of ICU survival were Pediatric Logistic Organ Dysfunction-2 score at CRRT initiation and admission diagnosis of respiratory failure.

CONCLUSIONS:

These are the first ML models to predict survival at ICU and hospital discharge in children and young adults receiving CRRT. RF outperformed other models for predicting ICU mortality. Future studies should expand the input variables, conduct a more sophisticated feature selection, and use deep learning algorithms to generate more precise models.

KEY POINTS.

Question: What specific question does the article address? State the hypothesis, goal, or purpose of the study.

This study aims to use machine learning to predict ICU and hospital survival in children receiving continuous renal replacement therapy (CRRT).

Findings: What is the key result? Mention the design (e.g., clinical trial, cohort study, case-control study, meta-analysis, predictive model). State the primary outcome(s) or finding(s) only. Omit secondary outcomes. Report basic numbers only but state whether results are statistically significant or not significant; do not include results of statistical tests or measures of variance.

Two predictive models were developed using machine learning techniques and the Worldwide Exploration of Renal Replacement Outcomes Collaborative in Kidney Diseases (WE-ROCK) database. A random forest model predicted ICU survival (area under the receiver operating characteristic curve [AUROC], 0.791; area under the precision-recall curve [AUPRC] 0.878), while a logistic regression model predicted hospital survival (AUROC, 0.777; AUPRC, 0.859). The Pediatric Logistic Organ Dysfunction-2 score at CRRT initiation was the most important predictor in both models.

Meaning: What are the key conclusions and implications based on the primary finding(s)?

Machine learning models may help predict ICU and hospital discharge survival in children and young adults undergoing CRRT.

Acute kidney injury (AKI) and volume overload (VO) are common in children and young adults admitted to ICUs (1), and those with the most severe forms require renal replacement therapy (RRT) (1–3). Continuous RRT (CRRT) is the preferred therapeutic modality for critically ill patients with hemodynamic instability. The mortality associated with CRRT ranges from 36% to 60% depending on the population studied (4–7). Given this high mortality rate, it is important for clinicians to identify patients at a higher risk of poor outcomes as it may allow them to provide targeted interventions and inform families at the bedside. This can be challenging, given the heterogeneous nature of the population and disease processes (8–10). Previous studies have identified that younger age, presence of sepsis, and certain underlying conditions such as pulmonary disease or hematopoietic stem cell transplantation are associated with higher mortality (11–13). Recent reports from the Worldwide Exploration of Renal Replacement Outcomes Collaborative in Kidney Diseases (WE-ROCK), which included 980 patients from 32 centers in seven countries showed an ICU mortality of 36%, with younger age, presence of comorbidity (particularly cardiac, hematopoietic stem cell transplants, or oncologic comorbidities), and sepsis associated with worse outcomes (14).

In recent years, machine learning (ML) algorithms have been used to predict mortality after CRRT initiation in critically ill adults as they can capture complex nonlinear interactions from large datasets (15, 16). However, there is no study evaluating this in children and young adults, likely as most studies examining CRRT risk factors are single center and have small sample size. We aimed to leverage the multicenter, multinational WE-ROCK study to build ML models to predict ICU survival after CRRT initiation in this population and compare the performance of different ML algorithms (17).

METHODS

Study Population

The initial study from the WE-ROCK group included 980 patients (birth to 25 yr old) from 32 institutions in seven countries (United States, Canada, United Kingdom, Italy, Spain, Austria, and Australia) who received CRRT for AKI or VO in the pediatric, neonatal, or cardiac ICU from January 2015 to December 2021 (14, 17, 18). Sixty-four patients were included from 2015 to 2018, while 916 patients were included 2018–2021. Given the small number of patients within the dataset who received CRRT between 2015 and 2018 and no significant change in CRRT practices these patients were in our final cohort. The exclusion criteria were as follows: 1) end-stage kidney disease (ESKD) defined as dialysis dependence; 2) infants with severe congenital anomalies of the kidney and urinary tract likely to result in ESKD; 3) patients receiving CRRT for a non-AKI/VO indication (i.e., ingestion or inborn errors of metabolism); 4) patients on concurrent extracorporeal membrane oxygenation; 5) patients receiving peritoneal dialysis during the same admission before CRRT initiation; and 6) patients who underwent CRRT via Carpediem (Mozarc Medical, Minneapolis, MN) were excluded due to the presence of an existing registry focusing on the device. The Institutional Review Board (IRB) at Cincinnati Children’s Hospital Medical Center granted approval for this collaborative study. Furthermore, each participating site’s IRB approved this multicenter investigation (Supplemental Table 1, http://links.lww.com/CCX/B438), and all procedures associated with this research were conducted in accordance with the ethical standards delineated in the Helsinki Declaration of 1975.

Study Variables

Demographic data, including sex, self-reported race, and ethnicity (from the electronic health record), age at ICU admission, and CRRT initiation were collected for all patients. Race and ethnicity were included as variables because of the potential for differences in outcomes. Data at CRRT initiation included baseline kidney function (creatinine and estimated glomerular filtration rate), presence of sepsis (defined as the presence of an infection and systemic inflammatory response syndrome within 24 hours of ICU admission), Vasoactive-Inotropic Score (VIS) (19), and Pediatric Logistic Organ Dysfunction-2 (PELOD-2) score 24 hours before CRRT initiation (20), cumulative fluid balance at CRRT initiation and daily for the first 7 days of CRRT, and loop diuretic use. Details of CRRT prescription, including device, modality, filter, dose, fluid type, and anticoagulation, were also collected. All features (excluding cumulative fluid balance) were collected before or at CRRT initiation and used within our analysis. Last, data were collected on outcomes of interest: survival to ICU and hospital discharge.

Data Preprocessing and Feature Selection

The data were analyzed using packages NumPy (21), pandas (22), and SciPy in Python (23), and non-normal distributions of the data were addressed by expressing categorical and continuous variables as numbers (proportions) and medians (quartiles 1–3), respectively. The Mann-Whitney U test was used to compare medians of continuous variables, while the Fisher exact test was employed for independence of categorical variables between the death and survival groups. Feature selection was conducted first through discussions by the research team, resulting in the inclusion of 61 of the 119 variables. Then of those variables selected, those with the p value of less than 0.2 in the univariate analysis were utilized for model development (Supplement Table 2, http://links.lww.com/CCX/B438).

Machine Learning Models

The ML models were trained and tested in a supervised manner using the binary outcomes in the dataset. We chose to focus on the utilization of classical ML algorithms, which are known to have better interpretability compared with other deep learning methodologies. The classical ML algorithms that were used in the training process, including logistic regression with L2 regularization (LR), decision trees (DTs), random forests (RFs), gradient boosting machines (GBMs), and support vector machines (SVMs) with linear kernels. Specifically, LR identifies relationships between independent variables and a dependent variable (binary outcome) and classifies data into one of the two outcome groups. DT is a hierarchical model that generates decision rules by minimizing uncertainty. RF uses random subsets of the data to create a multitude of DTs and combines them to produce a probability. GBM is an iterative algorithm that creates weight-based DTs to minimize gradient errors. SVM creates a hyperplane dividing the data into groups while optimizing the separation between the datapoints in each group (24). To assist with interpretation of ML methodology we have included a table of frequently used vocabulary and definitions (Supplemental Table 3, http://links.lww.com/CCX/B438).

The dataset was randomly divided into training (80%) and testing (20%) sets and baseline demographic data between the two cohorts was compared using Mann-Whitney U test (continuous variables) and Fisher exact test (binary/categorical variables) (Supplemental Table 2, http://links.lww.com/CCX/B438) for feature selection. The models were further trained using ten-fold nested cross-validation for both hyperparameters and model selection, and the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), F1 score (a measure of predictive performance of the test), accuracy, balanced accuracy, precision, sensitivity, and specificity were calculated for each model. The 95% CI of AUROC is computed with 2000 stratified bootstrap replicates. The best-performing model based on the training set was applied to the testing set to obtain the final model performance. Last, SHapley Additive exPlanations (SHAP) plots were used to determine the ranking of the top 20 predictors for the best performing model on the testing set.

RESULTS

Cohort Characteristics

Details of the cohort have been previously described (14). The original cohort included 980 patients and a total of 47 patients were excluded due to missing or incorrect data, resulting in 933 patients included in this study, of whom 508 (54%) were male. The median age at admission was 8.97 years (1.81–15.0 yr), with 89 patients (9.5%) between 18 and 25 years old. The median weight 27.5 kg (11.9–55.0 kg). The most common reason for ICU admission was shock/infection/major trauma in 350 patients (38%), and sepsis was present in 431 patients (46%) at the time of ICU admission. Most of the cohort (80%) had at least one comorbidity with the most common being oncological, present in 215 patients (23%). The median percent VO at CRRT initiation was 7.5% (interquartile range [IQR], 2.39–18.08%), and the median illness severity, measured by the PELOD-2 score within 24 hours of CRRT initiation, was 7 (IQR, 4–9). A total of 603 patients (65%) survived to ICU discharge. Multiple significant differences were identified between survivors and nonsurvivors (Table 1).

TABLE 1.

Key Demographic and Clinical Characteristics of Survivors and Nonsurvivors to ICU Discharge

Patient Characteristics	Overall, n = 933	ICU Survival		p b
		Survivor to ICU Discharge, n = 603a	Death Before ICU Discharge, n = 3301
Age (yr)	8.97 (1.81–15.0)	9.0 (2.06–14.93)	8.78 (1.21–15.28)	0.54
Sex (male)	508 (54.45)	320 (53.07)	188 (56.97)	0.28
Sepsis at ICU admission	431 (46.2)	258 (42.79)	173 (52.42)	0.01
Admission diagnosis
Shock/infection/major trauma	350 (37.51)	236 (39.14)	114 (34.55)	0.19
Respiratory failure	185 (19.83)	84 (13.93)	101 (30.61)	< 0.01
Post-surgical/minor trauma	44 (4.72)	31 (5.14)	13 (3.94)	0.52
CNS dysfunction	39 (4.18)	22 (3.65)	17 (5.15)	0.31
Pain/sedation management	8 (0.86)	6 (1.0)	2 (0.61)	0.72
Congenital heart disease	30 (3.22)	13 (2.16)	17 (5.15)	0.02
Post-surgical congenital heart disease	46 (4.93)	37 (6.14)	9 (2.73)	0.03
Heart failure and/or cardiomyopathy	37 (3.97)	22 (3.65)	15 (4.55)	0.49
Other	194 (20.79)	152 (25.21)	42 (12.73)	< 0.01
Comorbidities
None	186 (19.94)	146 (24.21)	40 (12.12)	< 0.01
Respiratory	121 (12.97)	74 (12.27)	47 (14.24)	0.42
Cardiac	183 (19.61)	105 (17.41)	78 (23.64)	0.03
Neurologic/neuromuscular	126 (13.5)	81 (13.43)	45 (13.64)	0.92
Kidney/urologic	88 (9.43)	66 (10.95)	22 (6.67)	0.04
Hematologic	118 (12.65)	73 (12.11)	45 (13.64)	0.54
Oncologic	215 (23.04)	119 (19.73)	96 (29.09)	< 0.01
Immunologic	148 (15.86)	69 (11.44)	79 (23.94)	< 0.01
Gastrointestinal	174 (18.65)	122 (20.23)	52 (15.76)	0.10
Endocrinologic	61 (6.54)	43 (7.13)	18 (5.45)	0.41
ICU admission weight (kg)	27.5 (11.9–55.0)	29.5 (12.85–57.55)	26.3 (10.0–52.0)	0.04
Body surface area	0.98 (0.51–1.55)	1.01 (0.55–1.59)	0.96 (0.45–1.51)	0.03
Baseline serum creatinine	0.44 (0.28–0.66)	0.47 (0.31–0.67)	0.40 (0.21–0.62)	< 0.01
Platelet count at ICU admission	97.5 (38.75–222.0)	113.0 (42.0–241.75)	84.0 (30.0–174.0)	< 0.01
Serum creatinine immediately before CRRT initiation	1.74 (0.92–3.3)	2.1 (0.99–3.7)	1.41 (0.82–2.49)	< 0.01
Sepsis at the time of CRRT initiation	400 (42.87)	229 (37.98)	171 (51.82)	< 0.01
Pre-CRRT Vasoactive-Inotropic Score	5.0 (0.0–20.0)	2.0 (0.0–13.0)	10.0 (0.0–25.46)	< 0.01
Lactate	2.2 (1.3–5.2)	2.0 (1.2–3.9)	3.3 (1.7–7.3)	< 0.01
Receiving invasive mechanical ventilation	585 (62.7)	352 (58.37)	233 (70.61)	< 0.01
WBC	9.5 (3.5–17.0)	9.85 (4.5–17.5)	8.55 (2.04–15.88)	< 0.01
Pediatric Logistic Organ Dysfunction-2 pre-CRRT	7.0 (4.0–9.0)	6.0 (4.0–8.0)	8.0 (6.0–10.75)	< 0.01
% Volume overload (ICU admit to CRRT initiation)	7.47 (2.39–18.08)	6.97 (2.01–16.63)	8.92 (3.54–22.08)	< 0.01
Total urine output in 24 hr before CRRT initiation (mL)	291.0 (64.0–864.0)	300.0 (69.0–880.0)	283.5 (61.75–802.5)	0.56

CKRT = continuous kidney replacement therapy.

^aStatistics presented: n (%) or median (interquartile range).

^bStatistical tests performed: Fisher exact test of independence between features and outcomes; Mann-Whitney U test for median differences between survival and death cohorts.

Bolded values identify significant variables.

Model Development

A total of 746 patients, constituting 80% of the entire patient pool, were included in the training dataset, whereas the remaining 187 patients (20%) were assigned to the test cohort. Within the training and testing cohorts for our ICU survival model, we observed a statistically significant difference in patients with neurologic/neuromuscular comorbid conditions (p = 0.036). In the cohorts used for predicting hospital survival, we identified significant differences in age (p = 0.046), reason for admission being CNS dysfunction (p = 0.034) and pain/sedation (p = 0.034), hematological comorbidity (p = 0.040), WBC count (p = 0.006), and urine output before CRRT initiation (p = 0.049) (Supplemental Table 4, http://links.lww.com/CCX/B438). The performance of each ML model for ICU and hospital survival is presented in Table 2.

TABLE 2.

Model Performance for Survival to ICU Discharge and Hospital Discharge

Model	Area Under the Curve	Area Under the Precision-Recall Curve	Accuracy	Balanced Accuracy	Precision	Recall	F1 Score
ICU survival
Gradient boost	0.754	0.834	0.738	0.694	0.773	0.843	0.806
Random forest	0.791	0.878	0.733	0.649	0.729	0.934	0.819
Decision tree	0.660	0.760	0.658	0.622	0.732	0.744	0.738
Logistic regression	0.755	0.853	0.722	0.654	0.738	0.884	0.805
Linear SVM	0.658	0.851	0.727	0.658	0.740	0.893	0.809
Hospital survival
Gradient boost	0.768	0.825	0.701	0.660	0.733	0.821	0.774
Random forest	0.768	0.847	0.695	0.645	0.717	0.846	0.776
Decision tree	0.595	0.775	0.572	0.546	0.661	0.65	0.655
Logistic regression	0.777	0.859	0.722	0.680	0.744	0.846	0.792
Linear SVM	0.677	0.843	0.722	0.677	0.741	0.855	0.794

SVM = support vector machine.

Bolded rows indicate the best preforming machine learning model for each outcome.

ICU Survival and Top Features

The RF exhibited an area under the curve (AUC) of 0.791 (95% CI, 0.725–0.851), an AUPRC of 0.878, and an F1 score of 0.819, representing the highest AUC among five ML models. The performances of GBM and LR were similar, with AUCs of 0.754 and 0.755, respectively (Table 2 and Fig. 1). The SHAP values showed that the PELOD-2 score (lower) at CRRT initiation and absence of respiratory failure were the top two significant predictors of ICU survival, followed closely by the serum creatinine at CRRT initiation (higher), norepinephrine dose (lower), and VIS at CRRT start (lower; Fig. 2).

Figure 1.

Receiver operating characteristic (ROC) curve and precision-recall curve for ICU survival. AUC = area under the curve, CRRT = continuous renal replacement therapy, SVM = support vector machine.

Figure 2.

SHapley Additive exPlanations (SHAP) plot showing the 20 most important variables and the contribution of each variable to our model’s prediction if ICU survival. BSA = body surface area, CRRT = continuous renal replacement therapy, PELOD-2 = Pediatric Logistic Organ Dysfunction-2, VIS = Vasoactive-Inotropic Score.

Hospital Survival and Top Features

The LR model was the best predictive model for hospital survival, with an AUC of 0.777 (0.706–0.840), an AUPRC of 0.859, and an F1 score of 0.792 (Table 2 and Fig. 3).

Figure 3.

Receiver operating characteristic (ROC) curve and precision-recall curve for hospital survival. AUC = area under the curve, CRRT = continuous renal replacement therapy, SVM = support vector machine.

The SHAP values showed that the lower PELOD-2 score at the CRRT initiation was the most important predictor of hospital survival. This was closely followed by serum creatinine at CRRT initiation (higher), admission diagnosis of shock/infection/major trauma, admission platelet count (higher), and absence of cardiac comorbidities (Fig. 4).

Figure 4.

SHapley Additive exPlanations (SHAP) plot showing the 20 most important variables and the contribution of each variable to our model’s prediction of hospital survival. CRRT = continuous renal replacement therapy, PELOD-2 = Pediatric Logistic Organ Dysfunction-2, VIS = Vasoactive-Inotropic Score.

In comparison to the ICU survival model, hospital survival model had lower performance across all measures, including a lower AUROC (0.791 vs. 0.777), AUPRC (0.878 vs. 0.859), and F1 score (0.819 vs. 0.792).

DISCUSSION

This investigation employed the WE-ROCK dataset and classical ML algorithms to predict outcomes for ICU and hospital survival. Among these techniques, RF and LR emerged as the top performers for ICU and hospital survival, respectively. By employing SHAP plots to explain the models, we identified 14 common factors that contributed to their success. The high-performing models achieved sensitivity of 0.710 and 0.701 and specificities of 0.712 and 0.700 for predicting ICU and hospital survival, respectively. Furthermore, their F1 scores, which offer a more comprehensive measure that balances true and false positives and is particularly useful in cases with imbalanced datasets, were 0.761 for predicting ICU survival and 0.746 for predicting hospital survival. Given our dataset’s 65% survival rate and 35% mortality rate, we argue that the F1 score is the most appropriate metric to use. Last, to our knowledge, this is the first ML model evaluating ICU and hospital survival in critically ill children and young adults receiving CRRT.

Utilizing SHAP plots to identify important features, we were able to divide these clinical variables into four major groups. The first group included variables associated with comorbid and admission conditions (respiratory indication for ICU admission, both gastrointestinal and immunological comorbidities, and a lack of a comorbid condition [protective]) (25). Second were variables associated with underlying renal dysfunction (baseline serum creatinine, serum creatinine at CRRT initiation, and urine output and VO at CRRT start). Next included variables associated with extrarenal organ dysfunction and illness severity (Pediatric Risk of Mortality [PRISM], PELOD-2 score at CRRT initiation, VIS, norepinephrine dose, WBC count, and platelet count) (26). And the final group had age and anthropometric variables (body mass index, body surface area, and weight). These clinically meaningful patient characteristics may be used in future ML models to improve model performance, but further external validation is required.

When comparing the ICU survival model to previously published literature in adults requiring CRRT, our models had better performance (16). Kang et al (16) created six different predictive models, all of which outperformed the existing illness severity scores (Acute Physiology and Chronic Health Evaluation II, Sequential Organ Failure Assessment, and mortality scoring system for acute kidney injury with CRRT). There are some similarities in the key features between the models developed by Kang et al (16) and in this article, with comorbid conditions, underlying renal dysfunction, respiratory failure, and hematological measures playing an important role in model prediction. Our hospital survival model performed slightly worse compared with the published literature. This has two likely explanations. First, most deaths occur before ICU discharge, and lapse of significant time between the end of WE-ROCK data collection (day 7 of CRRT) and hospital discharge. Second, there were some differences noted between the training and testing cohorts, which could have impacted the final model performance. Children and young adults with severe chronic illness tend to have prolonged hospitalization after their ICU stay to allow for rehabilitation unlike adults where they may be discharged to skilled nursing or rehabilitation facilities for the same. Given this, we theorize that more granular clinical data throughout CRRT and hospitalization would be able to better predict ICU and hospital survival (Supplemental Table 5, http://links.lww.com/CCX/B438).

In the realm of pediatric critical care nephrology, ML models have aimed to predict AKI and the need for CRRT (27–30). However, the number of these models is small compared with what has been developed for critically ill adults. The 27th Acute Disease Quality Initiative focusing on digital health and AKI discusses how digital health solutions can impact not only risk stratification and recognition, allowing for a more personalized approach to renal recovery (31), but also be used to design and validate models for risk stratification, and subsequently improve patient-centered outcomes (32). Employing these guidelines, we used the WE-ROCK dataset to better understand factors associated with survival and risk stratification of this vulnerable population. Our findings indicate feasibility in the application of ML in this population. In previous publications from the WE-ROCK registry, we have reported significant heterogeneity in the practice and prescription of CRRT in children (14). A potential initial step in recognizing modifiable factors in this vulnerable population could be through the application of unsupervised ML algorithms to identify subphenotypes in patients undergoing CRRT. This may improve prognostication as well as enable researchers to identify novel modifiable factors that could be implemented in clinical practice (33).

Artificial intelligence (AI) in pediatric critical care nephrology presents significant potential, particularly because ML models can address the dynamic and complex nature of critically ill children. Previously published literature has demonstrated the ability to predict AKI in a near-continuous manner, closely approximating real-time clinical conditions (28, 34). The incorporation of all electronic health record data from the entire duration of CRRT into future models has the potential to improve prediction accuracy and provide more dynamic predictions. We hope that AI may facilitate a more personalized approach to the prescription of CRRT, timely interventions, and allow us to improve outcomes in this vurnerable population.

The present study has several limitations. The WE-ROCK study collected data for the first 7 days of CRRT only and had limited laboratory data. The data collection form was created using the available knowledge and lacked granularity, which may lead to model overfitting and inability to identify other modifiable factors associated with survival. However, the features identified in this study have high clinical importance. Future studies using more granular data throughout the CRRT course that can be automated in the electronic health record may enable continuous model updating with relevant variables to adjust risk and predict all outcomes. In addition, there is known terminological and semiotic heterogeneity that could have led to minute inconsistencies in the data collected from the 30+ sites. Additionally, at the time of this analysis there were no countries from low-/lower-middle-income countries included within the dataset. High illness severity scores like PRISM-III at ICU admission are associated with mortality in critically ill children. Due to the large number of patients with missing PRISM-III score in our dataset, we could not include it within our final model, and instead included PELOD-2 within 24 hours of CRRT initiation. Future studies should use previously validated illness severity scores when developing predictive models for children receiving CRRT. Most features that were included in our final model were extracted before or at CRRT initiation and therefore limit the timeframe at which clinicians could use these predictive models. The differences between the training and testing cohorts could have diminished model performance. While the WE-ROCK study has some data on race and ethnicity, which may impact outcomes through various social determinants of health, it was not used for the prediction model. This was due to the variability in defining race and ethnicity across the participating centers and many international centers do not regularly use these social constructs in clinical research. Furthermore, the impact of development as a biological variable (DABV) may also impact outcomes, as well as considerations of sex differences that may or may not exist based on pubertal status (35). Future studies should use social determinants of health and DABV to assess its impact on survival and long-term morbidity in this high-risk population.

CONCLUSIONS

We developed and tested an ML model to predict both survival to ICU and hospital discharge in a multicenter cohort of children and young adults receiving CRRT. RF for ICU survival performed better than the other ML models (GBM, LR, DT, and SVM). We identified multiple clinical variables associated with illness severity and comorbidities that could assist clinicians in the prognostication of patients in this heterogeneous population. This is the first step in utilizing AI or ML to better understand the factors impacting outcomes in children receiving CRRT. Future studies should focus on two major aspects: 1) the utilization of more sophisticated ML techniques (deep learning/neural networks) to improve model performance and 2) use techniques such as federated learning or reinforcement learning to externally validate these findings.

ACKNOWLEDGMENTS

We thank T. Christine E. Alvarez, MHI, RN (Cincinnati Children’s Hospital Medical Center, Cincinnati, OH); Elizabeth Bixler, BS (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Erica Blender Brown, MA, CRA (Medical University of South Carolina, Charleston, SC); Cheryl L. Brown, BS (Cincinnati Children’s Hospital Medical Center, Cincinnati, OH); Ambra Burrell, BA (Nationwide Children’s Hospital, Columbus, OH); Anwesh Dash, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); Jennifer L. Ehrlich, RN, MHA (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); Simrandeep Farma, HBSc (Hospital for Sick Children, Toronto, ON, Canada); Kim Gahring, RN, BSN, CCRN (Children’s Hospital Colorado, Aurora, CO); Barbara Gales, RN (Mattel Children Hospital at UCLA, Los Angeles, CA); Madison R. Hilgenkamp (University of Nebraska Medical Center, Children’s Hospital & Medical Center, Omaha, NE); Sonal Jain, MS (Seattle Children’s Hospital, Seattle, WA); Kate Kanwar, BA, MS (Nationwide Children’s Hospital, Columbus, OH); Jennifer Lusk, BSN, RN, CCRN (Children’s Hospital Colorado, Aurora, CO); Christopher J. Meyer, BA, AA (Cincinnati Children’s Hospital Medical Center, Cincinnati, OH); Katherine Plomaritas, BSN, RN (University of Michigan, C.S. Mott Children’s Hospital, Ann Arbor, MI); Joshua Porter, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); Jessica Potts, BSN, RN (Children’s of Alabama/University of Alabama at Birmingham, Birmingham, AL); Alyssa Serratore, BNurs, GDipNP(PIC), RN, MsC (Royal Children’s Hospital, Melbourne, VIC, Australia); Elizabeth Schneider, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); Vidushi Sinha, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); P. J. Strack, RN, BSN, CCRN (Children’s Mercy Hospital, Kansas City, MO); Sue Taylor, RN (King’s College Hospital, London, United Kingdom); Katherine Twombley, MD (Medical University of South Carolina, Charleston, SC); Brynna Van Wyk, MSN, ARNP, CPNP (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); Samantha Wallace, MS (Indiana University School of Medicine, Riley Hospital for Children, Indianapolis, IN); Janet Wang, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); Megan Woods, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN); Marcia Zinger, RN (Cohen Children’s Medical Center, New Hyde Park, NY); Alison Zong, BS (University of Tennessee Health Science Center College of Medicine, Memphis, TN).

APPENDIX

The following individuals served as collaborators and investigators for the Worldwide Exploration of Renal Replacement Outcomes Collaborative in Kidney Diseases (WE-ROCK) studies. They collaborated in protocol development and review, data analysis, and participated in drafting or review of the article, and their names should be citable by PubMed. Emily Ahern, CPNP, DNP (Children’s Hospital Colorado, University of Colorado School of Medicine, Aurora, CO); Ayse Akcan Arikan, MD (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Issa Alhamoud, MD (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); Rashid Alobaidi, MD, MSc (Univeristy of Alberta, Edmonton, AB, Canada); Pilar Anton-Martin, MD, PhD (Le Bonheur Children’s Hospital, Memphis, TN); Shanthi S. Balani, MD (University of Minnesota, Minneapolis, MN); Matthew Barhight, MD, MS (Ann and Robert H. Lurie Children’s Hospital of Chicago, Chicago, IL); Abby Basalely, MD, MS (Cohen Children’s Medical Center, Zucker School of Medicine, New Hyde Park, NY); Amee M. Bigelow, MD, MS (Nationwide Children’s Hospital, The Ohio State University College of Medicine, Columbus, OH); Gabriella Bottari, MD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Andrea Cappoli, MD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Abhishek Chakraborty, MD (Le Bonheur Children’s Hospital, Memphis, TN); Eileen A. Ciccia, MD (Washington University School of Medicine, St. Louis Children’s Hospital, St. Louis, MO)^; Michaela Collins, BA (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH)^; Denise Colosimo, MD (Meyer Children’s Hospital, IRCCS, Florence, Italy); Gerard Cortina, MD (Medical University of Innsbruck, Innsbruck, Austria); Mihaela A. Damian, MD, MPH (Stanford University School of Medicine, Palo Alto, CA); Sara De la Mata Navazo, MD (Gregorio Marañón University Hospital, School of Medicine, Madrid, Spain); Gabrielle DeAbreu, MD (Cohen Children’s Medical Center, Zucker School of Medicine, New Hyde Park, NY); Akash Deep, MD (King’s College Hospital, London, United Kingdom); Kathy L. Ding, BS (University of Colorado, School of Medicine, Aurora, CO); Kristin J. Dolan, MD (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Lama Elbahlawan, MD (St Jude Children’s Hospital, Memphis TN); Sarah N. Fernandez Lafever, MD, PhD (Gregorio Marañón University Hospital, School of Medicine, Madrid, Spain); Dana Y. Fuhrman, DO, MS (University of Pittsburgh Medical Center Children’s Hospital of Pittsburgh, Pittsburgh, PA); Ben Gelbart, MBBS (Royal Children’s Hospital, University of Melbourne, Murdoch Children’s Research Institute, Melbourne, VIC, Australia); Katja M. Gist, DO, MSc (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH); Stephen M. Gorga, MD, MSc (University of Michigan Medical School, C.S. Mott Children’s Hospital, Ann Arbor, MI); Francesco Guzzi, MD (Santo Stefano Hospital, Prato, Italy); Isabella Guzzo, MD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Taiki Haga, MD (Osaka City General Hospital, Osaka, Japan); Elizabeth Harvey, MD (Hospital for Sick Children, Toronto, ON, Canada); Denise C. Hasson, MD (NYU Langone Health, Hassenfeld Children’s Hospital, New York, NY); Taylor Hill-Horowitz, BS (Cohen Children’s Medical Center, Zucker School of Medicine, New Hyde Park, NY); Haleigh Inthavong, BS, MS (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Catherine Joseph, MD (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Ahmad Kaddourah, MD, MS (Sidra Medicine and Weil Cornel Medicine, Qatar, Doha, Qatar); Aadil Kakajiwala, MD, MSCI (Children’s National Hospital, Washington, DC); Aaron D. Kessel, MD, MS (Cohen Children’s Medical Center, Zucker School of Medicine, New Hyde Park, NY); Sarah Korn, DO (Westchester Medical Center, Westchester, NY); Kelli A. Krallman, BSN, MS (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH); David M. Kwiatkowski, MD, Msc (Lucile Packard Children’s Hospital, Palo Alto, CA); Jasmine Lee, MSc (Hospital for Sick Children, Toronto, ON, Canada); Laurance Lequier, MD (Univeristy of Alberta, Edmonton, AB, Canada); Tina Madani Kia, BS (Univeristy of Alberta, Edmonton, AB, Canada); Kenneth E. Mah, MD, MS (Stanford University School of Medicine, Palo Alto, CA); Eleonora Marinari, MD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Susan D. Martin, MD (Golisano Children’s Hospital at University of Rochester Medical Center, Rochester, NY); Shina Menon, MD (Stanford University School of Medicine, Palo Alto, CA; Lucile Packard Children’s Hospital, Palo Alto, CA; Seattle Children’s Hospital, University of Washington, Seattle, WA); Tahagod H. Mohamed, MD (Nationwide Children’s Hospital, The Ohio State University College of Medicine, Columbus, OH); Catherine Morgan, MD, MSc (Univeristy of Alberta, Edmonton, AB, Canada); Theresa A. Mottes, APRN (Ann and Robert H. Lurie Children’s Hospital of Chicago, Chicago, IL); Melissa A. Muff-Luett, MD (Children’s Hospital & Medical Center, University of Nebraska Medical Center, Omaha, NE); Siva Namachivayam, MBBS (Royal Children’s Hospital, University of Melbourne, Murdoch Children’s Research Institute, Melbourne, VIC, Australia); Tara M. Neumayr, MD (Washington University School of Medicine, St. Louis Children’s Hospital, St. Louis, MO); Jennifer Nhan, MD, MS (Children’s National Hospital, Washington, DC); Abigail O’Rourke, MD (Cohen Children’s Medical Center, Zucker School of Medicine, New Hyde Park, NY); Nicholas J. Ollberding, PhD (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH); Matthew G. Pinto, MD (Maria Fareri Children’s Hospital at Westchester Medical Center, Valhalla, NY); Dua Qutob, MD (Sidra Medicine and Weil Cornel Medicine, Doha, Qatar); Valeria Raggi, MD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Stephanie Reynaud, MD (Dalhousie University, Halifax, NS, Canada); Zaccaria Ricci, MD (Meyer Children’s Hospital, IRCCS, Florence, Italy); Zachary A. Rumlow, DO (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); María J. Santiago Lozano, MD, PhD (Gregorio Marañón University Hospital, School of Medicine, Madrid, Spain); Emily See, MBBS (Royal Children’s Hospital, University of Melbourne, Murdoch Children’s Research Institute, Melbourne, VIC, Australia); David T. Selewski, MD, MSCR (Medical University of South Carolina, Charleston, SC); Carmela Serpe, MSc, PhD (Bambino Gesù Children Hospital, IRCCS, Rome, Italy); Alyssa Serratore, RN, MsC (Royal Children’s Hospital, University of Melbourne, Murdoch Children’s Research Institute, Melbourne, VIC, Australia); Ananya Shah, BS (University of Colorado, School of Medicine, Aurora, CO); Weiwen V. Shih, MD (Children’s Hospital Colorado, University of Colorado School of Medicine, Aurora, CO; University of Colorado, School of Medicine, Aurora, CO); H. Stella Shin, MD (Children’s Healthcare of Atlanta, Emory University, Atlanta, GA); Cara L. Slagle, MD (Indiana University School of Medicine, Riley Hospital for Children, Indianapolis, IN); Sonia Solomon, DO (Maria Fareri Children’s Hospital at Westchester Medical Center, Valhalla, NY); Danielle E. Soranno, MD (Indiana University School of Medicine, Riley Hospital for Children, Indianapolis, IN); Rachana Srivastava, MD (Mattel Children’s Hospital at UCLA, Los Angeles, CA); Natalja L. Stanski, MD (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH); Michelle C. Starr, MD, MPH (Indiana University School of Medicine, Riley Hospital for Children, Indianapolis, IN); Erin K. Stenson, MD (Children’s Hospital Colorado, University of Colorado School of Medicine, Aurora, CO; University of Colorado, School of Medicine, Aurora, CO); Amy E. Strong, MD, MSCE (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); Susan A. Taylor, MSc (King’s College Hospital, London, United Kingdom); Sameer V. Thadani, MD (Baylor College of Medicine, Texas Children’s Hospital, Houston, TX); Amanda M. Uber, DO (Children’s Hospital & Medical Center, University of Nebraska Medical Center, Omaha, NE); Brynna Van Wyk, ARNP, MSN (University of Iowa Stead Family Children’s Hospital, Carver College of Medicine, Iowa City, IA); Tennille N. Webb, MD, MSPH (Children’s of Alabama/University of Alabama at Birmingham, Birmingham, AL); Huaiyu Zang, PhD (Cincinnati Children’s Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH); Emily E. Zangla, DO (University of Minnesota, Minneapolis, MN); Michael Zappitelli, MD, MSc (Hospital for Sick Children, Toronto, ON, Canada).