This is a retrospective cohort study of children (<18 years of age) who received CRRT and were admitted to any of the intensive care units at Texas Children hospital from 9/2016-10/2018. Children who received therapy with extracorporeal membranous oxygenation, ventricular assist devices, and/or pacemakers, or those without an arterial line were excluded. This study was approved by the Baylor College of Medicine Institutional Review Board with a waiver of consent (H-32217).

Hemodynamic measurements were extracted every two seconds for 60 min before and after CRRT connection (initiation and subsequent restart) using archived high-resolution physiological data (SickBay Medical Informatics Corporation, Houston, TX). Non-physiological data were identified and eliminated using a local-scale median absolute deviation from the local median over a 60-value interval (Supplementary Fig. S1). Subsequently, the median value was calculated for one-minute epochs, and this minute-by-minute hemodynamic data were used for this analysis.

Per institutional protocol, continuous venovenous hemodiafiltration (CVVHDF) was prescribed to all patients using HF1000 membranes. Regional citrate anticoagulation was utilized with a post-filter calcium infusion. Circuits were primed using either packed red blood cells diluted to 35% hematocrit (PRBCs) or normal saline, depending on the patient's weight. When a patient's circuit was primed with PRBCs, rapid restart was attempted. In this procedure, the prime blood from the old circuit was directly infused into the new circuit before it is reconnected to the patient. This approach is designed to minimize overall blood product use and exposure.

We collected data on patient age at ICU admission, weight, height, body surface area, primary reasons for ICU admission, comorbid conditions, baseline renal function (defined as the lowest measurements obtained 90 days prior to admission), presence of CKD (baseline glomerular filtration rate <90 ml/min/1.73 m), fluid overload (FO) at CRRT initiation, and indication for CRRT initiation. Patients with an FO exceeding 10% at the initiation of CRRT were classified as having FO as an indication for CRRT. We also collected data on the patients ionized calcium (pre- and post-connection), pre-connection patient arterial pH, vasoactive inotrope score (VIS), hematocrit, pediatric logistic organ dysfunction-2 (PELOD 2) score at ICU admission, and each filter change. Lastly, we collected CRRT prescription data, which included blood flow, priming solution, rapid versus reprimed, location of dialysis line, prescribed clearance (ml/min/1.73 m), prescribed total ultrafiltration rate in the first hour after CRRT connection (ml/kg/h), achieved total ultrafiltration rate in the first hour after CRRT connection (ml/kg/h), and effluent dose (ml/kg/h). Our primary outcome was MAKE30, which is a composite outcome that includes (1) diminished renal function measured by an estimated glomerular filtration rate <75% of the baseline, (2) continual need for dialysis (any modality), or (3) death.

Building on our previous work, IDH was defined as a >20% decrease from the baseline mean arterial pressure (MAP) sustained for at least two consecutive minutes. Baseline MAP was defined as the mean MAP during the hour prior to CRRT connection. Data on the total number of CRRT connections that met the IDH criteria were collected. Given the varying number of CRRT connections during our observation period, for each patient, we calculated IDH burden by dividing the number of CRRT connections meeting the IDH criteria by the total number of CRRT connections censored at 30 days after CRRT initiation. Finally, patients were classified as having any IDH if they developed IDH at any of the observed CRRT connections.

Connections with greater than 25% missing hemodynamic data were excluded. After removal of these connections, a total of 328 (2.4%) MAP readings were missing and required imputation. Missing values were imputed using the k-nearest neighbor algorithm based on MAPs over the 60-minute observation period using the weighted average and a k = 10. After imputation, the final dataset was compared with the original data using the mean and standard deviation, and we confirmed that there were no significant differences before and after imputation

Then, MAP values in mmHg were converted to age-based percentiles using linear interpolation.

Using the completed dataset, we visualized MAP trajectories using line graphs in 25 (10%) of our final cohort. We then assigned each of these 25 trajectories to two groups (increase or decrease) based on our visual assessment of the MAP trend over the 60 minutes following CRRT connection (Supplementary Fig. S2 and Supplementary Table S1).

We compared hierarchical and K-means clustering approaches by first generating elbow plots for each method to determine the optimal number of clusters. The clustering algorithm that yielded the highest total within-cluster sum of squares which indicate greater cluster/group cohesion, was selected for further analysis (Supplementary Figs. S3 and S4). Then we evaluated two distance metrics: Euclidean distance and dynamic time warping (DTW) by assessing the silhouette coefficient for each. Prior studies have suggested that an average silhouette coefficient greater than 0.5 reflects a meaningful separation between clusters, with values closer to +1 indicating stronger inter-cluster distinctness. The distance metric associated with the highest average silhouette coefficient was selected (Fig. 1 and Supplementary Fig. S5). Based on this methodology, our final unsupervised machine learning model consisted of K-means clustering with a k = 2 using dynamic time warping as the distance measure. Once our final model was selected, we used the MAP percentile at 1-minute intervals for the 60 minutes immediately after CRRT connection for clustering.

To further evaluate model performance, we used the Adjusted Rand Index (ARI) in two distinct ways. First, to assess the concordance between expert-defined and algorithm-derived classifications, we compared our clinician-assigned "gold standard" groupings with the K-means cluster assignments. Second, to evaluate the stability of the clustering algorithm, we ran K-means clustering across 100 different random initializations (seeds) and compared each result to the original clustering solution. The ARI was calculated for both evaluations, with a value of 1.0 indicating perfect agreement, values between 0.5 and 0.9 indicating substantial agreement, and a value of 0 suggesting agreement no better than random chance.

Continuous variables were presented as median with interquartile range (IQR), while categorical variables were expressed as frequencies and counts. Logistic regression was used to analyze associations between IDH burden, clinical variables, and MAKE30. Variables with a p < 0.2 in unadjusted analysis, along with factors deemed clinically important, were evaluated for potential inclusion in the final multivariable logistic regression model. To optimize the model fit, we employed the Aikiki Information Criterion minimization technique. The associations within our final adjusted model were reported as beta coefficients along with their corresponding 95% confidence intervals (CIs).

Given that each patient had multiple CRRT connections, upon patient clustering, we assigned individuals to their predominant cluster (the cluster to which the patient had the highest number of connections). In instances of equal distribution, the patient was allocated to the cluster with the lowest total number of events. Chi-square test to compare categorical variables, while the Mann-Whitney rank sum test for continuous variables was used to identify significant differences between clusters. R version 2024.04.0 with the packages dplyr v.1.1.4, tidyr 1.3.1, gtsummary v.1.7.2, dtwclust v6.0.0, and DMwR2 v.0.0.2 were employed for our analysis.