Design setting

This retrospective observational study received approval from the Cleveland Clinic Institutional Review Board (IRB #20-390). All procedures were conducted in accordance with the ethical standards of the responsible institutional and/or regional review committees and with the principles of the 1975 Declaration of Helsinki.

Participants

Children ≤ 18 years of age who were admitted to the paediatric cardiac intensive care unit at Cleveland Clinic Children’s after cardiothoracic surgery for CHD between February 2023 and November 2024 were enrolled in this study. Patients older than 18 years were excluded from study. Patients were excluded only if required data elements could not be retrieved from the electronic medical record and/or the Paediatric Cardiac Critical Care Consortium registry for model development. Analyses were performed using complete cases after excluding patients with non-retrievable data; no imputation was performed. Reporting followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guideline (Supplementary Table 1). ^{Reference von Elm, Altman and Egger18}

Data source and measurements

Clinical and perioperative data were obtained retrospectively from the institutional electronic medical record (EPIC Systems, Verona, WI) and the Paediatric Cardiac Critical Care Consortium registry database. Laboratory values, haemodynamic parameters, and echocardiographic findings were extracted from routinely recorded clinical documentation. Data accuracy was verified by two independent investigators through cross-checking random samples of entries against the source charts.

Variables

Patient characteristics included chronological age, operative weight, prematurity, documented genetic anomalies, single-ventricle physiology, the presence of a bidirectional shunt on echocardiography, preoperative endotracheal intubation, and preoperative vasoactive/inotropic support; together, these variables reflected preoperative physiological status.

In the operative and postoperative periods, procedural complexity was classified as high risk for the Risk Adjustment for Congenital Heart Surgery (RACHS-1) scores >3 and Society of Thoracic Surgeons-European Association for Cardio-Thoracic Surgery (STAT) categories >2. Procedure-specific metrics comprised total cardiopulmonary bypass duration, aortic cross-clamp time, and circulatory-arrest time, as well as the number of prior cardiac operations and whether a heart transplant was performed. Intraoperative transesophageal or epicardial echocardiography findings of residual shunt or lesion and the decision to leave the sternum open postoperatively (delayed sternal closure) were also recorded. Variables for the multivariable analysis were selected based on the univariate results.

In this study, low cardiac output syndrome was defined according to the Paediatric Cardiac Critical Care Consortium registry criteria: ^{Reference Tandon, Bhattacharya and Morca1,Reference Gaies, Pasquali and Donohue5} (1) a vasoactive-inotropic score >15 at any time; (2) a threefold increase in vasoactive-inotropic score within any 48-hour interval, including postoperative escalation of support, with the resultant vasoactive-inotropic score ≥10; (3) an arteriovenous oxygen (AVO₂) difference >40% measured invasively or by near-infrared spectroscopy; or (4) documentation of low cardiac output syndrome in a physician note. Paediatric Cardiac Critical Care Consortium prospectively records the earliest qualifying time point for each case.

Statistical analysis

All eligible patients meeting inclusion criteria during the study period were included using a consecutive sampling approach to minimise selection bias. Continuous variables are summarised as mean ± SD or median (IQR) according to distribution, and categorical variables as counts. Normality was assessed using the Shapiro–Wilk test (α = 0.05); the null hypothesis of normality was retained when p > 0.05. Group comparisons were performed with the χ² or Fisher exact test for categorical data, the Student t-test for normally distributed continuous data, and the Wilcoxon rank-sum test for non-normal data. Propensity-score distributions between low cardiac output syndrome and non-low cardiac output syndrome were compared using a shift plot and the Mann–Whitney U test. The shift plot computes a shift function for two (in)dependent groups using the robust Harrell–Davis quantile estimator with bias-corrected bootstrap confidence intervals. ^{Reference Rousselet and Pernet19} Univariate screening for candidate predictors of low cardiac output syndrome employed point-biserial, tetrachoric, polychoric, or Spearman correlations, with variables retained for multivariable modelling when p < 0.10 and each category represented ≥5 % of the cohort. Independent associations with low cardiac output syndrome were then examined with a multivariable logistic-regression model; non-significant terms (p > 0.05) were sequentially removed by backward elimination. Model discrimination is reported with adjusted pseudo-R ² (AOR) and odds ratios with 95 % confidence intervals. All tests were two-sided, and a p-value < 0.05 denoted statistical significance. Analyses were performed in Stata 17 (StataCorp, College Station, TX, and SAS 9.3©). Multicollinearity among candidate predictors was assessed via pairwise Spearman correlations and variance inflation factors; all variance inflation factors for predictors retained in the trimmed model were below 2.5 (Supplementary Table S2).

Concordance with prior literature and novelty

In prior paediatric series, low cardiac output syndrome has been observed in nearly half of patients following congenital heart surgery and remains the leading cause of postoperative mortality in children. ^{Reference Song, Dang and Dong11} A key contribution of our study is endpoint standardisation: low cardiac output syndrome was defined using Paediatric Cardiac Critical Care Consortium registry criteria, which improves comparability across studies and facilitates adoption in Paediatric Cardiac Critical Care Consortium-aligned settings. Our cohort’s 24% incidence is broadly comparable and underscores the clinical burden. Importantly, our model is intended for use at OR exit/early paediatric cardiac intensive care unit arrival—before low cardiac output syndrome is typically recognised at the bedside—thereby supporting early triage and resource planning. In practice, higher-risk strata can prompt earlier escalation of surveillance (e.g., higher-frequency haemodynamic reassessment, earlier lactate/echo review, and consideration of advanced monitoring) and inform bedside staffing intensity during the first postoperative hours. Prior published work identified risk factors of low cardiac output syndrome, including preoperative haemodynamic reserve and surgical history, including inotrope requirement, intubated status, and previous operations, consistent with our findings that preoperative inotropes and previous surgeries track with higher low-cardiac-output-syndrome risk. ^{Reference Du, Chen and Song12,Reference Wang, Fu and Bai13,Reference Drennan, Burge and Szyld15} While these predictors are not novel in direction, we translate them into a parsimonious and interpretable risk estimate using variables that are routinely available at OR exit (prior operations, operative weight, preoperative inotrope use, RACHS category, and CPB time), enabling automated calculation at paediatric cardiac intensive care unit arrival without additional testing and supporting straightforward EHR implementation. In infant-focused studies, longer aortic cross-clamp time and lower body weight are independent predictors. ^{Reference Zou, Yu and Wang20} In our model, higher operative weight was protective, and a time-related intraoperative load (longer CPB) conferred incremental risk, aligning with that pattern. These observations complement evidence that low cardiac output syndrome severity captured at the bedside (e.g., low cardiac output syndrome score) relates strongly to morbidity and paediatric cardiac intensive care unit resource utilisation, ^{Reference Ulate, Yanay and Jeffries21} and they reinforce the value of combining preoperative/intraoperative features with early postoperative dynamics in risk stratification. To enhance clinical usability, we reported triage-relevant performance metrics, and our distributional analysis (Figure 1) further clarifies rule-out utility alongside residual risk—motivating future dynamic updating with early postoperative physiology rather than implying low cardiac output syndrome is fully captured by static factors. Notably, postoperative multimodal physiology is not required for the model to be clinically useful at paediatric cardiac intensive care unit arrival; rather, it represents a complementary second layer for dynamic updating of residual risk. This rationale is supported by our prior neonatal work, where supervised ML models incorporating hourly postoperative clinical and laboratory data during the first 48 postoperative hours achieved strong forecasting performance for low cardiac output syndrome across short-term horizons, illustrating the incremental value of time-varying physiology beyond static predictors. ^{Reference Baloglu, Wang and Marino22} The concentration of postoperative mechanical circulatory support within the low cardiac output syndrome group provides external face validity for the Paediatric Cardiac Critical Care Consortium low cardiac output syndrome phenotype as a marker of severe circulatory failure in our cohort, despite the therapy-dependent nature of its underlying definition.

Preoperative profile

Our finding that preoperative inotrope requirement and preoperative endotracheal intubation independently predict low cardiac output syndrome is consistent with the broader literature that treats preoperative instability as a bedside proxy for limited circulatory reserve and heightened postoperative risk. ^{Reference Du, Chen and Song12,Reference Wang, Fu and Bai13,Reference Drennan, Burge and Szyld15} In large paediatric cohorts and meta-analyses, younger age and lower operative weight repeatedly emerge as core risk factors, ^{Reference Du, Chen and Song12,Reference Wang, Fu and Bai13} which matches our model’s observation that higher weight is protective. The presence of syndromes highlights compounding pulmonary comorbidities that can worsen haemodynamics and complicate management. ^{Reference Delany, Gaydos and Romeo23,Reference Yamaguchi, Tomonari and Utsunomiya24} Overall, our findings are consistent with prior paediatric reports, reinforcing their relevance to contemporary clinical experience. ^{Reference Du, Chen and Song12,Reference Wang, Fu and Bai13,Reference Drennan, Burge and Szyld15,Reference Delany, Gaydos and Romeo23}

Intra and postoperative profile

Surgical complexity, captured by longer CPB duration and higher RACHS/STAT, likely amplifies circulatory stress and the inflammatory cascade, predisposing to low cardiac output syndrome; these associations are consistently documented across paediatric studies and a recent meta-analysis. ^{Reference Du, Chen and Song12,Reference Wang, Fu and Bai13} The classic postoperative time course described by Wernovsky (cardiac index nadir ∼9–12 h after surgery) contextualises why low cardiac output syndrome often declares itself hours after the OR when reperfusion injury and vasoplegia peak. ^{Reference Wernovsky, Wypij and Jonas3} The observed temporal concentration of low cardiac output syndrome onset aligns closely with this trajectory and supports the use of preoperative and intraoperative variables as proximate predictors for the events captured in our cohort. Targeted pharmacologic prevention has biological plausibility but risk remains concentrated in high-complexity cases. ^{Reference Hoffman, Wernovsky and Atz14} Markers of severity used in practice, such as low cardiac output syndrome score and VIS, track morbidity and resource utilisation, ^{Reference Ulate, Yanay and Jeffries21} whereas certain single physiologic readouts may be insufficiently specific for new organ injury. ^{Reference Claessens, Jansen and Breur17} Consistent with our cohort (higher “open sternum” rates in low cardiac output syndrome), delayed sternal closure is typically deployed for haemodynamic instability and is associated with greater morbidity. ^{Reference Hurtado-Sierra, Calderón-Colmenero and Curi-Curi16} Beyond the OR, fluid overload is a powerful and modifiable risk factor linked to mortality/morbidity in paediatric cardiac surgery. ^{Reference Lex, Tóth and Czobor25} Looking ahead, converging evidence argues for continuous, multimodal monitoring and analytics-assisted early warning: non-invasive cardiac output monitoring in CHD and emerging machine-learning models illustrate how integrating pre/intra/postoperative signals (including lactate and blood-pressure dynamics) can improve discrimination and timeliness of low cardiac output syndrome detection. ^{Reference Tandon, Bhattacharya and Morca1,Reference Hong, Xu and Ge26,Reference Abbasi, Li and Dekle27}

Interpreting physiological markers

Interpreting physiological markers for impending low cardiac output syndrome remains difficult because there is still no uniform definition and single variables perform inconsistently across settings. ^{Reference Alten and Gaies28} Recent reviews emphasise the ongoing lack of consensus definitions, complicating comparisons and surveillance strategies, and argue for multimodal assessment instead of reliance on any single metric. ^{Reference Schoonen, van Klei, van Wolfswinkel and van Loon10} In practice, scores derived from vasoactive support correlate with adverse outcomes after paediatric cardiac surgery, but they are therapy-dependent surrogates rather than direct measures of oxygen delivery. ^{Reference Butts, Scheurer and Atz4,Reference Alten and Gaies28,Reference Gaies, Jeffries and Niebler29,Reference Koponen, Karttunen and Musialowicz30} Ulate and colleagues proposed the low cardiac output syndrome score to fold clinical signs (e.g., tachycardia, oliguria, toe temperature, near-infrared spectroscopy, and lactate) together with pharmacologic support; in their infant cohort, cumulative low cardiac output syndrome score discriminated composite morbidity well (AUC ≈ 0.83), yet the construct still leaned heavily on treatment intensity and lacked concurrent validation against direct cardiac output or oxygen-delivery measurements. ^{Reference Ulate, Yanay and Jeffries21} For example, postoperative cerebral near-infrared spectroscopy, while useful for trend monitoring, has not consistently predicted adverse outcomes, underscoring the limitations of relying on any single physiologic marker for low cardiac output syndrome risk stratification. ^{Reference Claessens, Jansen and Breur17} Taken together, these data support our approach: multivariable, integrative prediction that combines preoperative and intraoperative risk factors with early postoperative physiology offers a more robust path to actionable low cardiac output syndrome risk stratification than reliance on any single marker alone. ^{Reference Schoonen, van Klei, van Wolfswinkel and van Loon10}

Building on this rationale, integrated real-time analytics can fuse multiple physiological and laboratory streams to estimate risk states (e.g., inadequate oxygen delivery) continuously. Model-based indices computed every few seconds from routinely available inputs have been shown to track the probability of dangerously low venous saturation and to correlate with lactate and adverse events; external validations report strong discrimination for detecting low SvO₂ even when some inputs are missing. ^{Reference Abbas, Hussain and Shahbaz31,Reference Holman, Baronov and McMurray32} In parallel, paediatrics has seen a broader shift from single-signal monitoring to proactive risk surveillance: collaborative, low-tech prevention bundles in cardiac ICUs achieved ∼30% reductions in risk-adjusted in-hospital cardiac arrest that were sustained across centres, while machine learning models using routine electronic health record data can flag arrest risk hours in advance. ^{Reference Alten, Cooper and Klugman33–Reference Kenet, Pemmaraju and Ghate35} Taking these together, this evidence supports our rationale for multivariable prediction: combining preoperative and intraoperative risk factors with early postoperative physiology is more likely to surface actionable low cardiac output syndrome risk than any single marker alone. ^{Reference Singh, Villaescusa and da Cruz36}

Propensity-score distributions and clinical implications

Using the trimmed model’s predicted probabilities as propensity scores, we observed a left-skewed distribution among non-low-cardiac-output-syndrome patients, whereas the low-cardiac-output-syndrome distribution did not deviate from normality (Shapiro–Wilk p = 0.14). As shown in Figure 1 (top panel), non-low cardiac output syndrome patients cluster near the lowest propensity scores, whereas low cardiac output syndrome patients span a broader and generally higher range of scores. Quantile-wise contrasts via a shift function confirmed systematic between-group differences (Mann–Whitney U p < 0.001; see Figure 1). In Figure 1, the red connectors link decile-matched quantiles between groups, and the lower panel (shift function) summarises low cardiac output syndrome–no low cardiac output syndrome differences across the score distribution with confidence intervals, illustrating that separation persists across quantiles and is most pronounced toward higher quantiles. Clinically, a heavy concentration of very low scores in non-low cardiac output syndrome supports strong rule-out performance (reflected in our high negative predictive value), while the presence of low cardiac output syndrome cases even at lower propensities highlights residual risk that static pre/intra-operative models cannot fully capture, consistent with low cardiac output syndrome as a multifactorial, time-evolving syndrome and with meta-analytic evidence that risk reflects disease severity plus intra- and postoperative factors. ^{Reference Wang, Fu and Bai13,Reference Alten, Cooper and Blinder37} This overlap—low cardiac output syndrome events occurring despite lower propensities—helps reconcile strong rule-out performance with incomplete separation between groups. These patterns argue for dynamic updating with early postoperative physiology. Contemporary, model-based indices that continuously estimate the probability of inadequate oxygen delivery (e.g., probability that mixed/central venous saturation falls below a threshold) have shown correlation with measured venous saturations and provide second-to-second risk signals suitable for integration with pre-/intra-operative features. ^{Reference Ulate, Yanay and Jeffries21,Reference Loomba, Villarreal and Flores38} In parallel, system-level prevention work demonstrates that proactive surveillance frameworks can reduce serious events in cardiac ICUs, underscoring the value of moving from single-marker monitoring to integrated, real-time risk surveillance. ^{Reference Alten, Cooper and Klugman33}

Limitations

This study has several limitations. First, it represents a single-centre, retrospective analysis, which may limit generalisability and introduce spectrum or case-mix effects. Second, although the vasoactive-inotropic score and physician documentation are widely accepted components of the Paediatric Cardiac Critical Care Consortium low cardiac output syndrome definition, these measures reflect therapy delivered rather than therapy needed; institutional choice of inotropic agent, threshold for escalation, and timing of weaning vary across centres and providers and may introduce phenotype misclassification in either direction. The model presented here is therefore calibrated to current practice patterns at our institution and may require recalibration prior to deployment in centres with different vasoactive thresholds or weaning protocols. Third, although we included a broad range of preoperative and intraoperative variables, by design, postoperative physiological streams (e.g., continuous haemodynamics and lactate kinetics) and direct measures of cardiac output were not incorporated, restricting the model’s ability to capture dynamic recalibration. Finally, the modest sample size and number of low cardiac output syndrome events may constrain statistical power, particularly for multivariable modelling, and external validation in larger, multicentre cohorts will be required to confirm generalisability and calibration.

Applications and future work

Our findings support clinical use of the low cardiac output syndrome risk–scoring model for (i) rapid risk stratification at OR exit/early paediatric cardiac intensive care unit arrival, (ii) resource planning in higher-risk cases (e.g., closer nurse/physician surveillance and earlier access to advanced monitoring), and (iii) triggering haemodynamic-optimisation bundles so that patients likely to develop low cardiac output syndrome receive earlier support. Looking ahead, generalisability and time-aware recalibration are essential. We recommend (1) multicentre, prospective external validation with predefined decision thresholds; and (2) dynamic updating of the pre- and intraoperative model using continuous postoperative high-fidelity physiologic data (arterial waveform features, ECG, urine output, and lactate kinetics). ^{Reference Loomba, Sourour and Flores39} This integration is feasible and clinically relevant: single postoperative signals can be insufficient on their own (e.g., cerebral near-infrared spectroscopy not independently associated with new brain injury in infants with critical CHD), whereas model-based, real-time indices that estimate the probability of inadequate oxygen delivery show good discrimination and correlate with venous saturations. In parallel, non-invasive cardiac-output monitoring in CHD continues to broaden the physiologic data available for continuous risk recalibration. As part of an ongoing prospective initiative within the Cleveland Clinic Children’s Center for Artificial Intelligence (C4AI), the model is being prepared for silent deployment within the institutional electronic health record system, with prospective evaluation of calibration and temporal drift; subsequent integration as clinical decision support, in which exceedance of a prespecified high-risk threshold would trigger a tiered surveillance bundle; and a quality improvement evaluation comparing pre- and post-implementation periods planned. Each stage of this stepwise approach is contingent on formal prospective validation and institutional ethics approval. Subsequent iterations of the model will incorporate direct haemodynamic indices (e.g., cardiac index and central venous oxygen saturation) and harmonised multicentre low cardiac output syndrome phenotype definitions in order to reduce reliance on therapy-driven surrogates.