Biomarker-based risk model to predict persistent multiple organ dysfunctions after congenital heart surgery: a prospective observational cohort study

Background Multiple organ dysfunction syndrome (MODS) is an important cause of post-operative morbidity and mortality for children undergoing cardiac surgery requiring cardiopulmonary bypass (CPB). Dysregulated inflammation is widely regarded as a key contributor to bypass-related MODS pathobiology, with considerable overlap of pathways associated with septic shock. The pediatric sepsis biomarker risk model (PERSEVERE) is comprised of seven protein biomarkers of inflammation and reliably predicts baseline risk of mortality and organ dysfunction among critically ill children with septic shock. We aimed to determine if PERSEVERE biomarkers and clinical data could be combined to derive a new model to assess the risk of persistent CPB-related MODS in the early post-operative period. Methods This study included 306 patients < 18 years old admitted to a pediatric cardiac ICU after surgery requiring cardiopulmonary bypass (CPB) for congenital heart disease. Persistent MODS, defined as dysfunction of two or more organ systems on postoperative day 5, was the primary outcome. PERSEVERE biomarkers were collected 4 and 12 h after CPB. Classification and regression tree methodology were used to derive a model to assess the risk of persistent MODS. Results The optimal model containing interleukin-8 (IL-8), chemokine ligand 3 (CCL3), and age as predictor variables had an area under the receiver operating characteristic curve (AUROC) of 0.86 (0.81–0.91) for differentiating those with or without persistent MODS and a negative predictive value of 99% (95–100). Ten-fold cross-validation of the model yielded a corrected AUROC of 0.75 (0.68–0.84). Conclusions We present a novel risk prediction model to assess the risk for development of multiple organ dysfunction after pediatric cardiac surgery requiring CPB. Pending prospective validation, our model may facilitate identification of a high-risk cohort to direct interventions and studies aimed at improving outcomes via mitigation of post-operative organ dysfunction. Supplementary Information The online version contains supplementary material available at 10.1186/s13054-023-04494-7.


Background
Cardiopulmonary bypass (CPB) potentiates a systemic inflammatory response in all patients, the degree of which varies based on many factors [1][2][3][4][5][6][7][8][9]. An exaggerated response, as seen in systemic inflammatory response syndrome (SIRS), can be detrimental and contributes to the development of multiple organ dysfunction (MODS), prolonged length of stay, and worse outcomes [5][6][7]. Almost all pediatric cardiac surgery patients meet criteria for organ dysfunction in the early postoperative period with ubiquitous inotropic and/or mechanical ventilator support, but children with optimal surgical interventions will begin to wean from postoperative support within the first few days. Failure to wean may represent persistent or progressive organ dysfunction, with risk of mortality increasing in conjunction with number of organ systems involved [10,11]. Identifying patients at increased risk for persistent MODS due to an exaggerated inflammatory response to CPB could help guide clinical management, provide prognostic enrichment in future trials, and, ultimately, improve outcomes.
Sepsis and CPB both cause cellular injury and release of molecules that activate the innate and adaptive immune responses resulting in pro-inflammatory mediator upregulation [1,3]. Research focusing on innate and adaptive immune gene expression and profiling in pediatric sepsis generated the Pediatric Sepsis Biomarker Risk Model (PERSEVERE) [12][13][14][15][16][17][18][19][20]. PERSEVERE and, more recently, PERSEVERE II have been utilized as risk-stratification tools to estimate probability of mortality and organ dysfunctions in pediatric septic patients [18]. Research on sepsis and CPB-mediated inflammation has identified significant overlap in inflammatory biomarker activation, including PERSEVERE biomarkers [5,[21][22][23][24][25][26][27]. We therefore posited that PERSEVERE biomarkers could be used to derive a unique risk model for early prediction of persistent MODS after CPB in pediatric patients.

Patients, samples, and data collection
The study was approved by the institutional review board at Cincinnati Children's Hospital Medical Center. All patients under the age of 18 years old undergoing surgery requiring CPB for correction of congenital heart disease between November 2016 and November 2020 were screened for eligibility. Patients were only included for their index surgery to prevent re-enrollment of patients requiring reoperation for residual lesions while still recovering from their initial surgery. For patients with single ventricle physiology, each surgical stage was treated as a separate index surgery, i.e., stage 1 palliative surgery, Glenn operation, Fontan operation, and/or biventricular repair. Due to the short time frame between stage 1 and Glenn, Glenn candidates were screened prior to re-enrollment and were excluded if they met criteria for organ dysfunction at time of screening. Patients undergoing CPB for heart or lung transplantation, patients requiring immunosuppression, and patients with suspected or proven infection were excluded. Threehundred and fifty-nine patient encounters (293 unique patients) were consented for the study. Of these, 306 encounters were included in the analysis, because both 4-and 12-h biomarker samples were collected within the specified time. Baseline demographic, clinical, and laboratory data needed to calculate severity of illness scoring and determine organ dysfunction were extracted from the electronic medical record (EMR). To minimize clinically unnecessary blood draws, laboratory data to assess for organ dysfunction were only collected at discretion of the managing clinical team.

Definitions
The Society of Thoracic Surgery-European Association for Cardiothoracic Surgery (STAT) mortality category [28,29] was used to account for risk related to surgical complexity. Pre-and postoperative severity of illness was assessed using Pediatric Risk of Mortality score III (PRISM III) [30]. Organ dysfunction was defined via adaption of Goldstein criteria [31] to account for differences in the postoperative congenital heart disease population when compared to the pediatric sepsis population, Additional file 1. Persistent MODS was defined a priori as dysfunction of 2 or more organ systems on postoperative day 5. As an additional measure of organ dysfunction, daily Pediatric Logistic Organ Dysfunction-2 (PELOD-2) scores were calculated preoperatively and for the first 5 postoperative days [32,33].

Clinical and surgical management
All patients received methylprednisolone (30 mg/kg) as part of the CPB circuit prime. Neonates and patients in the hospital prior to their scheduled operation received an additional dose of methylprednisolone (30 mg/kg) the morning of surgery (prior to CPB initiation). Choice of anesthesia was not standardized and left to the decision of the cardiac anesthesiologist. All patients received either modified ultrafiltration and/or continuous ultrafiltration intraoperatively, based on surgeon preference. The need for additional steroids and use of postoperative peritoneal dialysis was left to the discretion of the clinical team.

Statistical analysis
Descriptive statistical analyses were performed using R (version 4.0.4). Demographic, clinical, and biomarker data were described using medians with interquartile ranges (IQR), means with standard deviations, or frequencies with percentages as appropriate. Comparisons of data for patients with and without persistent MODS were performed using the Kruskal-Wallis, Chi-squared, or Fisher's exact tests as appropriate. Multivariate regression analysis, controlling for clinical data, was performed to examine the relationship between biomarker concentrations at 4 and 12 h and risk of development of MODS.
Classification and regression tree (CART) analysis was used to determine biomarker cut-points and derive a decision tree (Salford Predictive Modeler v6.6, Salford Systems, San Diego, CA) [35]. Candidate prediction variables for derivation of the decision tree were as follows: all seven PERSEVERE biomarkers at 4 and 12 h time points, change in PERSEVERE biomarker levels from 4 to 12 h, age in months (included as both continuous and dichotomous variables), single ventricle status, history of prematurity, CPB time, maximum vasoactive inotropic score (VIS), and STS-EACTS mortality category. Clinical predictor variable selection was based on extant literature [36][37][38][39]. Tuning parameters determined a priori included: tenfold cross-validation, at least one of the paired terminal daughter nodes contains ≥ 5% of the subjects in the root node, and no predictor variables repeated within one of the two main branches. Performance of the decision tree was determined by generating a classification table of true versus predicted status and calculation of discrimination metrics including sensitivity, specificity, positive and negative predictive values, and area under the receiver operating curve (AUROC). We compared our prediction model, which we will refer to as PERSEVERE-CPB, to PRISM III and STS-EACTS mortality category, as they are widely accepted and validated risk assessment and severity of illness scoring systems of this patient population, using the AUROC, sensitivity, and specificity. We further compared PER-SEVERE-CPB to the 24-h postoperative PELOD-2 score, as PELOD-2 is a validated scoring system for organ dysfunction [32].
Using risk categories (referred to as PERSEVERE-CPB risk category) stratified the cohort into risk category based on high-, intermediate-, and low-risk terminal nodes of our model. We then evaluated the association of risk category with administration of postoperative steroids for hypotension and clinical outcomes.
Finally, we performed an uncontrolled subanalysis comparing biomarker concentrations in subjects who received dialysis (peritoneal or continuous renal replacement therapy) within the first 24 h after surgery to assess the potential effect of dialysis on biomarker concentration.

Results
Demographics, clinical characteristic, and biomarker concentrations of patients with and without persistent MODS are shown in Tables 1, 2 and Fig. 1, and Additional file 2. Of the 306 subjects with biomarkers drawn at both 4 and 12 h after separation from CPB, 43 (14.1%) had persistent MODS on POD 5. The cohort with persistent MODS was significantly younger, had a history of prematurity, had higher illness severity before and immediately after CPB, received more organ support, was more likely to receive steroids for post-operative hypotension, and had worse clinical outcomes. In multivariate logistic regression models, accounting for age less than 12 months, STAT mortality category, CPB time, and single ventricle status, IL-8 concentration at both 4 and 12 h was independently associated with risk of persistent MODS, as did 12-h concentrations of GZMB and CCL3, as shown in Table 2.

Biomarker-based risk prediction model
Our newly derived PERSEVERE-CPB model is shown in Fig. 2. PERSEVERE-CPB included IL-8 concentration at 12 h, the change in serum concentration of CCL3 from 4 to 12 h, and infant age category (< 12 months). There were two low-risk terminal nodes (terminal nodes 1 and 3) in which subjects had < 2% risk of developing persistent MODS. There was one intermediate-risk node with 23 patients (20.5%) who developed persistent organ dysfunction (terminal node 2). There was one high-risk node with persistent organ dysfunction in 72% of patients (terminal node 4). PERSEVERE-CPB performed well at determining risk of persistent MODS with model characteristics shown in Table 3. IL-8 concentration at 12 h functioned as the upper tier decision rule, thus having the most predictive weight. Age less than 12 months was the second most important predictive variable, followed by change in the serum concentration of CCL3 from 4 to 12 h.   The portion of the cohort falling into the high-risk PERSEVERE-CPB category (terminal node 4 of model) were more likely to receive steroids for post-operative hypotension compared to those falling into the intermediate-and low-risk categories (35%, 22%, 2%, respectively; p < 0.001). The high-risk cohort also experienced longer duration of ventilator and vasoactive support, longer CICU and hospital stays, and had higher in-hospital mortality compared to those falling into intermediateand low-risk categories, Table 4.

Biomarker concentrations in patients receiving dialysis
Dialysis, either continuous renal replacement therapy (CRRT) or PD, was used in 12 patients in the first 24 h after separation from CPB, with 9 being infants. Peritoneal dialysis catheters drained ascites without active dialysis in the remaining 34 neonates. Use of dialysis was associated with increased IL-8 at both 4 and 12 h post-CPB. CCL-3 concentrations were higher in the dialysis group, but only 12 h concentrations in the entire cohort were significant, Additional file 3.

Discussion
Using inflammatory biomarkers and established clinical risk factors, we have derived a decision tree that is able to stratify patients by risk for developing persistent multiple organ dysfunction syndrome at post-operative day 5 after cardiopulmonary bypass surgery for congenital heart disease. Of the clinical risk factors and biomarkers included in this study, interleukin 8 (IL-8) concentration was the most important predictor of persistent MODS.  The model functions exceptionally well in identifying low risk patients, as illustrated by a high negative predictive value and low negative likelihood ratio. However, given the low positive predictive value and low positive likelihood ratio, PERSEVERE-CPB over-selects for MODS (approximately 31% false positive rate), which may limit its utility and result in overtreatment of the intermediate and high risk. Despite only a modest ability to predict MODS, our model enables the clinician to increase vigilance in a smaller cohort of patients, which has added importance as those falling into the high-risk PERSEVERE-CPB strata experienced worse clinical outcomes (longer duration of ventilator and vasoactive support, longer duration of stay, higher in-hospital morality) compared to the intermediateand low-risk groups. This model has the potential to allow for early identification of patients categorized as low risk to receive standard of care supportive therapies, and those at intermediate or high risk to receive early targeted clinical interventions aimed at reducing the risk of MODS. Additionally, separation of low and higher risk cohorts may allow for prognostic enrichment in future clinical trials of interventions aimed at mitigating organ dysfunctions. However, without a rapid point of care PERSEVERE biomarker panel that allows for real time risk stratification, utility of PERSE-VERE-CPB is limited. There are ongoing efforts at our institution focused on the development of a rapid point of care PERSEVERE biomarker panel. Once available, future studies will focus on timing and implementation of PERSEVERE-CPB in efforts to improve postoperative outcomes, including reduction in MODS. Future work with also focus on the addition of real-time physiologic    and laboratory data to the model may improve the precision and specificity of this model. For assessing risk of persistent MODS, PERSEVERE-CPB performed well when compared to existing pediatric critical care and cardiac surgery risk-assessment tools (STAT, PRISM III, PELOD 2). In particular, PER-SEVERE-CPB performed similarly to the postoperative day one PELOD-2 score for predicting development of persistent MODS. Although STAT and PRISM III were primarily validated to predict risk of mortality and not MODS, the low mortality rate in our cohort did not allow us to develop a biomarker-based predictive model for in-hospital mortality.
IL-8 level functioned as the upper level decision rule, indicating that it played a key role in determination of risk for MODS. Almost 42% of patients who developed persistent MODS fell into terminal node 4, with an elevated 12 h IL-8 concentration. IL-8 is one of the more studied biomarkers of inflammation in patients after CPB. It is a neutrophil chemoattractant, plays a pivotal role in neutrophil activation, and is produced in large quantities by endothelial cells [40]. Elevated postoperative IL-8 has been associated with markers of low cardiac output (low mixed venous oxygen concentration and higher inotropic score) [41], development of postoperative acute kidney injury [26,42,43], increased duration of mechanical ventilation [22,43,44], and longer ICU length of stay [6]. The pathophysiologic role IL-8 plays in neutrophil/endothelium activation, bypass-mediated inflammation, and development of MODS warrants further examination, with obvious potential as a therapeutic target. In comparison, CCL3, or macrophage inflammatory protein 1α (MIP-1α), has not been extensively studied in bypass-mediated inflammation. During acute inflammation, CCL3 aids in the recruitment of leukocytes and plays a role in neutrophil infiltration [45,46]. Since both PERSEVERE and PERSEVERE-II have demonstrated CCL3 plays a major role in discrimination of both mortality and multiple organ failure in severe pediatric sepsis [47], further investigation into the role of CCL3 in CPB-mediated inflammation and its contribution to development of organ dysfunction is warranted.
Age less than 12 months at time of surgery functioned as the second level decision rule in PERSEVERE-CPB. Younger age is known to be associated with increased morbidity after pediatric cardiac surgery [36,37,48], which is not a surprise given that infants and neonates undergo the most complex and highest risk surgeries. Future efforts to create risk models specific to infants and neonates could help determine if there are modifiable risk factors or potential therapeutic targets or if their increased risk is attributable to complexity of surgery and cardiac physiology (such as single ventricle physiology) alone.
Perioperative steroids are used in children undergoing CPB to blunt the bypass-mediated inflammatory response [49]. Non-uniform perioperative steroid administration in our population is a limitation of this study: while all patients received steroids, preoperatively hospitalized neonates and infants received steroids both before and during CPB, whereas all other patients received steroids only during CPB. It is unclear how additional steroid administration might have impacted these results, if at all. We acknowledge that perioperative steroids blunt the bypass-mediated inflammatory response; it is possible additional dosing may have resulted in even lower biomarker concentrations. Interestingly, the majority of the high-risk cohort (17 out of 20 subjects) were hospitalized neonates and infants which may suggest that inflammation has a bigger impact in outcome in this subset of patients, despite receiving two doses of steroids. Perhaps also supporting this theory, the high-risk cohort was more likely to receive steroids for hypotension in the first 24 h postoperative, which may reflect an enhanced inflammatory response leading to higher degree or longer lasting vasoplegia (Table 4).
Another limitation of the study was placement of peritoneal dialysis catheters in neonates, which is a standard practice at our institution. Use of peritoneal dialysis has been shown to decrease inflammatory cytokines after bypass and in other inflammatory states [50]. Unlike prior studies, use of dialysis was associated with increased IL-8 at both 4 and 12 h post-CPB in both the entire cohort and the neonatal subpopulation. CCL-3 concentrations were higher in the dialysis group, but only 12 h concentrations in the entire cohort were significant, Additional file 3. It is likely in this uncontrolled subanalysis that higher biomarker concentrations reflect treatment bias as opposed to a potentiating biomarker effect of peritoneal dialysis, as younger and higher risk patients/surgeries (with presumably more inflammatory CPB-response) are standardized to receive peritoneal catheters and undergo dialysis postoperatively. We cannot surmise the direct impact of peritoneal drainage and dialysis in this cohort without paired analysis and baseline biomarker levels. Future studies comparing postoperative inflammatory biomarker concentrations over time, use of dialysis, and correlation with risk of persistent MODS, particularly in the neonates and infants who, in this study, comprise a majority of the most at risk population, are warranted.
Also limiting our study, 3 subjects who developed MODS had a residual lesion or complication of care that contributed to prolonged need for mechanical ventilation and inotropic/vasopressor support. It would have been preferable to include only patients with MODS resulting from biologic and physiologic consequences of surgery and their intrinsic response to inflammation. Lastly, although the small number of events in this study prevented validation beyond a tenfold cross-validation procedure, we hope to be able to enhance this in a future, multicenter study. Cross-validation AUC for our model showed acceptable ability to predict persistent MODS, comparable to postoperative PRISM III and PELOD-2.

Conclusions
Using known clinical risk factors and biomarkers of inflammation originally identified as key markers of inflammation in pediatric patients with septic shock, we have created a simple, biologically plausible model that accurately predicts risk of persistent organ dysfunction in pediatric patients after cardiac surgery for congenital heart disease. IL-8 concentration was the most predictive variable for development of MODS after CPB in our patient population; future efforts to better define CPBrelated IL-8 pathophysiology and modifiable risk factors for IL-8 elevation after CPB are warranted.