Pre‐ and Immediate Postoperative Prediction Model for Organ Dysfunction or Death Early After Cardiac Surgery: A Post Hoc Analysis of a Multicenter Randomized Trial

Abstract

Background

Development of organ dysfunction or death is still common in patients undergoing cardiac surgery. Yet, current risk stratification tools fail to adequately incorporate both preoperative vulnerability and immediate postoperative physiological derangements. This study aims to develop a predictive model integrating these critical timepoints to identify high‐risk patients for presence of organ dysfunction or death 48 hours after surgery.

Methods

This is a post hoc analysis of an international, multicenter, randomized, controlled trial in patients undergoing cardiac surgery (n=1394). Prespecified patient characteristics (age, Clinical Frailty Scale, at nutrition risk, combined procedures, urgent surgery, moderate–severe chronic kidney disease, left ventricular ejection fraction, European System for Cardiac Operative Risk Evaluation II, cardiopulmonary bypass duration, sex, Charlson Comorbidity Index, and Sequential Organ Failure Assessment score) were included in logistic regression models employing bootstrap validation.

Results

A total of 434 (31.1%) patients had organ dysfunction or died 48 hours after surgery. The preoperative model identified Clinical Frailty Scale, nutrition risk, urgent surgery and European System for Cardiac Operative Risk Evaluation II as significant predictors of organ dysfunction or death 48 hours after surgery (optimism‐corrected area under the receiver operating characteristic curve, 0.644 [95% CI, 0.610–0.678]). Incorporation of postoperative variables (Sequential Organ Failure Assessment score at intensive care unit admission, and cardiopulmonary bypass duration) improved predictive performance (area under the receiver operating characteristic curve, 0.773 [95% CI, 0.745–0.801]).

Conclusions

Incorporation of variables collected the day of surgery substantially improved the ability to predict organ dysfunction or death 48 hours after surgery compared with using presurgical variables only. This pragmatic, clinically actionable model may enable targeted resource allocation and personalized interventions and may provide a stratification tool for future research.

Registration

URL: clinicaltrials.gov; Unique Identifier: NCT02002247.

Nonstandard Abbreviations and Acronyms

CPB

cardiopulmonary bypass

EuroSCORE II

European System for Cardiac Operative Risk Evaluation II

SOFA

Sequential Organ Failure Assessment

SUSTAIN CSX

Sodium Selenite Administration in Cardiac Surgery

Clinical Perspective.

What Is New?

• In contrast to traditional cardiac surgery risk models, which use static preoperative scores and, thus, are missing critical interactions between baseline vulnerability and postoperative stress, our innovative model integrates preoperative frailty, surgical complexity, and early postoperative physiological response and shows that the Sequential Organ Failure Assessment score serves as the crucial link connecting baseline risk to ultimate outcomes.

What Are the Clinical Implications?

• Our findings enable (1) early high‐risk detection during perioperative evaluation, (2) targeted interventions for modifiable factors (eg, nutritional optimization in frail patients), and (3) dynamic monitoring adjusted to individual risk trajectories.

Cardiac surgery trials face a critical challenge: traditional composite endpoints like major adverse cardiac/cerebrovascular events fail to capture the spectrum of postoperative organ dysfunction that drives prolonged stays at an intensive care unit (ICU) and enhanced resource use. ¹ Although these ischemic endpoints are valuable in cardiology, patients undergoing cardiac surgery suffer from distinct risks—including acute kidney injury, respiratory failure, and vasopressor dependence—that are better reflected by persistent organ dysfunction. ¹ , ² This disconnection contributes to the well‐documented difficulties in conducting successful cardiac surgery trials, where slow recruitment, operator variability, and low event rates often lead to underpowered studies. ³ , ⁴ The problem is compounded by modern advances in surgical technique and perioperative care.

As demonstrated by the SUSTAIN CSX (Sodium Selenite Administration in Cardiac Surgery) trial, ⁴ improved baseline characteristics have reduced the incidence of persistent organ dysfunction or death within 30 days, making it increasingly difficult to detect treatment effects without using enrichment strategies in patient selection. Yet, prolonged ICU stays—frequently triggered by persistent organ dysfunction—remain a major burden, accounting for >30% of cardiac surgery costs and correlating with long‐term functional decline. ⁵ , ⁶ Early identification of high‐risk patients is therefore both a scientific imperative and a clinical necessity.

In this context, the first 48 hours after surgery represent a pivotal juncture: patients showing unresolved organ dysfunction at this time point face 3 to 5 times higher mortality and prolonged hospitalization. ⁴ , ⁷ , ⁸ , ⁹ Thus, the 48‐hour time point serves as a critical window to identify patients who are at increased risk for presence of organ dysfunction based on predictive characteristics, and, based on that, to assess the effectiveness of interventions and to adjust treatment plans accordingly, ultimately aiming to improve patient prognosis.

Hence, the objective of this post hoc analysis is to develop a predictive model using pre‐ and immediate postoperative patient characteristic variables to identify a subpopulation among patients undergoing major cardiac surgery who are at increased risk of development of organ dysfunction or death within 48 hours after surgery.

METHODS

The data that support the findings of this analysis are available from the corresponding author upon reasonable request. The study adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis + Artificial Intelligence reporting guidelines.

Design and Patients

The present report is a post hoc analysis of the SUSTAIN CSX trial, an international, multicenter, randomized, double‐blind, controlled trial in patients undergoing cardiac surgery (n=1394 included in primary analysis), which investigated the effect of perioperative selenium supplementation on the occurrence of persistent organ dysfunction or death within 30 days after surgery across 23 sites in Canada and Germany. The trial was approved by the ethics committees of Queens University, Canada, and RWTH Aachen University, Germany, the German Federal Institute for Drugs and Medical Devices, and by all participating centers. There was no patient or public involvement in this trial. Written informed consent to participate in the trial was obtained from all patients before surgery. The trial protocol and primary results of the SUSTAIN CSX trial have been published elsewhere. ⁴ , ¹⁰

Patients were included if they were ≥18 years, scheduled to undergo elective or urgent cardiac surgery with the use of cardiopulmonary bypass (CPB) and cardioplegic arrest. Patients with single procedures, low risk for worse postoperative outcomes, hypersensitivity to sodium selenite or to any of the diluent vehicle, total bilirubin more than 2.0 mg/dL, disabling neuropsychiatric disorders, pregnancy or lactation period, current antioxidant use, cardiac transplant, planned ventricular assist device implantation, correction of complex congenital anomalies, or planned use of hypothermic cardiocirculatory arrest were excluded.

The current analysis was conceived after completion of the SUSTAIN CSX trial.

Intervention

Patients were randomly assigned (1:1 ratio) to receive the intervention (2000 μg/L) of intravenous selenium within 30 minutes after induction of anesthesia and before initiation of CPB, then 2000 μg/L of intravenous selenium immediately on admission to the postoperative ICU, then 1000 μg/L of intravenous selenium each successive morning while in ICU or placebo (saline) at the same time points for a maximum of 10 days.

Outcome and Data Collection

The primary outcome of the SUSTAIN CSX trial was days free from persistent organ dysfunction or death in the first 30 days after surgery, whereas the primary outcome of the current analysis is presence of organ dysfunction or death on the second day after surgery. Persistent organ dysfunction was defined in accordance to the initial definition as need for life‐sustaining therapies—including mechanical ventilation, noninvasive ventilation, any vasopressor therapy, mechanical circulatory support, continuous kidney replacement therapy, or intermittent hemodialysis—that developed postoperatively at any time during the day ⁵ and has been evaluated in a blinded manner. This outcome has been used as primary endpoint for many subsequent international studies. Between January 2015 and January 2021 the following variables were collected in a blinded manner by qualified, trained, and experienced medical personnel (ie, physicians, nurses) according to their respective authorities and responsibilities: (1) preoperatively—age, sex, ethnicity, Charlson Comorbidity Index (mortality prediction of a patient who may have a range of concurrent conditions), European System for Cardiac Operative Risk Evaluation (EuroSCORE II), Clinical Frailty Scale (standardized 9‐point rating tool), nutrition risk status (evaluated by screening tool routinely used at the participating sites), kidney function, left ventricular ejection fraction (LVEF), surgical procedures; (2) postoperatively—CPB time, Sequential Organ Failure Assessment (SOFA) score (standardized medical coding including respiratory, cardiovascular, hepatic, coagulatory, renal, and neurological measures), clinical outcome parameters (eg, daily presence of components of persistent organ dysfunction and mortality); and (3) pre‐ and postoperative—laboratory measures. ⁴ , ¹⁰

Statistical Analysis

In the present post hoc analysis, the sample size was fixed by the data set of the SUSTAIN CSX trial. We examined the ability of specific pre‐ (evaluated before surgery) and immediate postoperative (evaluated immediately after surgery) patient variables, which have been collected within the methodology of the SUSTAIN CSX trial by blinded researchers, to predict presence of organ dysfunction or death 48 hours after surgery. The 10 presurgery variables considered included (1) age, (2) sex, (3) Charlson Comorbidity Index, (4) EuroSCORE II, (5) Clinical Frailty Scale, (6) at nutrition risk (reduced oral intake or recent weight loss), (7) moderate–severe chronic kidney disease, (8) LVEF, (9) combined procedures, and (10) urgent surgery; and the two immediate postsurgical variables were CPB duration and SOFA score at ICU admission. The selected variables were primarily chosen based on their prespecified role in subgroup analyses of the SUSTAIN CSX trial ⁴ and their established association with adverse outcomes. ¹⁰ , ¹¹ , ¹² Additionally, clinically relevant parameters (ie, age, Charlson Comorbidity Index, SOFA score at ICU admission) were included due to their documented prognostic value in prior studies. ¹³ , ¹⁴ , ¹⁵ This analysis develops a prediction model for presence of organ dysfunction or death 48 hours after surgery allowing for a smooth nonlinear relationship between the continuous variables and the log of the odds (logit). To avoid overfitting, we did the following: (1) limited candidate predictors to the nine variables prespecified for the subgroups in the primary analysis of the SUSTAIN CSX trial plus three additional predictors (sex, Charlson comorbidity index, and SOFA score at ICU admission), (2) made all decisions regarding variable transformation and spline expansion of numeric variables before considering the outcome, and (3) did not consider interactions because there was no a priori expectation of effect modification.

Frailty index, Charlson comorbidity index, and SOFA score at ICU admission were modeled as linear on the logit since their relationship with outcome could safely be assumed monotonic, and these were discrete with a majority of patients having five or fewer distinct values. Each of the remaining three numeric variables (LVEF, EuroSCORE II, and CPB duration) were modeled using restricted cubic splines with knots at 5%, 27.5%, 50%, 72.5%, and 95% as suggested by Harrell. ¹⁶ However, we also considered models where all numeric predictors were modeled as linear on the logit. Before modeling, the distribution of the numeric predictors was examined and a natural log transformation for distributions that looked approximately log‐normal was considered. The log transformation was not based on the association with the outcome but rather to avoid outliers in the long right tail that could result in a small subset of observations being overly influential. The five remaining binary predictors were modeled using dummy variables coded as 1 for yes and 0 for no.

For model selection, backward stepwise selection with an exit and reentry criteria of P=0.157 was used to approximate the Akaike information criterion. ¹⁷ The four basis terms that defined the cubic spline for each numeric variable were tested, removed, or reentered collectively even if individual basis terms did not meet the inclusion/exclusion threshold. The three basis terms that added nonlinearity were tested collectively to assess nonlinearity.

We compared the performance of models selected by backward stepwise selection to models with all candidate predictors included (full models), as well as models including restricted splines to models only using linear terms, and most important, models including only presurgical variables to models including presurgical plus immediately postsurgical variables. For each of the aforementioned eight models, the concordance‐index (c‐index; also known as area under the receiver operating characteristic curve) was used to measure the discriminative ability of the model and the Brier score; and rescaled generalized R‐square were used to measure the overall fit. We also reported receiver operating characteristic curves (sensitivity versus 1‐specificity at each possible probability cut‐point), calibration plots with corresponding Spiegelhalter's test (against the null of correct calibration), and the distribution of the estimated probability of event among participants with and without the event (persistent organ dysfunction or death at 48 hours after surgery).

We employed multiple imputation to use the full available sample while accounting for uncertainty due to sporadic missing predictor variables. We generated 20 randomly imputed data sets using fully conditional specification as implemented in the multiple imputation procedure of the SAS/STAT software. The fully conditional specification used linear regression for the continuous predictors and the discriminant function method for binary variables. All predictors considered in the prediction modeling as well as mortality at day 30 and persistent organ dysfunction or death at day 2, were included in the imputation models with variables ordered from least (no missing) to most missing data (LVEF). Values are randomly drawn from their expected distribution given the values of the other predictors to avoid bias under missing at random conditional on the other variables included in the imputation model. The logistic regression model fitting and selection was then performed separately within each of the imputed data sets and final parameter estimates with standard errors and model performance measures were combined using Rubin's rules. ¹⁸ The reported area under the receiver operating characteristic curve, calibration curves, and histograms of estimated event probabilities were each based on the true event status (never missing) and the average predicted event probability across the 20 imputations. The estimated event probabilities varied so minimally across the 20 imputations that we felt reporting imputation specific graphics was unnecessary.

We then used bootstrap validation based with 50 bootstraps within each of the 20 imputed data sets as recommended by Steyerberg. ¹⁹ Thus, the validation was based on 1000 bootstrap samples derived from 20 imputed samples. The 1000 samples were used to estimate the optimism of the c‐index, Brier score, and rescaled generalized R‐square. The optimism was then subtracted from each measure to yield the optimism corrected performance measure. ²⁰ Finally, we used the following bootstrap approach to estimate 95% CIs for the raw and optimism corrected performance measures. For each of the eight models, we reran the entire process that selected the model and estimated the optimism corrected performance measures as described previously separately within 500 bootstrap samples (1000 for the two key models using variable selection but no splines). We observed that the bootstrap distribution of all raw and optimism corrected performance measure were approximately normal, and because it was practical to run only a limited number of bootstrap samples due to memory and central processing unit run time, we used the SD of the bootstrap sample to derive 95% confidence limits by adding and subtracting 1.96 times the bootstrap SD to the performance measure. This approach was more stable than the percentile approach in our case. Collinearity of the linear predictors were assessed by the variance inflation factor with missing values imputed by the expectation–maximization algorithm. We re‐assessed the performance of our two recommended models within age and sex subgroups. All analysis was performed using SAS V9.4 (SAS Institute Inc., Cary NC, USA). The statistical analysis code is available from the corresponding author upon request.

RESULTS

Patient Cohort and Characteristic Variables

All 1394 patients included in the primary analysis of the SUSTAIN CSX trial were included in this post hoc analysis (Figure S1). At 48 hours after surgery, 13 (0.9%) of the 1394 patients had died and an additional 421 (30.2%) had presence of organ dysfunction for an overall presence of organ dysfunction or death rate of 434 (31.1%). Of the 960 patients alive without presence of organ dysfunction or death 48 hours after surgery, 6 (0.6%) patients died by postoperative day 30. In comparison, 45 (10.7%) of the 421 patients alive with presence of organ dysfunction 48 hours after surgery died by postoperative day 30. All 22 patients alive with presence of organ dysfunction on day 30 had presence of organ dysfunction 48 hours after surgery. Patients alive without presence of organ dysfunction 48 hours after surgery had 30‐ and 180‐day mortality rates of 0.6% and 2.0%, respectively, compared with mortality rates of 13.4% and 19.5% in the 434 patients with presence of organ dysfunction or death 48 hours after surgery.

The characteristics of patients with and without presence of organ dysfunction or death 48 hours after surgery are shown in Table 1. Patients with presence of organ dysfunction or who died within 48 hours after surgery were more often female (P=0.005), had older age (P=0.01), were more often in need for urgent surgery (P=0.005), were at nutrition risk (P=0.002), and had higher SOFA score at ICU admission (P<0.0001), EuroSCORE II (P<0.0001), Clinical Frailty Scale (P<0.0001), prevalence of moderate or severe chronic kidney disease (P=0.01), and longer CPB duration (P<0.0001) compared with patients who were alive without organ dysfunction 48 hours after surgery. In addition, LVEF was lower in patients with presence of organ dysfunction compared with patients who were alive and those who died within 48 hours after surgery (P<0.0005).

Table 1.

Variables of Patients With and Without Presence of Organ Dysfunction and Those Who Died 48 Hours After Surgery

Patient variables	Without presence of organ dysfunction 48 h after surgery (n=960)	With presence of organ dysfunction 48 h after surgery (n=421)	Mortality at 48 h after surgery (n=13)	P value*
Age, y	67.7±10.7	69.3±9.7	71.8±4.8	0.01†‡
Sex				0.005‡
Male	730 (76%)	308 (73%)	5 (38%)
Female	230 (24%)	113 (27%)	8 (62%)
Surgical urgency				0.005
Elective	815 (85%)	327 (78%)	8 (62%)
Urgent	142 (15%)	92 (22%)	5 (38%)
Emergent	3 (0%)	2 (0%)	0 (0%)
Charleson Comorbidity Index	1.4±1.4	1.5±1.4	1.8±1.3	0.06†‡
Functional Comorbidity Score	1.5±1.4	1.5±1.3	1.8±1.0	0.61
Sequential Organ Failure Assessment score at ICU admission	7.2±2.6	9.7±2.9	10.5±4.3	<0.0001†‡
European System for Cardiac Operative Risk Evaluation II	10.5±8.3	14.3±10.8	20.2±8.7	<0.0001†‡
At nutrition risk	206/960 (21%)	128/421 (30%)	3/13 (23%)	0.002
Clinical Frailty Scale	2.7±0.9	3.0±1.1	3.2±0.9	<0.0001†‡
Combined procedures	758/953 (80%)	315/420 (75%)	10/13 (77%)	0.17
Moderate–severe chronic kidney disease	338/960 (35%)	168/421 (40%)	9/13 (69%)	0.01‡
Left ventricular ejection fraction <39%	67/808 (8%)	58/365 (16%)	1/13 (8%)	0.0005
Cardiopulmonary bypass >143.6 min	336/954 (35%)	233/418 (56%)	7/13 (54%)	<0.0001†‡
Initiation of temporary mechanical assist device in the operating room	0 (0%)	12 (3%)	1 (8%)	<0.0001
Initiation of temporary mechanical assist device in the first 48 h after surgery	0 (0%)	32 (8%)	3 (23%)	<0.0001
ICU readmission rate	44 (5%)	31 (7%)	0 (0%)	0.07
Hospital readmission rate	170 (18%)	95 (23%)	0 (0%)	0.02
Moderate renal disease (creatinine clearance 50–85 mL/min)	282 (29%)	139 (33%)	7 (54%)	0.07
Severe renal disease (creatinine clearance <50 mL/min) off dialysis	56 (6%)	29 (7%)	2 (15%)	0.30
Dialysis (regardless of serum creatinine level)	1 (0)	0 (0%)	0 (0%)	0.80

ICU indicates intensive care unit.

^*The global P value is tested against the null hypothesis that all 3 groups have the same mean or proportion. When the global P value <0.05, unadjusted pairwise comparisons were made. Significant (P<0.05) pairwise comparisons are denoted as follows: ^†Survivors without persistent organ dysfunction vs survivors with persistent organ dysfunction, ^‡Survivors without persistent organ dysfunction vs decedents.

CPB time is described by procedure in Table S1.

Fourteen percent of the participants were missing LVEF only and an additional 3% were missing additional or other predictors, leaving 83% of the participants without any missing data (Table S2).

Use of Preoperative Candidate Predictors to Model Presence of Organ Dysfunction or Death 48 Hours After Surgery

The first model for predicting the presence of organ dysfunction or death 48 hours after surgery included 10 preoperative predictors using a total of 19 degrees of freedom when restricted cubic splines were included (Table 2). Backwards stepwise selection removed age, sex, Charlson Comorbidity Index, combined procedures, moderate–severe chronic kidney disease, and LVEF (for model with splines but not linear model) from the prediction model. Following this, the remaining preoperative predictor variables Clinical Frailty Scale, at nutrition risk, urgent surgery, and EuroSCORE II. It is notable that the tests for nonlinearity were never significant. We also note that when we reselected the model considering only linear predictors (ie, no splines) LVEF was retained in the model.

Table 2.

Model Selection Results Using Preoperative Candidate Predictors With Restricted Cubic Splines

Predictor	Predictor form	Df	% of bootstrap samples selecting variable	P value full model	P value for nonlinearity in full model	P‐value in selected model	P value for nonlinearity in selected model	VIF if linear†
Age	RCS	4	59%	0.338	0.353			1.14
Sex	Binary	1	19%	0.764				1.10
Charlson Comorbidity Index	Linear	1	24%	0.741				2.32
Clinical Frailty Scale	Linear*	1*	94%*	0.004*		0.001*		1.11*
At nutrition risk	Binary*	1*	81%*	0.025*		0.016*		1.05*
Combined procedures	Binary	1	44%	0.224				1.03
Urgent surgery	Binary*	1*	52%*	0.159*		0.073*		1.06*
Moderate–severe chronic kidney disease	Binary	1	31%	0.611				2.39
Left ventricular ejection fraction‡	RCS	4	76%	0.396	0.872			1.12
EuroSCORE II	RCS on log scale*	4*	100%*	<0.0001*	0.688*	<0.0001*	0.897*	1.39*

DF indicates degrees of freedom; EuroSCORE II, European System for Cardiac Operative Risk Evaluation II; RCS, restricted cubic spline; and VIF, variance inflation factor.

^*Variables were selected for the model.

^† VIF is 1/(1−R²) where R² is the proportion of variance of the given variable that is explained by all other candidate predictors in a linear regression model where all variables are binary or linear with log transformation for EuroSCORE II. VIF values <5 are generally considered not indicative of high collinearity. VIF estimates are based on single imputation estimated by the expectation–maximization algorithm whereas all other estimates use multiple imputation.

^‡ This variable was selected when the model did not use restricted cubic splines and was selected in a minority of imputed data sets.

In Table 3, we report the performance measures of models including all candidate predictors with and without using restricted cubic splines for numeric variables as well as the corresponding models after applying variable selection. We found that although the model using restricted cubic splines generally fit slightly better than the model without splines before optimism correction, after optimism correction the models allowing only for a linear association with the logit consistently outperformed the more general models using restricted cubic splines to allow for nonlinear associations. After optimism correction the full models consistently outperformed the corresponding selected model, but only trivially. Thus, we recommend the use of the selected linear model which had an optimism corrected c‐index of 0.644 (95% CI, 0.610–0.678). The model parameters with instructions on how to estimate the probability of death or organ dysfunction 48 hours after surgery are provided in Table S3.

Table 3.

Performance of the Models

	Presurgery predictors only				Pre‐ and immediate postoperative predictors
	Full model		Selected model		Full model		Selected model
Performance metric	With splines	Linear only	With splines	Linear only	With splines	Linear only	With splines	Linear only
Model degrees of freedom	19	10	7	5	24	12	13	9
Naïve c‐index	0.667 (0.637 to 0.697)	0.661 (0.631 to 0.691)	0.660 (0.627 to 0.693)	0.657 (0.625 to 0.689)	0.788 (0.763 to 0.813)	0.783 (0.758 to 0.808)	0.784 (0.757 to 0.811)	0.782 (0.755 to 0.809)
Optimism c‐index	0.026 (0.022 to 0.030)	0.013 (0.010 to 0.016)	0.023 (0.018 to 0.028)	0.013 (0.010 to 0.016)	0.019 (0.016 to 0.022)	0.009 (0.008 to 0.010)	0.018 (0.011 to 0.025)	0.010 (0.008 to 0.012)
Optimism corrected c‐index	0.641 (0.607 to 0.675)	0.647 (0.614 to 0.680)	0.636 (0.599 to 0.673)	0.644 (0.610 to 0.678)*	0.769 (0.742 to 0.796)	0.773 (0.747 to 0.799)	0.766 (0.737 to 0.795)	0.773 (0.745 to 0.801)*
Naïve R2	0.097 (0.060 to 0.134)	0.091 (0.056 to 0.126)	0.090 (0.050 to 0.130)	0.088 (0.051 to 0.125)	0.303 (0.253 to 0.353)	0.293 (0.243 to 0.343)	0.296 (0.241 to 0.351)	0.291 (0.239 to 0.343)
Optimism R2	0.039 (0.037 to 0.041)	0.020 (0.018 to 0.022)	0.035 (0.031 to 0.039)	0.020 (0.017 to 0.023)	0.044 (0.040 to 0.048)	0.021 (0.019 to 0.023)	0.040 (0.036 to 0.044)	0.022 (0.019 to 0.025)
Optimism corrected R2	0.059 (0.021 to 0.097)	0.071 (0.035 to 0.107)	0.055 (0.015 to 0.095)	0.068 (0.031 to 0.105)	0.259 (0.207 to 0.311)	0.271 (0.220 to 0.322)	0.256 (0.200 to 0.312)	0.269 (0.216 to 0.322)
Naïve Brier Score (lower better)	0.199 (0.189 to 0.209)	0.200 (0.190 to 0.210)	0.200 (0.189 to 0.211)	0.200 (0.190 to 0.210)	0.165 (0.154 to 0.176)	0.166 (0.155 to 0.177)	0.166 (0.155 to 0.177)	0.167 (0.156 to 0.178)
Optimism Brier Score	−0.006 (−0.006 to −0.006)	−0.003 (−0.003 to −0.003)	−0.006 (−0.007 to −0.005)	−0.003 (−0.004 to −0.002)	−0.007 (−0.008 to −0.006)	−0.004 (−0.004 to −0.004)	−0.007 (−0.008 to −0.006)	−0.004 (−0.005 to −0.003)
Optimism corrected Brier Score	0.205 (0.195 to 0.215)	0.203 (0.193 to 0.213)	0.206 (0.195 to 0.217)	0.204 (0.194 to 0.214)	0.172 (0.161 to 0.183)	0.170 (0.159 to 0.181)	0.173 (0.162 to 0.184)	0.171 (0.160 to 0.182)
Spiegelhalter P value	0.98	0.99	0.96	0.98	0.98	0.99	0.97	0.97

Values in parenthesis are 95% confidence limits estimated as ±1.96 times SD of the bootstrap sample.

^*Recommended model selection.

Receiver operating characteristic curves show that the sensitivity versus specificity were virtually identical for the full and selected models whether we included restricted cubic splines for some numeric variables or used only linear terms for all numeric variables (Figure S2). The calibration plot (Figure 1) and corresponding Spiegelhalter test (P=0.98) indicate that the recommended selected linear model is well calibrated even if only modestly predictive.

Figure 1. Calibration plot (preoperative selected linear predictors).

Use of Preoperative and Immediate Postoperative Candidate Predictors to Model Presence of Organ Dysfunction or Death 48 Hours After Surgery

The performance of the models improved considerably when the immediate postoperative predictors CPB duration and SOFA score at ICU admission were considered (Table 3). The selected model included SOFA score at ICU admission, Clinical Frailty Scale, at nutrition risk, combined procedures, urgent surgery, EuroSCORE II, and duration of CPB when restricted cubic splines were used (Table 4). Age and LVEF were also retained when only linear terms were considered. Again, we found that the performance of the full and selected models were similar (Figure S3), and the linear model performed better than the model with splines after we corrected for optimism (Table 3). Thus, we again recommend the selected model with linear predictors only (Table S4) which yielded an optimism corrected c‐index of 0.773 (95% CI, 0.745–0.801) for predicting persistent organ dysfunction or death 48 hours after surgery (Table 3). Figure 2 and the corresponding Spiegelhalter test (P=0.97) again suggest good calibration.

Table 4.

Model Selection Results (Preoperative and Immediate Postoperative Variables) With Restricted Cubic Splines

Predictor	Predictor form	Df	% of bootstrap samples selecting variable	P value in full model	P value for non linearity in full model	P value in selected model	P value for nonlinearity in selected model	VIF if linear†
Age‡	RCS	4	65%	0.256	0.347			1.20
Sex	Binary	1	47%	0.252				1.11
Charlson Comorbidity Index	Linear	1	21%	0.581				2.33
Sequential Organ Failure Assessment score at ICU admission	Linear*	1*	100%*	<0.001*		<0.0001*		1.14*
Clinical Frailty Scale	Linear*	1*	68%*	0.070*		0.035*		1.12*
At nutrition risk	Binary*	1*	73%*	0.047*		0.038*		1.05*
Combined procedures	Binary*	1*	58%*	0.087*		0.115*		1.05*
Urgent surgery	Binary*	1*	59%*	0.098*		0.058*		1.06*
Modeate‐severe chronic kidney disease	Binary	1	23%	0.573				2.41
Left ventricular ejection fraction‡	RCS	4	78%	0.346	0.723			1.11
EuroSCORE II	RCS on log scale*	4*	96%*	0.016*	0.511*	<0.0001*	0.424*	1.45*
CPB duration	RCS on log scale*	4*	100%*	<0.001*	0.187*	<0.0001*	0.169*	1.20*

CPB indicates cardiopulmonary bypass; DF, degrees of freedom; EuroSCORE II, European System for Cardiac Operative Risk Evaluation II; RCS, restricted cubic spline; and VIF, variance inflation factor.

^*Variables were selected for the model.

^† VIF is 1/(1−R²) where R² is the proportion of variance of the given variable that is explained by all other candidate predictors in a linear regression model where all variables are binary or linear with log transformation for EuroSCORE II and CPB duration. VIF values <5 are generally considered not indicative of high collinearity. VIF estimates are based on single imputation estimated by the expectation–maximization algorithm while all other estimates use multiple imputation.

^‡ This variable was selected when the model did not use restricted cubic splines and was selected in a minority of imputed data sets.

Figure 2. Calibration plot (preoperative and immediate postoperative selected linear predictors).

Figure 3 provides receiver operating characteristic curves comparing the selected linear model based only on presurgical variables to the enhanced model including presurgical as well as immediate postsurgical variables. All model performance measures increased considerably when immediate postsurgical variables were considered (Table 3). Figure 4 provides a histogram of the estimated probability of death or persistent organ dysfunction 48 hours after surgery for the 2 models among patients with and without the event. Among true positives the estimated probability of being positive increased from 36% to 47%, whereas among the true negatives the estimated probability decreased from 29% to 24%.

Figure 3. ROC curve comparing the selected linear models including only presurgical variables and presurgical plus immediately postsurgical variables.

ROC indicates receiver operating characteristics.

Figure 4. Distribution of estimated probability of persistent organ dysfunction or death 48 hours after surgery from the selected linear models, including only presurgical variables and presurgical plus immediately postsurgical variables.

PODS indicates persistent organ dysfunctions.

There were no notable differences in performance when applying the overall selected linear models based on presurgical and pre‐presurgical as well as immediate postsurgical variables to age and sex subgroups (Table S5).

DISCUSSION

The findings of our study demonstrate that integrating preoperative vulnerability markers (Clinical Frailty Scale, at nutrition risk, EuroSCORE II, urgent surgery, combined procedures) with immediate postoperative characteristics (CPB duration, SOFA score at ICU admission) significantly helps to predict the presence of organ dysfunction or death 48 hours after cardiac surgery compared with considering preoperative predictors only.

Current risk stratification models remain incomplete, failing to incorporate immediate postoperative markers such as early SOFA scores and CPB duration that could enable real‐time risk assessment. This gap hinders both trial efficiency (by diluting cohorts with low‐risk patients) and bedside decision‐making (by delaying targeted interventions). ¹⁰ , ²¹

Our proposed solution—integrating both preoperative vulnerabilities and immediate postoperative responses—offers a transformative approach to identify high‐risk patients when interventions matter most. This precision strategy achieves three fundamental advances: (1) it enables predictive enrichment by selecting patients most likely to benefit from specific therapies, (2) it reduces population heterogeneity that obscures treatment effects, and (3) it enhances trial efficiency through optimized sample sizes and validated composite endpoints. ³ , ⁵ By capturing the continuum of risk from preoperative status through surgical stress response, we create new opportunities for both precision medicine and accelerated therapeutic development in cardiac surgery.

Frailty (eg, evaluated by the Clinical Frailty Scale) is of increasing interest in the field of cardiac surgery and has previously been validated as an independent predictor for poor clinical outcomes and mortality in patients undergoing cardiac surgery. ²² , ²³ , ²⁴ A recently published systematic review and meta‐analysis by He et al. ²⁴ including 9 studies showed that frailty was associated with increased in‐hospital, short‐term, and long‐term mortality. In this context, our data confirm that frailty has a strong association with presence of organ dysfunction 48 hours after surgery and, thus, might be a reliable predictor for higher event rates of a complicated and prolonged course of disease.

In the present trial, being at nutrition risk has been confirmed as a reliable predictor of worse patient outcome. Besides a generally known high prevalence of preexisting malnutrition among critically ill patients, including those admitted to the ICU after cardiac surgery procedures, ²⁵ , ²⁶ , ²⁷ there exists increased risk for iatrogenic malnutrition during their stay at the ICU, which is also associated with worse clinical and functional outcomes. ²⁸ The latter is a result of disease‐specific and phase‐specific metabolic, endocrinologic, and inflammatory alterations going along with the onset of critical illness. In this context, assessing nutrition risk early in the hospital/ICU admission process, thus, allows health care providers to implement timely and appropriate nutritional interventions. This proactive approach can help to optimize patient care, improve metabolic function, enhance immune response, and ultimately lead to better clinical outcomes in critically ill patients. ²⁹ , ³⁰

In routine clinical practice, the EuroSCORE II is already implemented to predict perioperative complications in cardiac surgery patients. Several validation studies ³¹ , ³² , ³³ showed reliable prediction of postoperative mortality and morbidity, whereas others showed some predictive inaccuracies of the EuroSCORE II. ³⁴ The latter was also the case for the SUSTAIN CSX trial, which defined EuroSCORE II >5% as an inclusion criterion and, thereby, intended to select patients undergoing cardiac surgery at high risk of a prolonged and complicated postoperative course. Unfortunately, even though the EuroSCORE II >5% was associated with increased CPB and operation times, the patients in the SUSTAIN CSX trial had a rather short ICU length of stay after cardiac surgery, ⁴ indicating that EuroSCORE >5% is no longer a valid cut‐off for estimating high mortality risk in patients after cardiac surgery.

Besides the EuroSCORE II, the Society of Thoracic Surgeons risk score (not available from the underlying data set) represents another tool to predict patients' risk for worse outcomes after cardiac surgery. The Society of Thoracic Surgeons risk score provides predictions for a range of outcomes, including stroke, hospital stay duration, and renal failure, making it a comprehensive tool. However, it is only applicable to specific surgeries, including isolated coronary artery bypass graft, isolated aortic valve replacement, isolated mitral valve replacement, isolated mitral valve repair, as well as coronary artery bypass graft combined with aortic valve replacement or mitral valve surgery. ³⁵ , ³⁶ Considering these limitations, the purpose of the prediction model developed in the present work aims to be applicable to all cardiac surgery procedures.

In general, patients in need of urgent surgery are exposed to high perioperative stress, which might increase the risk for worse postoperative outcomes. ³⁷ And, even if the advantage of performing combined procedures is that multiple operations can be avoided, there is still a risk of postoperative complications as the body is exposed to high levels of stress during a prolonged surgery duration. ¹² Even though CPB is an essential component of most cardiac surgical procedures, prolonged CPB time is often a surrogate marker for surgical complexity. The duration of CPB directly correlates with the intensity of the systemic inflammatory response syndrome triggered by ischemia/reperfusion and the extracorporeal circuit. As CPB time increases, the risk of organ‐specific complications, such as acute kidney injury, respiratory failure, and neurologic impairment, rises significantly. Prolonged CPB time has been consistently associated with an increased risk of adverse postoperative outcomes, including presence of organ dysfunction, prolonged mechanical ventilation, infection, and increased mortality. ³⁸ In this context, early identification of patients at risk for extended bypass duration could allow for preemptive optimization and tailored intra‐ and postoperative management.

Further, our findings confirm that the SOFA score is a reliable prognostic tool, as it correlates with increased risk for a prolonged and complicated postoperative course, for example, characterized by presence of organ dysfunction or death. ³⁹ , ⁴⁰ , ⁴¹ , ⁴² A higher SOFA score at ICU admission is associated with an increased risk of death, making it a valuable metric for clinicians in the intensive care setting. Thus, implementation of the SOFA score as diagnostic and prognostic tool in clinical practice can enhance decision‐making processes and improve patient management in critical care environments.

Risk prediction models analyzing the association of defined patient characteristic variables with single outcome parameters such as ICU length of stay and mortality, respectively, have been evaluated manifold in recent clinical studies. ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ Here, pre‐ and immediate postoperative variables from the SUSTAIN CSX data set have been used to develop a pragmatic predictive model for presence of organ dysfunction or death 48 hours after surgery—a tool designed to:

1. enrich clinical trials by identifying patients most likely to benefit from investigational therapies,

2. guide early ICU care by flagging high‐risk patients for intensified monitoring, and

3. complement existing scores (eg, EuroSCORE) with dynamic, organ‐specific insights.

This approach aligns with successful “treatable traits” paradigms in sepsis and critical care, ⁶ , ²¹ offering a more tailored strategy to address cardiac surgery's trial challenges: low event rates, heterogeneous populations, and rising trial costs. ⁴⁷

However, this study has several limitations. Age is also a variable in EuroSCORE II; however, the variance inflation factor for all the variables is always <1.2, and therefore, collinearity is not an issue in this analysis. Further, because our second model uses pre‐ and immediate postoperative variables, it can be used only for trials that enroll/randomize patients 48 hours after cardiac surgery. Also, LVEF was missing for 15% of the patients, though there were minimal missing data among the other predictors. Finally, this preliminary model should be independently validated in future prospective studies.

Nevertheless, the current analysis shows potential generalizability, especially among developed countries with considerable medical advancement, as the data are derived from a multicenter clinical trial with large sample sizes in these countries.

CONCLUSIONS

This post hoc analysis of a multicenter clinical trial identifies a high‐risk subpopulation using actionable, routinely available variables, bridging a critical gap between trial efficiency and personalized care. The proposed model suggests that a combination of preoperative and immediate postoperative patient characteristics—including SOFA score, Clinical Frailty Scale, nutrition risk, presence of combined procedures, urgency of surgery, EuroSCORE II, and CPB time—may help predict the risk of presence of organ dysfunction or death following cardiac surgery. A simplified model including only linear terms performed as well as a more complex model including restricted cubic splines for some numeric predictors. Incorporating these variables into clinical decision‐making and the design of future trials could enable identification of high‐risk patients who may benefit most from targeted interventions rather than considering preoperative characteristics only.

Sources of Funding

The SUSTAIN CSX trial has been funded by the Lotte and John Hecht Memorial Foundation, biosyn Arzneimittel GmbH (Fellbach, Germany), and the Canadian Institute of Health Research, who had no role in designing the protocol, conducting the trial, or analyzing the data. The present post hoc analysis, however, did not receive any external funding.

Disclosures

None.

Supporting information

Tables S1–S5

Figures S1–S3

Data S1: TRIOPDAI Checklist