Derivation and Validation of a Nomogram to Predict In-Hospital Complications in Children with Tetralogy of Fallot Repaired at an Older Age.

Background We aimed to develop and validate a prediction model for in-hospital complications in children with tetralogy of Fallot repaired at an older age. Methods and Results A total of 513 pediatric patients from the Tianjin data set formed a derivation cohort, and 158 pediatric patients from the Hefei and Xiamen data sets formed validation cohorts. We applied least absolute shrinkage and selection operator analysis for variable selection and logistic regression coefficients for risk scoring. We classified patients into different risk categorizations by threshold analysis and investigated the association with in-hospital complications using logistic regression. In-hospital complications were defined as death, need for extensive pharmacologic support (vasoactive-inotrope score of ≥20), and need for mechanical circulatory support. We developed a nomogram based on risk classifier and independent baseline variables using a multivariable logistic model. Based on risk scores weighted by 11 preoperative and 4 intraoperative selected variables, we classified patients as low, intermediate, and high risk in the derivation cohort. With reference to the low-risk group, the intermediate- and high-risk groups conferred significantly higher in-hospital complication risks (adjusted odds ratio: 2.721 [95% CI, 1.267-5.841], P=0.0102; 9.297 [95% CI, 4.601-18.786], P<0.0001). A nomogram integrating the ARIAR-Risk classifier (absolute and relative low risk, intermediate risk, and aggressive and refractory high risk) with age and mean blood pressure showed good discrimination and goodness-of-fit for derivation (area under the receiver operating characteristic curve: 0.785 [95% CI, 0.731-0.839]; Hosmer-Lemeshow test, P=0.544) and external validation (area under the receiver operating characteristic curve: 0.759 [95% CI, 0.636-0.881]; Hosmer-Lemeshow test, P=0.508). Conclusions A risk-classifier-oriented nomogram is a reliable prediction model for in-hospital complications in children with tetralogy of Fallot repaired at an older age, and strengthens risk/benefit-based decision-making.

Clinical Perspective

What Is New?

We developed a nomogram model integrating the ARIAR‐Risk classifier (absolute and relative low risk, intermediate risk, and aggressive and refractory high risk) with age at surgery and mean blood pressure to predict in‐hospital complications after tetralogy of Fallot repair and further validated models in both derivation and validation cohorts, suggesting good discrimination and goodness‐of‐fit.

What Are the Clinical Implications?

A risk‐classifier–oriented nomogram is a reliable prognostic tool for the assessment of in‐hospital complications in children with tetralogy of Fallot repaired at an older age and is easy to implement in clinical practice.

Introduction

Tetralogy of Fallot is the most common cyanotic congenital heart defect globally and carries high morbidity and mortality risk if not treated immediately and properly.1, 2, 3 However, infants and children who have undergone surgical repair with cardiopulmonary bypass (CPB) are at high risk for significant morbidity and mortality.4, 5, 6 Although these potential adverse outcomes have been well described, a critical need remains to develop a predictive model by integrating baseline, clinical, and procedural factors that are indicative of illness severity and short‐term outcome. The identification of these variables could aid in appropriate risk stratification, monitoring, and clinical management for the treatment of tetralogy of Fallot.7, 8

Various models have been developed to predict postoperative mortality and morbidity in cardiac surgery.9, 10 In contrast, no predictive models combine baseline, clinical, and procedural factors to predict poor outcomes after tetralogy of Fallot repair, given the heterogeneities of congenital heart diseases.11, 12

We performed a study incorporating pre‐ and intraoperative characteristics with the aim of identifying and validating a risk classifier that predicts in‐hospital complications in Chinese children with tetralogy of Fallot repaired at an older age defined as over 6 months‐1 year. Moreover, we integrated risk‐classifier and independent baseline predictors to generate a nomogram, backed by internal and external validation, to advance clinical evaluation for patient therapeutics and to strengthen risk/benefit–based decision‐making.

Methods

Data Availability

The data, analytic methods, and study materials will not be made available to other researchers for purposes of reproducing the results or replicating the procedure. Because this study uses data from human subjects, the data and everything pertaining to them are governed by the TedaICH Data Protection Agency and can only be made available to additional researchers if a formal request is filed with the TedaICH authorities.

Study Design and Participants

Between January 1, 2012, and July 31, 2018, 513 consecutive pediatric patients with tetralogy of Fallot at Teda International Cardiovascular Hospital (Tianjin, China) formed the derivation cohort. We used an independent data set of 158 pediatric patients (validation cohort) from Anhui Provincial Hospital (Hefei, China) between January 1, 2004, to July 31, 2018, and Xiamen Cardiovascular Hospital (Xiamen, China) between January 1, 2012, to July 31, 2018, to externally validate this model (Figure S1). We included pediatric patients aged 10 days to 18 years who underwent complete repair of tetralogy of Fallot with CPB. Both derivation and validation cohorts included only tetralogy of Fallot patients, excluding those with pulmonary atresia with ventricular septal defect or double‐outlet right ventricle. Ethics and regulatory approval for the study was obtained from each local ethics committee. Written informed consent was obtained from each patient before surgery.

Candidate Predictors

Consistent data for each patient were collected from the medical records, and all candidate predictors were selected on the basis of detailed literature reviews and clinical evidence within the confines of data availability. Baseline characteristics included continuous and categorized age at surgery,13 sex, weight, height, body mass index, heart rate, respiratory rate, systolic and diastolic blood pressure, blood pressure difference, and mean arterial pressure. The clinical profiles included Tet spell history, systemic arterial saturation, cyanosis degree, right bundle‐branch block, New York Heart Association (NYHA) class (modified Ross scoring for infants14), and hematocrit. Anatomic profiles included ventricular septal defect subtypes, defect scale (the ratio of the defect to aortic root diameter), overriding aorta, predominantly interventricular shunting, right ventricular outflow tract (RVOT) pressure gradient, RVOT obstruction level, McGoon index, left ventricular (LV) ejection fraction, indexed left atrial diameter, indexed right atrial diameter, indexed LV end‐diastolic diameter, indexed RV end‐diastolic diameter, indexed LV end‐diastolic volume, collateral arteries, and patent ductus arteriosus. Surgical profiles included repair approach, RVOT obstruction repair, transannular patch, pulmonary patch, and tricuspid valve detachment. Extracorporeal profiles included cannulation approach, reoxygenation level, cardioplegia, and CPB temperature and duration. These detailed and specific definitions are listed in Table S1.

Study Outcomes

The primary clinical end point was in‐hospital complications, including death, need for extensive pharmacologic cardiovascular support (highest vasoactive‐inotropic score of >20 points), and need for additional mechanical circulatory support within the first 72 hours after operation, whichever occurred first. Vasoactive‐inotrope score was calculated daily as per Gaies et al,15 where vasoactive‐inotrope score=dopamine dose (lg/kg per minute)+dobutamine dose (lg/kg per 9 minute)+[100 × epinephrine dose (lg/kg per minute)]+[10 x milrinone dose (lg/kg per minute)]+[10 000 × vasopressin dose (U/kg per minute)]+[100 × norepinephrine dose (lg/kg per minute)]. All outcomes were adjudicated independently by an event collaborative team.

Statistical Analysis

Before data analysis, predictor variables in the derivation and validation cohorts were inspected for missing values. Among the predictors, the proportion of missing data ranged from 0 to 31.7%. To include these data from the analyses, we imputed missing data by multiple imputations by chained equations, using the mice package for R, in which predictive mean matching is embedded with the cases (k)=5 default. Patients with missing outcome measures and lost demographic and surgical records were excluded from both the derivation cohort (28/573, 4.9%) and the validation cohort (47/262, 17.9%; Figure S1).

Data are presented as frequencies (percentages) for categorical variables and medians (interquartile ranges [IQRs]) for continuous variables. Differences between groups were assessed using the χ2 test or Fisher exact test for categorical variables and the Student t test or the Mann–Whitney U test for continuous variables. The pre‐ and postimputation data sets were compared with the Kruskal–Wallis equality‐of‐populations rank test for the derivation and validation cohorts.

Model derivation was performed according to Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidance.16 We included a set of predefined prediction variables of preoperative variables comprising clinical and anatomical profiles and intraoperative variables comprising surgical and extracorporeal profiles (Tables S1 and S2). We applied least absolute shrinkage and selection operator (LASSO) analysis in a penalized logistic regression model (R package glmnet) to select the most useful prediction variables from all pre‐ and intraoperative candidates in the derivation cohort.17 Subsequently, we constructed the prediction scoring model by assigning each patient a risk score for in‐hospital complications based on the product of the expression levels for the variables selected by the LASSO analysis and the respective regression coefficients weighted by logistic regression analysis in the derivation cohort. Then we fitted the dose–response relationship between the risk score and in‐hospital complications–related morbidity using generalized additive models and further found the optimal cutoff point using Empowerstats software (X&Y Solutions). The thresholds for the scores that were output from the predictive model that was used to classify patients into different risk categories were defined as the scores that gave the largest log‐likelihood value in a 2‐piecewise regression model.18 The Dunnett method was used for multiple comparisons of in‐hospital complication rates against a control group (lowest risk category).

We further assessed the association of risk categorizations with in‐hospital complications using logistic regression for baseline characteristics. We also performed tests for linear trend by entering the median value of each category of risk score as a continuous variable in this model. We developed a in‐hospital complications nomogram starting from a multivariable logistic regression model that allowed us to obtain in‐hospital complications probability estimates. We included a set of predefined predictor variables of the risk‐classifier and baseline characteristics (age group at surgery, sex, weight, height, body mass index, heart rate, respiratory rate, systolic blood pressure, diastolic blood pressure, pulse pressure difference, and mean arterial pressure)19 and applied a stepwise procedure based on the Akaike information criterion (AIC) for variable selection and a multifractional polynomial for variable selection.20

We carried out internal validation of the model development process using a bootstrap resampling process (1000 bootstrap samples per model) to provide an unbiased estimate of model performance. To assess the external validity of model performance, using an independent and external data set of 158 pediatric patients (Hefei and Xiamen cohorts), we examined the overall accuracy (Nagelkerke's R 2), calibration (calibration plots and Hosmer–Lemeshow calibration test) and discrimination (area under receiver operating characteristic curve [AUC]).21, 22, 23 We also plotted decision curves to assess the net benefit of nomogram‐assisted decisions.24 To investigate the cumulative incidence of in‐hospital complications by age at surgery, we used Kaplan–Meier estimates with age as the time scale.25

To investigate whether the predictive strength of nomogram covariates would change by varying the year of surgery, we extracted the linear predictor from the nomogram and investigated the interaction between such linear predictors and the year of surgery with the Wald test.26 We considered a 2‐sided P value of <0.05 to be statistically significant. We performed the statistical analyses using Stata v14 (StataCorp) and R software (v3.2.0; R Foundation for Statistical Computing).

Results

The derivation cohort consisted of 513 patients who had complete surgery, with a median age of 30.7 months (IQR: 13.5–60.8 months) and 221 (43.1%) female patients. The validation cohort consisted of 158 patients, with a median age of 28.7 months (IQR: 13.4–55.1 months) and 73 (46.2%) female patients. The occurrence of in‐hospital complications was 14.8% (76/513) in the derivation cohort and 13.9% (22/158) in the validation cohort, without a significant difference between cohorts (P=0.782).

Baseline clinical and surgical characteristics after imputation in the derivation and validation cohorts are listed in Table 1 and Tables S2 and S3. Based on LASSO analysis, we identified a composite panel that consisted of 11 preoperative and 4 intraoperative variables associated with in‐hospital complication morbidity in the derivation cohort, with the optimal λ penalty (AUC=0.785; Table 1; Figure 1). A risk score was calculated for each patient using a formula derived from the expression levels of these 15 variables weighted by their regression coefficients:Risk score=−2.88745+(−0.15197×cyanosis: minor)+(0.36081×cyanosis: major)+(−0.13475×Tet spell history: absence)+(−0.53315×NYHA class: grades I–II)+(−0.23192×right bundle‐branch block: absence)+(0.02408×prehematocrit: %)+(−0.00824×indexed LV end‐diastolic volume: mm/m2)+(0.00357×indexed right atrial diameter: mm/m2)+(−0.06019×interventricular shunting: bidirectional)+(−0.17454×overriding aorta: ≤50%)+(−0.13612×McGoon index: >1.5)+(0.04306×collaterals: major)+(−0.25836×RVOT obstruction repair: infundibular outflow patch)+(−0.00103×tricuspid valve detachment: absence)+(−0.25053×reoxygenation level: lower)+(0.00769×CPB duration: minute).

Table 1

Selected Baseline and Perioperative Variables of Development and Validation Cohorts After Imputation

Variables	Derivation Cohort (n=513)	Validation Cohort (n=158)	P Value
Baseline variables
Age at surgery, mo	30.7 (13.5–60.8)	28.7 (13.4–55.1)	0.664
Age group at surgery, %			0.429
Infant (<1 y)	216 (42.1)	66 (41.8)
Toddler and preschool (2–5 y)	190 (37.0)	64 (40.5)
School‐aged child (6–12 y)	92 (17.9)	21 (13.3)
Adolescent (13–18 y)	15 (2.9)	7 (4.4)
Mean BP, mm Hg	47.33 (42.67–50.67)	46.8 (42.1–49.3)	0.484
Preoperative variables
Hematocrit, %	43.6 (38.1–50.5)	43.2 (38.1–49.5)	0.797
Indexed LVEDV, mL/m2	27.1 (20.7–35.7)	27.5 (21.9–37.3)	0.364
Indexed RA diameter, mm/m2	49.7 (40.9–56.2)	49.1 (39.5–55.6)	0.593
Cyanosis degree, %			0.902
None	137 (26.7)	45 (28.5)
Minor	174 (33.9)	53 (33.5)
Major	202 (39.4)	60 (38.0)
Tet spell history, %			0.932
Absence	300 (58.5)	93 (58.9)
Presence	213 (41.5)	65 (41.1)
NYHA functional class, %			0.575
I–II	479 (93.4)	150 (94.94)
III–IV	34 (6.6)	8 (5.1)
Right bundle‐branch block, %			0.567
Absence	463 (90.3)	145 (91.8)
Presence	50 (9.8)	13 (8.2)
Interventricular shunting, %			0.807
Predominantly left to right	58 (11.3)	17 (10.7)
Predominantly bidirectional	432 (84.2)	132 (83.5)
Predominantly right to left	23 (4.5)	9 (5.7)
Aortic overriding, %			0.305
≤50%	438 (85.4)	140 (88.6)
>50%	75 (14.6)	18 (11.4)
McGoon index, %			0.731
>1.5	353 (68.8)	111 (70.3)
≤1.5	160 (31.2)	47 (29.8)
Collateral circulation, %			0.367
Minimal	383 (74.7)	109 (69.0)
Minor	73 (14.2)	27 (17.1)
Major	57 (11.1)	22 (13.9)
Intraoperative variables
Reoxygenation level, %			0.086
Lower (≤250 mm Hg)	262 (51.1)	93 (58.9)
Higher (>250 mm Hg)	251 (48.9)	65 (41.1)
Tricuspid valve detachment, %			0.858
Absence	409 (79.7)	127 (80.4)
Presence	104 (20.3)	31 (19.6)
RVOTO repair options, %			0.821
Parietal muscle resection	23 (4.5)	8 (5.1)
Infundibular outflow patch	76 (14.8)	19 (12.0)
Valve‐sparing repair	130 (25.3)	43 (27.2)
Transannular patch	284 (55.4)	88 (55.7)
CPB duration, min	121 (90–156)	122 (95–154)	0.786
Postoperative outcome
In‐hospital complications, %	76 (14.8)	22 (13.9)	0.782
Postoperative inotropic score, point	9.00 (6.00–13.50)	8.70 (6.00–12.43)	0.256
Risk score, point	−2.097 (−2.558 to 1.581)	−2.158 (−2.647 to 1.647)	0.403

Continuous data are presented as median (interquartile range), and dichotomous data are presented as n (%). BP indicates blood pressure; CPB, cardiopulmonary bypass; LVEDV, left ventricle end‐diastolic volume; NYHA, New York Heart Association; RA, right atrial; RVOTO, right ventricular outflow tract obstruction.

Figure 1

LASSO model profile plots. A, Coefficient profile plots showing how size of the coefficients of preoperative and intraoperative factors shrinks with increasing value of the λ penalty, with the factors and their regression coefficients selected for the model based on the optimal λ for the LASSO model. B, Penalty plot for the LASSO model; color error bars indicate standard error. C, The optimal λ penalty of the LASSO model with a maximum AUC of 0.785. AUC indicates area under the receiver operating characteristic curve; LASSO, least absolute shrinkage and selection operator.

A dose dependency of in‐hospital complications risk was identified for increasing risk score (odds ratio [OR]: 4.243 [95% CI, 2.881–6.247], P<0.0001; Figure 2A, Figure S2A). Patients with in‐hospital complications had a significantly higher risk score than those without in‐hospital complications (median: −1.300 [IQR: −2.059 to 0.944] versus −2.196 [IQR: −2.643 to 1.714], P<0.001; Figure S2B). Based on threshold analysis, we generated 2 cutoff values of −2.1488 and −1.5957 to classify such scores into low risk (−2.149 or fewer points), intermediate risk (−2.149 to −1.596 points), and high risk (−1.596 or more points) for the probability of in‐hospital complications in the derivation cohort. Subsequently, triple‐risk (ie, low, intermediate, high) categories were further subdivided into the following subcategories based on threshold analysis of subgroups (Figure 2A): absolute low risk (−2.816 or fewer points), relative low risk (−2.816 to −2.149), intermediate risk (−2.149 to −1.596), aggressive high risk (−1.596 to −0.813), and refractory high risk (−0.813 or more points) subcategories (ARIAR‐Risk classifier). Before multivariable adjustment, both the triple‐risk classifier and the ARIAR‐Risk classifier showed a strong independent predictive factor for in‐hospital complications in the derivation cohort (Figure S3A and S3B). After multivariable adjustment for baseline characteristics, with reference to the low‐risk group, the intermediate‐ and high‐risk groups conferred significantly higher risk of in‐hospital complications in the derivation cohort (adjusted OR: 2.721 [95% CI, 1.267–5.841], P=0.0102; 9.297 [95% CI, 4.601–18.786), P<0.0001, respectively; Figure 2B). Similarly, with reference to the intermediate‐risk group, the low‐risk groups conferred significantly lower risk of in‐hospital complications (adjusted OR: 0.092 [95% CI, 0.018–0.716] for absolute low risk; 0.486 [95% CI, 0.222–1.066] for relative low risk), and the high‐risk groups conferred a significantly higher risk of in‐hospital complications (adjusted OR: 2.701 [95% CI, 1.373–5.316] for aggressive high risk; 8.442 [95% CI, 3.191–22.336] for refractory high risk; Figure 2C). We performed multiple comparisons for ARIAR‐Risk categories and triple‐risk categories, and the results are summarized in Table S4. The AUC comparison showed that the ARIAR‐Risk classifier exhibited prediction performance superior to the triple‐risk classifier (AUC: 0.753 [95% CI, 0.697–0.809] versus 0.733 [95% CI, 0.675–0.791]; P=0.0014) in predicting in‐hospital complications (Figure 3). Consequently, the ARIAR‐Risk classifier was selected as the major variable instead of the triple‐risk classifier to develop the nomogram model.

Figure 2

Risk‐based classifier profiles in a derivation cohort. A, Dose‐response relationship between risk score and morbidity from in‐hospital complications and threshold points for classifying patients into risk categorizations. B and C, Adjusted association between the triple‐risk (ie, low, intermediate, high) classifier (B) and the ARIAR‐Risk classifier (C) and probability and OR of in‐hospital complications with 95% CIs. ARIAR‐Risk indicates absolute and relative low risk, intermediate risk, and aggressive and refractory high risk; OR, odds ratio.

Figure 3

Comparison of the ARIAR‐Risk classifier and the triple‐risk (ie, low, intermediate, high) classifier. AUC indicates area under the receiver operating characteristic curve. ARIAR‐Risk indicates absolute and relative low risk, intermediate risk, and aggressive and refractory high risk.

In addition to the ARIAR‐Risk classifier, AIC‐based stepwise‐selected variables (patient age group and mean blood pressure) on the multivariable logistic model had a significant effect on in‐hospital complications (all Wald tests P<0.05) in the derivation cohort: age group at surgery (infant [referent]; toddler and preschool, AUC: 0.502 [95% CI, 0.265–0.952]; school age child, AUC: 0.871 [95% CI, 0.413–1.837]; adolescent, AUC: 1.619 [95% CI, 0.417–6.285]) and mean blood pressure (per mm Hg, OR: 0.974 [95% CI, 0.948–1.001]). Based on the ARIAR‐Risk classifier and the independent risk factors selected, we developed a nomogram using a multivariable logistic model to predict the probability of in‐hospital complications after surgery for a patient (Figure 4A) based on the AIC‐selected logistic model: 0.19653−0.68956×(age group: toddler and preschool)−0.13795×(age group: school age child)+0.48166×(age group: adolescent)−0.02633×mean blood pressure (mm Hg)−2.30699×(ARIAR‐Risk: absolute low risk)−0.76712×(ARIAR‐Risk: relative low risk)+1.02621×(ARIAR‐Risk: aggressive high risk)+2.09761×(ARIAR‐Risk: refractory high risk).

Figure 4

Derivation and validation of an in‐hospital complications nomogram. A, In‐hospital complications nomogram. This nomogram provides a method to calculate the probability of cumulative incidence of developing postoperative in‐hospital complications after complete repair of tetralogy of Fallot, on the basis of a patient's combination of covariates. AUCs for the derivation cohort (B), internal validation with bootstrapping (C), and external validation cohort (D) of the in‐hospital complications nomogram. Calibration plots for the derivation cohort (E), internal validation with bootstrapping (F), and external validation cohort (G) of the in‐hospital complications nomogram. Shaded area is 95% CIs of the cumulative incidence. Dashed line is the reference line, which would indicate where an ideal nomogram would lie. Decision curves for the in‐hospital complications nomogram in the derivation cohort (H), internal validation with bootstrapping (I), and external validation cohort (J). Gray solid line indicates net benefit of a strategy of treating all patients. Blue dotted line indicates net benefit of treating no patients. Color solid line indicates net benefit of a strategy of treating patients for the derivation cohort, for internal validation with bootstrapping, and for the external validation cohort according to the nomogram predictions. AUC indicates area under the receiver operating characteristic curve; BP, blood pressure.

In addition, we further developed another nomogram by integrating the ARIAR‐Risk classifier and multifractional polynomial–selected variables (patient age group and mean blood pressure divided by 100). We compared multifractional polynomial–selected and AIC‐selected nomogram models and found that the 2 models had similar prediction performance in terms of AUC (multifractional polynomial–selected model, AUC: 0.785 [95% CI, 0.730–0.839]; AIC‐selected model, AUC: 0.785 [95% CI, 0.731–0.839]), so we finally selected the AIC‐based nomogram model for validation, taking into consideration the clinical significance, in which there is no need for mean blood pressure divided by 100.

The overall accuracy of the in‐hospital complications nomogram was moderate for the derivation cohort (Nagelkerke's R 2=0.206) and for the validation cohort (Nagelkerke's R 2=0.298). The discrimination of the nomogram was moderate to good at most for the derivation cohort (AUC: 0.785 [95% CI, 0.731–0.839]), for internal validation with bootstrapping (AUC: 0.784 [95% CI, 0.729–0.837]), and for the external validation cohort (AUC: 0.759 [95% CI, 0.636–0.881]; Figure 4B–4D). The specificity and sensitivity were 0.7735 and 0.7105, respectively, for the derivation cohort; 0.7231 and 0.7500, respectively, for internal validation with bootstrapping; and 0.8382 and 0.6500, respectively, for the external validation cohort (Table 2).

Table 2

Performances of Nomogram for Derivation Cohort, Internal Validation With Bootstrapping, and External Validation Cohort

	Derivation Cohort	Internal Validation With Bootstrapping	External Validation Cohort
Specificity	0.774	0.723	0.838
Sensitivity	0.711	0.750	0.650
Accuracy	0.764	0.727	0.814
Positive likelihood ratio	3.136	2.709	4.018
Negative likelihood ratio	0.374	0.346	0.418
Positive predictive value	0.353	0.320	0.371
Negative predictive value	0.939	0.943	0.942

The calibration plots of the nomogram for in‐hospital complication probability showed moderate to good performance at most (Figure 4E–4G). The Hosmer–Lemeshow calibration test was not significant both for the derivation cohort (χ2=6.933, P=0.544) and the external validation cohort (χ2=7.264, P=0.508); both indicate a good fit. The decision curves for in‐hospital complication probability in the derivation cohort, internal validation with bootstrapping, and external validation cohort (Figure 4H–4J) showed relatively good performance for the model in terms of clinical application. If the threshold probability in clinical decision was more than 10%, then use of the nomogram model to detect in‐hospital complications showed a greater advantage than assuming that all patients would develop in‐hospital complications or that no patients would develop in‐hospital complications.

Figure 5A through 5C shows the cumulative probability of in‐hospital complications by age at surgery and how this probability depends on the overall, triple‐risk, and ARIAR‐Risk classifiers. Patients with high risk scores had the highest cumulative incidence, whereas intermediate‐ and low‐risk patients had significantly lesser cumulative risk patterns (both P for log‐rank <0.0001). As observed with the nomogram model, no significant change was detected for the prognostic strength of in‐hospital complication predictors by varying the year of surgery (Wald test for interaction: mean arterial pressure×year of surgery, P=0.563), even if the ARIAR‐Risk classifier and age group at surgery were modeled as continuous scales (risk score×year of surgery, P=0.568; age at surgery×year of surgery, P=0.374; Figure S4A–S4C).

Figure 5

Cumulative risks of in‐hospital complications with increasing age at surgery. Kaplan–Meier curves by overall (A), triple‐risk (ie, low, intermediate, high) classifier (B), and ARIAR‐Risk classifier (C). Shaded area is 95% CIs of the cumulative incidence.

Given the heterogeneity of age, we divided the derivation cohort into 2 subgroups (406 younger and 107 older patients) based on the age threshold of 60 months to investigate whether the nomogram model performed equally well in older and younger patients. The discrimination of the nomogram was better for the younger subgroup (AUC: 0.791 [95% CI, 0.734–0.849]) than the older subgroup (AUC: 0.759 [95% CI, 0.700–0.819]). The Hosmer–Lemeshow calibration test was not significant for either the younger subgroup (χ2=3.215, P=0.920) or the older subgroup (χ2=8.713, P=0.367). The accuracy of this model was moderate for the younger subgroup (Nagelkerke's R 2=0.225) and older subgroup (Nagelkerke's R 2=0.163).

Discussion

In this multicenter retrospective cohort study, we developed and validated a novel predictive tool based on 11 preoperative and 4 intraoperative variables selected to improve the ability to predict in‐hospital complications in Chinese children with tetralogy of Fallot repaired at an older age. Our results showed that the triple‐risk classifier developed in this study categorized patients into low‐, intermediate‐, and high‐risk groups of patients who had significantly different probabilities of in‐hospital complications. Furthermore, the optimized ARIAR‐Risk classifier showed significantly better predictive performance than the triple‐risk classifier at predicting in‐hospital complications in tetralogy of Fallot repaired at an older age. We built a nomogram based on the ARIAR‐Risk classifier and independent baseline variables to predict individual risk of in‐hospital complications, and it showed good discrimination and goodness‐of‐fit.

Given that tetralogy of Fallot is a heterogeneous disease and in‐hospital complications result from multifactorial synergies,1, 4 exploring the key determinants involving initiation and derivation of in‐hospital complications might help to improve prognostic and therapeutic strategies.7, 8, 27 In the current study, we identified a panel of 15 perioperative variables that effectively predict in‐hospital complications in children with tetralogy of Fallot repaired at an older age. Among these candidates, the presence of a lower McGoon ratio, right‐to‐left shunting, lower indexed LV end‐diastolic volume, right bundle‐branch block, and major aortopulmonary collaterals were previously shown to be associated with more severe RVOT obstruction and thus to contribute to worse outcome.28 In patients with preexisting major cyanosis, Tet spell history, higher hematocrit, and inferior NYHA class, the cardiac autoregulatory capacity is likely impaired, thus rendering heart more susceptible to surgical strikes with bypass and compromising functional recovery.8

Interestingly, our results showed that higher overriding is associated with increased risk of in‐hospital complications; this finding could be explained by the fact that patients with higher overriding frequently require larger Dacron patches to close ventricular septal defect areas, resulting in myocardial deformations and geometric impairments.29, 30 Isolated infundibular outflow patch,31 absence of tricuspid valve detachment, and shorter CPB duration were found to relate to less invasive strikes, which result in better preservation of cardiac structural integrity and minimization of systemic inflammatory responses mainly induced by ischemia/reperfusion injury. Clinically promising results have been achieved by lowering reoxygenation via adjusting oxygen pressure during bypass in patients with cyanotic heart diseases; this approach likely alleviated hypoxia/reoxygenation injury to reduce the risk of major organ dysfunctions in cyanotic patients.32, 33 Consequently, this prediction model may help clinicians identify high‐risk patients following repaired tetralogy of Fallot.

Although some risk factors and biomarkers have been associated with in‐hospital complications following pediatric cardiac surgery,12, 34 this specific nomogram is still lacking for prediction of the risk of in‐hospital complications for an individual, especially for repaired tetralogy of Fallot; therefore, we incorporated the ARIAR‐Risk classifier. This classifier provides insights into pathophysiologic, anatomic, and procedural heterogeneities of tetralogy of Fallot repair and independent baseline variables that reflect sociodemographic and baseline clinical heterogeneities to develop a nomogram for predicting in‐hospital complications. This nomogram was further verified in external validation cohorts and exhibited good model performance; it might provide a simple and accurate method for predicting prognoses in pediatric patients with repaired tetralogy of Fallot.7, 8, 35

Apart from 15 clinical, anatomic–physiologic, and procedural candidates, we identified age at surgery as an independent predictor of in‐hospital complications. Our analysis demonstrated an increasing trend for in‐hospital complications with increasing age and decreasing mean blood pressure. Baseline lower mean blood pressure is indicative of reduced blood supply to organs and end‐organ damage, such as the brain and the heart, which leads to serious consequences.36 Three months to 3 to 4 years is a good age at which to operate on tetralogy of Fallot in clinical practice; however, the median patient was approximately 5 years of age in our study. Five is relatively old for repair in much of the world, where most cases are repaired in infancy or shortly thereafter because older age is associated with a stiffer right ventricle, higher diastolic dysfunction, and increased chance of mortality and morbidity due to prolonged cyanosis.37 Given the heterogeneity of age, we divided the derivation cohort into younger and older subgroups to investigate whether the nomogram model performed equally well in older and younger patients. Our results suggested that the discrimination of the nomogram model was better for the younger subgroup than for the older subgroup. Taking into consideration the sample cohort age, the generalizability of our results must be interpreted in the context of demographic and clinical characteristics, especially for age at surgery in our current study.

Our findings should be interpreted in the context of the study's limitations. First, surgical strategy selection in the study was determined by balancing the risks and benefits associated with each procedure in conjunction with the available baseline factors and the preferences of cardiac surgeons involved; therefore, the specific expertise may differ from those of other practitioners, potentially limiting the generalizability of these results to other institutions. In addition, although our sample size of 513 patients in the derivation set was small compared with established prediction tools based on large numbers of congenital heart disease patients, the power of a study is driven by the number of events, not simply by the total number of patients. With a prevalence of ≈15% of patients who developed in‐hospital complications in our derivation cohort, the size of our study was adequate to power the derivation of risk prediction models. In addition, some missing data were imputed, limiting the generalizability of the study, especially with validation in a small cohort; prospective large‐scale studies are needed to further verify the current findings.

Conclusions

To the best of our knowledge, this study is the first to incorporate per‐ and intraoperative characteristics to develop and validate a predictive model for in‐hospital complications in children with tetralogy of Fallot repaired at an older age. Our findings show that the triple‐risk classifier can effectively classify patients into different risk groups of in‐hospital complications, thereby improving predictive value for the assessment of patient prognosis. Moreover, we showed that the optimized ARIAR‐Risk classifier might have significantly better predictive performance than the triple‐risk classifier for identifying patients who would develop in‐hospital complications following tetralogy of Fallot repair at an older age. A nomogram including the ARIAR‐Risk classifier might help clinicians in directing personalized therapeutic regimen selection for patients with tetralogy of Fallot repaired at an older age. However, attention should be paid to the fact that the patients included in this study underwent surgical repair of tetralogy of Fallot at an older age that is different from much of the rest of the world.

Sources of Funding

This work was mainly supported by the Fundamental Research Funds for the Central Universities (No. 3332018189) and National Clinical Key Specialty Construction Projects of China.

Disclosures

None.

Table S1. Definition of Preoperative and Intraoperative Variables in the Derivation Cohort

Table S2. Baseline and Perioperative Characteristics of Patients in the Derivation Cohort Used for Nomogram Construction: Pre‐ and Postimputation

Table S3. Selected Baseline and Perioperative Characteristics of Patients in the Validation Cohort Used for Nomogram Validation: Pre‐ and Postimputation

Table S4. Comparison of Multiple Rates for Each Risk Group

Figure S1. Flow charts of the derivation (A) and validation (B) cohorts.

Figure S2. A, The distribution plots of risk score of in‐hospital complications among the derivation cohort. B, Violin plots for risk score of in‐hospital complications in patients with or without in‐hospital complications. The width of the colored shape indicates the probability density of patients with a given result. The gray notched box plots represent the median (yellow horizontal line), 95% CI of the median (notch), interquartile range (25th to 75th percentiles; box), and the upper 1.5 times the interquartile range (solid vertical line).

Figure S3. Association between risk‐based classifier and risk of in‐hospital complications. Unadjusted association between triple‐risk (ie, low, intermediate, high) classifier (A) and ARIAR‐Risk classifier (B) and probability and odds ratios of in‐hospital complications with 95% CIs. ARIAR‐Risk indicates absolute and relative low risk, intermediate risk, and aggressive and refractory high risk.

Figure S4. Distribution of selected variables according to year of surgery. Patient's age at surgery (A), mean BP (B), and risk score (C) according to year of surgery in the derivation cohort. Data are presented as medians with Tukey whiskers (boxes represent interquartile ranges, and bars represent 50% extreme quartiles). BP indicates blood pressure.

Click here for additional data file.