Understanding Mortality After Congenital Heart Surgery: What Do Procedure-Specific Factors Add?

Abstract

BACKGROUND

Although collected by The Society of Thoracic Surgeons Congenital Heart Surgery Database (STS-CHSD) since 2013, the value of procedure-specific factors (PSFs) in improving our understanding of expected mortality has not been studied. We evaluated the contribution of PSFs across a large cohort.

METHODS

Included were benchmark operations (BMO) for which PSFs are captured across 115 United States centers (2016-2022). We assessed the influence of PSFs on operative mortality discrimination and determined which specific PSFs had the most impact. Advanced modeling strategies were used given the numerous covariates and relatively low mortality rates. A baseline model (including standard STS-CHSD clinical risk variables) was compared with a baseline plus PSF model, with results reported overall and stratified by BMO.

RESULTS

Among 37,282 included BMOs, overall operative mortality was 2.6%. The proportion of BMOs with at least 1 PSF recorded as present ranged from 2.7% for ventricular septal defect to 85% for Norwood. Model discrimination increased from 0.088 (baseline) to 0.095 (baseline + PSF) overall, with better discrimination for 5 of 9 BMOs. Results for individual PSFs varied by BMO and analytic method. No PSFs were retained in the final models for ventricular septal defect in any scenario. The greatest number of PSFs were retained for Norwood and truncus arterious repair.

CONCLUSIONS

Incorporating PSFs into expected mortality models improved model discrimination for some BMOs, and certain PSFs had an important impact on mortality estimates, whereas others did not. These data can aid in reducing data collection burden and support ongoing refinement of risk models.

Understanding center performance for children undergoing congenital heart surgery is challenging due to the wide heterogeneity of diagnoses, procedures, and patient comorbidities, coupled with substantial variability in the types of patients treated or case-mix across centers. To support numerous quality improvement, benchmarking, and reporting initiatives, many data sources have added more detailed variables attempting to better assess patient risk and “expected” center outcomes based on their case-mix and risk profile of treated patients.^1–9

In 2013, The Society of Thoracic Surgery Congenital Heart Surgery Database (STS-CHSD) introduced procedure-specific factors (PSFs), later expanded to 113 PSFs for 11 procedures, including 82 PSFs for 9 STS-CHSD benchmark operations (BMOs).⁹ Although PSFs provide granular anatomic information, their collection adds significant time, complexity, and cost to data extraction. Importantly, PSFs are not currently included in STS-CHSD risk models, and their potential to enhance our understanding of expected mortality risk beyond the standard clinical risk variables has not yet been studied.

To address these issues, we evaluated the impact of PSFs on estimation of expected mortality beyond the standard clinical variables. We assessed the influence of PSFs on model discrimination and studied which specific PSFs impacted expected mortality.

MATERIAL AND METHODS

DATA SOURCE AND POPULATION.

We used STS-CHSD data between January 1, 2016, and June 30, 2022, restricting to the 9 BMOs for which PSFs were collected, based on the STS algorithm assigned primary procedure. We excluded coarctation (no PSFs collected) and the Cone procedure, (not a BMO) (Appendix 1). Detailed inclusion and exclusion criteria are provided in Supplemental Figure 1a and Supplemental Table 1. This study was approved by the Harvard Faculty of Medicine Institutional Review Board (IRB21-1365, initial approval 12/10/2021) with waiver of Health Insurance Portability and Accountability Act authorization and informed consent.

PROCEDURE-SPECIFIC FACTORS.

All 82 available PSFs were evaluated, and 4 were combined where clinically indicated. For instance, for atrioventricular canal (AVC) repair, a single variable was used to describe ventricular dominance and commitment of the common atrioventricular valve to the dominant ventricle (Supplemental Table 2).

MISCLASSIFICATION OF PSFs.

PSFs were not available due to misclassification in 2 situations. The first likely occurred when the center did not select the option for submitting PSF, and thus, the PSF fields were not filled irrespective of whether the center’s choice of primary procedure was a BMO or not and whether it matched the STS primary procedure or not. The second likely occurred during STS data processing when the STS-CHSD algorithm-assigned primary procedure was different from the center-chosen primary procedure. In both instances, the center did not submit PSFs. We created a new variable, “PSF missing due to misclassification” for these cases and included it in our models.

MISSING PSF.

Here the center did not submit PSF data despite selecting the option for submitting PSFs. The center-selected BMO was either in agreement with the STS algorithm-defined primary procedure or was a different BMO requiring PSFs. These included instances where either all or some of the PSF data were missing. In such cases, we assumed that center-assigned PSFs were absent and imputed them as “no” (Supplemental Figure 1b; Supplemental Table 2). We conducted a sensitivity analysis treating all PSF-missing cases as missing and conducting multiple imputations rather than treating them as “no” (Appendix 1).

OTHER MODEL VARIABLES.

In addition to the PSFs, we evaluated the standard clinical risk variables included in the most recently published version of the STS-CHSD mortality risk model⁵ (Appendix 1).

PRIMARY OUTCOME.

The primary outcome was operative mortality^1,2 as defined by the STS-CHSD as death occurring in the hospital (index hospital or acute care center after transfer), within 30 days of the index surgery if discharged home, or within 183 consecutive days after transfer to a long-term care facility.

ANALYSIS.

We assessed the impact of PSFs on expected mortality estimation by addressing 3 key questions:

1. Do PSFs improve mortality discrimination beyond the standard clinical risk variables?

2. When standard clinical risk variables are forced into the model, which PSFs are identified as important, and does the number of important PSFs vary across BMOs?

3. When variable selection is entirely data-driven, which standard clinical risk variables and PSFs are retained or excluded by the model?

Owing to the rarity of mortality and the low number of cases per BMO, all BMOs were combined into a single model. Additionally, the large number of standard clinical risk variables and the low operative mortality rate meant that logistic regression models could fail to converge due to complete separation. We adopted modeling strategies to (1) minimize bias caused by complete separation, small sample sizes, and low mortality rates, and (2) use PSF variable selection algorithms.^10,11 All models were trained and tested on independent random data splits, with 70% of the data used for training and 30% reserved for testing (testing cohort). Results are reported from the testing cohort overall and stratified by BMO.

To address the first question, we used logistic regression with a Firth correction to reduce bias from the small number of operative deaths. Two models were estimated: one using only standard clinical risk variables (baseline model) and another adding all PSFs to the baseline model (baseline + PSF model). The Firth baseline model is not identical to the current STS risk model used in harvest analysis, although all the risk variables are included. We screened for collinearity among covariates and removed standard clinical risk variables with negligible variability. Model performance was assessed overall and within each BMO by measuring discrimination (mean predicted mortality of nonsurvivors – mean predicted mortality of survivors; higher values indicate better discrimination), area under the receiver operating characteristic (AUROC) curve (1 is ideal), and the distribution of estimated mortality (median, interquartile range [IQR]; larger IQR is better).

Calibration was assessed using the Spiegelhalter Z statistic (0 is ideal) and calibration curves plotting predicted mortality against observed mortality in 0.05 increments. A model with perfect calibration would align closely with a 45° diagonal line. A model may discriminate well—rank the risk of dying among those who die higher than among those who survive, yet not be well calibrated (eg, the predicted risks are not close to the observed risks).

For the second and third questions, we used regression approaches that either forced the standard clinical risk variables to remain in the model or permitted entirely data-driven selection. We applied 3 approaches. We used LASSO (least absolute shrinkage and selection operator), which allowed selection of individual PSFs in the model. We used group LASSO which enabled retention or elimination of entire groups of PSFs. If all PSFs for a BMO were excluded, this suggested PSFs did not enhance model performance for that BMO. Exclusive LASSO allowed structured variable selection, requiring the selection of at least one PSF for each BMO. All analyses were conducted using R 4.3.2 software (The R Project for Statistical Computing).

RESULTS

STUDY POPULATION CHARACTERISTICS.

The cohort included 37,282 BMOs over a 6.5-year study period from 115 centers. Neonates comprised 24% (n = 8991) of the cohort, and cohort characteristics are summarized in Table 1.⁵ The overall operative mortality rate was 2.6% (n = 954), varying from 0.4% (46 of 11,613) for VSD to 12.8% (500 of 3917) for Norwood (Supplemental Table 3).

TABLE 1.

Study Population Characteristics of Entire Cohort of 37,282

Characteristic	Data Values (N = 37,282)
Mortality	954 (2.56)
Female sexa	16,453 (44.18)
Age, mean (SD), y	1.29 (3.37)
Neonate (0 to 28 days)	8991 (24.12)
Infant (29 days to 1 year)	20,812 (55.82)
Child (>1 year to <18 years)	7254 (19.46)
Adult (≥18 years)	225 (0.6)
Prematurity among neonates and infantsa	5626 (15.11)
Weight, mean (SD), kga	7.78 (9.21)
Preoperative factors	14,794 (39.68)
Mechanical circulatory support	78 (0.21)
Shock	161 (0.43)
Neurologic deficit	726 (1.95)
Renal dysfunction or dialysis	420 (1.13)
Mechanical ventilator support	4372 (11.73)
Any other preoperative factors	7458 (20)
Any prior cardiothoracic proceduresa	4726 (12.69)
Number of prior cardiothoracic procedures, mean (SD)	0.28 (0.86)
Any noncardiac congenital anatomic abnormality	7441 (19.96)
Tracheoesophageal fistula	235 (0.63)
Congenital diaphragmatic hernia	139 (0.37)
Omphalocele	105 (0.28)
Intestinal malrotation	440 (1.18)
Anal atresia	358 (0.96)
Gastroschisis	<11
Hirschsprung disease	122 (0.33)
Any chromosomal abnormality or syndromes	13,659 (36.64)
Chromosomal abnormality/syndrome risk groupb
Risk group 1 (lowest)	1881 (5.05)
Risk group 2	6404 (17.18)
Risk group 3	871 (2.34)
Risk group 4	468 (1.26)
Risk group 5 (highest)	15 (0.04)

^aWeight was missing in <11 patients; sex in 44, prematurity in 43, and prior cardiothoracic surgery in 53 of this cohort;

^bRisk group represents chromosomal abnormality and syndromes grouped based on risk of mortality from lowest to highest, as previously described.⁵ Data are presented as n (%), or as mean (SD) unless indicated otherwise.

MISSING AND MISCLASSIFIED PSFS.

Overall, 8016 BMOs (21.5%) had misclassified PSFs, and 1226 (3.3%) had at least 1 missing PSF (Supplemental Figure 2; Supplemental Table 2). Of the 8016 misclassified, no PSFs were submitted either because the “need PSF” option was not selected, and thus the fields to submit PSFs were not available (n = 7719), a non-BMO procedure was selected as primary procedure by site (n = 248), or the site chose a different BMO from the STS (n = 49). The instance where the STS did not assign BMO as primary procedure (although the center defined it as a BMO) occurred in 200 cases (0.15%). At the center level, all but 2 institutions had at least 1 misclassified BMO, with 20 of 115 (17%) centers having ≥50% of their cases misclassified. One large hospital accounted for 12.3% of the misclassified BMOs. The Norwood operation was the least likely to be misclassified. The baseline characteristics of and outcomes with and without PSFs due to missingness or misclassification for the overall BMO cohort (n = 37,282) are provided in Supplemental Table 2.

PSF FREQUENCY AND OUTCOME.

The proportion of BMOs with at least 1 PSF reported as present varied from 2.7% for VSD to 85% for Norwood. The 5 PSFs with the highest associated mortality were grade 3/4 atrioventricular valve regurgitation (Norwood), intact atrial septum/obstructed pulmonary venous return (Norwood); moderate/severe truncal valve stenosis (truncus); presence of sinusoids (Norwood); and moderate/severe truncal valve regurgitation (truncus). The Figure highlights the 2 PSFs with the highest mortality and the 2 most common PSFs within each BMO (details of all PSFs by BMO are in Supplemental Table 4). Most PSFs associated with mortality >10% were concentrated in the Norwood and truncus groups (Supplemental Table 5).

FIGURE 1.

Highest mortality and most common procedure-specific factors (PSFs) across benchmark operations (BMOs). The left panel lists the 2 PSFs with highest mortality for each BMO and the right panel lists the 2 most common PSFs by BMO. In both panels, the total length of the bar represents the frequency (%) of each specified PSF within the BMO indicated. The percentage who died is represented in maroon, and percentage alive is represented by the blue portion of the bar. *Where N < 10, we are unable to specify mortality status due to privacy regulations, and the total bar is presented in grey. (ASO, arterial switch operation; AV, atrioventricular; AVC, complete atrioventricular canal; AVVR, atrioventricular valve regurgitation; Cx, circumflex coronary artery; LV, left ventricle; PA, pulmonary artery; RCA, right coronary artery; RV, right ventricle; RVOT, right ventricular outflow tract; TOF, tetrology of Fallot; VSD, ventricular septal defect.)

IMPACT OF PSF ON MORTALITY DISCRIMINATION.

This was assessed in the testing cohort of 11,186. Adding PSFs to the overall model improved model discrimination from 0.088 (baseline) to 0.095 (baseline + PSF), with an increase in the range of predicted deaths (IQR, 0.4%-2.1% baseline vs 0.3%-2.2% baseline + PSF). The AUROC curves did not differ (0.85 for both), nor did the Spiegelhalter Z statistic (Table 2; Supplemental Figure 2). When stratified by BMO, the findings varied (Supplemental Figures 2–4). The range of predicted probability of mortality in the baseline + PSF model was greater for all BMO except VSD, there was better discrimination for arterial switch operations (ASOs), AVC, Fontan, Glenn/hemi-Fontan and Norwood, and higher AUROC curves for Norwood, Glenn/hemi-Fontan, and Fontan for the baseline + PSF model. For example, the discrimination for truncus improved from 0.058 to 0.073, and the AUROC curve for Glenn/hemi-Fontan increased from 0.70 to 0.80 when PSFs were added to the baseline model.

TABLE 2.

Comparison of Baseline vs Baseline Plus Procedure-Specific Factor Model Discrimination Overall and Across Benchmark Operations Limited to the Testing Cohort of 11,186

Benchmark Operation	No.	Mortality (%)	Model Discrimination (higher is better)		Spiegelhalter Z Statistic (0 is ideal)		AUROC Curve (1 is ideal)		Predicted Probability of Mortality (wider range is better)
			Baseline	Baseline + PSF	Baseline	Baseline + PSF	Baseline	Baseline + PSF	Baseline Median (IQR)	Baseline + PSF Median (IQR)
All	11,186	2.50	0.088	0.095	−0.80	‒1.20	0.85	0.85	0.008 (0.004–0.021)	0.008 (0.003–0.022)
ASO	882	2.27	0.006	0.015	1.27	0.62	0.58	0.54	0.013 (0.009–0.019)	0.012 (0.008–0.021)
ASO+VSD	377	4.51	0.031	0.022	−1.02	−1.04	0.66	0.67	0.047 (0.035–0.068)	0.045 (0.031–0.069)
AVC	1379	2.03	0.010	0.015	1.20	0.82	0.69	0.69	0.013 (0.009–0.019)	0.013 (0.009–0.020)
Fontan	1000	1.10	0.008	0.020	−0.46	‒0.71	0.73	0.79	0.007 (0.007–0.012)	0.009 (0.005–0.014)
Glenn/hemi-Fontan	565	1.77	0.029	0.049	−0.49	−0.75	0.70	0.80	0.015 (0.009–0.024)	0.014 (0.008–0.025)
Norwood	1175	11.66	0.039	0.054	−1.48	−1.56	0.63	0.67	0.113 (0.088–0.159)	0.112 (0.083–0.165)
TOF	2068	0.82	0.039	0.037	−1.18	−1.35	0.78	0.77	0.007 (0.005–0.010)	0.006 (0.004–0.011)
Truncus	252	8.33	0.073	0.058	−0.78	−0.23	0.72	0.60	0.068 (0.051–0.123)	0.064 (0.042–0.127)
VSD	3488	0.54	0.015	0.013	1.74	1.61	0.92	0.90	0.002 (<0.001–0.004)	0.002 (<0.001–0.004)

Results from Firth models in the testing cohort. Bolded values indicate better performance. ASO, arterial switch operation; AUROC, area under the receiver operating characteristic; AVC, complete atrioventricular canal; PSF, procedure-specific factor; TOF, tetrology of Fallot; truncus, truncus arteriosus; VSD, ventricular septal defect.

IMPACT OF PSF ON MORTALITY CALIBRATION.

This was assessed in a testing cohort of 11,186. For the overall model, the calibration plots (Supplemental Figure 3) favored the baseline model. The Spiegelhalter Z statistic was <1.96 SDs away from 0, suggesting no real difference with the baseline model. Calibration improved for AVC, Glenn/hemi-Fontan, Norwood, tetrology of Fallot, and VSD with the inclusion of PSFs (Table 2; Supplemental Figures 2, 3).

RESULTS WITH STANDARD CLINICAL RISK VARIABLES RETAINED.

Results were variable across operations and across analytic methods. Forcing retention of all the standard clinical risk variables in the models resulted in no PSFs selected for inclusion by the LASSO or group LASSO in the overall model. However, the number of PSFs retained ranged from 0 to 5 across the BMO with exclusive LASSO (0 for VSD, 4 for truncus, and 5 for Norwood) (Table 3).

TABLE 3.

Procedure-Specific Factors Selected Across Analytic Methods, Stratified by Retention of Standard Clinical Risk Variables

Variable	Standard Clinical Risk Variables Forced in			Entirely Data-Driven Variable Selection
	LASSO	Group LASSO	Exclusive LASSO	LASSO	Group LASSO	Exclusive LASSO
Standard clinical risk factors (No. retained of 54 total standard clinical risk variables) PSFs (No. and name of PSF variables retained by BMO)			54	20	19	2
ASO (14 total PSFs)	0	0	2 • Intramural coronary • Malaligned commissures	2 • Intramural coronary • Single coronary	All 14	2 • Intramural coronary • Single coronary
ASO/VSD (16 total PSFs)	0	0	2 • Circumflex from RCA • Aorto pulmonary diameter mismatch	4 • Circumflex from RCA • Aorto pulmonary diameter mismatch • Malaligned commissure • Double coronary loop	All 16	2 • Aorto pulmonary diameter mismatch • Circumflex from RCA
AVC (6 total PSFs)	0	0	2 • Single PM/parachute LAVV • Double orifice LAVV	0	0	2 • Single PM/parachute LAVV • Double orifice LAVV
Fontan (8 total PSFs)	0	0	1 • Grade 3/4 AVVR	1 • Grade 3/4 AVVR	All 8	2 • Grade 3/4 AVVR • LV dominance
Glenn/hemi-Fontan (8 total PSFs)	0	0	3 • RV dominance • Indeterminate ventricular dominance • Hypoplastic branch PAs	1 • RV dominance	All 8	3 • RV dominance • Indeterminate ventricular dominance • Hypoplastic branch PAs
Norwood (16 total PSFs)	0	0	5 • RV dominance • Grade 3/4 AVVR • Aortic atresia • RV-PA conduit • Systemic artery to PA shunt	8 • Grade 3/4 AVVR • Aortic atresia • Intact atrial septum/obstructed pulmonary venous return • Sinusoids • Aberrant RSCA • RV dominance • Systemic artery to PA shunt • Balanced ventricles	All 16	5 • RV dominance • Grade 3/4 AVVR • Systemic artery to PA shunt • RV • PA conduit • Aortic atresia
TOF (4 total PSFs)	0	0	2 • Multiple VSD repair • Hypoplastic branch PAs	0	All 4	1 • Multiple VSD repair
Truncus (4 total PSF)	0	0	4 • ≥Moderate truncal regurgitation • Abnormal coronary • Truncus type 3 • ≥Moderate truncal stenosis	4 • ≥Moderate truncal regurgitation • Abnormal coronary • Truncus type 3 • ≥Moderate truncal stenosis	All 4	3 • Abnormal coronary • ≥Moderate truncal regurgitation • ≥Moderate truncal stenosis
VSD (2 total PSFs)	0	0	0	0	0	0

Results from testing cohort across analytic methods and BMOs. A specified factor could have odds ratio of >1 or <1, but the significance of individual risk factors was not assessed. Supplemental Figure 4 provides directionality of odds ratios for the standard clinical risk variables and for PSFs. A complete listing of all PSFs for each BMO is available in Supplemental Table 4. ASO, arterial switch operation; AVC, complete atrioventricular canal; AVVR, atrioventricular valve regurgitation; BMO, benchmark operation; IQR, interquartile range; LASSO, least absolute shrinkage and selection operator; LAVV, left atrioventricular valve ; LV, left ventricle; PA, pulmonary artery; PM, papillary muscle; PSF, procedure-specific factor; RCA, right coronary artery; RSCA, right subclavian artery; RV, right ventricle; TOF, tetrology of Fallot; truncus, truncus arteriosus; VSD, ventricular septal defect.

RESULTS WITH DATA-DRIVEN VARIABLE SELECTION.

When methods were used that allowed entirely data-driven variable selection and permitted standard clinical risk variables to drop from the models, results were variable. Of the 54 standard clinical variables, the number that remained across various analytic methods included 20 (LASSO), 19 (group LASSO), and 2 (exclusive LASSO) (Supplemental Table 6). More PSFs were retained but varied by BMO. For VSD no PSFs were retained across any of the 3 analytic methods. For the group LASSO, all PSFs were dropped for AVC but remained for all other BMO. Across methods, the greatest number of PSFs retained by BMO were 2 of 14 PSFs for ASO, 4 of 16 PSFs for ASO/VSD, 2 of 6 PSFs for AVC, 2 of 8 PSFs for Fontan, 3 of 8 PSFs for Glenn/hemi-Fontan, 8 of 16 PSFs for Norwood, 1 of 4 PSFs for tetrology of Fallot, and 4 of 4 PSFs for truncus (see Table 3 for details regarding the specific PSFs retained).

SENSITIVITY ANALYSIS.

Our results were robust to multiply imputing missing PSFs rather than imputing to “no” (Supplemental Table 7).

COMMENT

Despite being collected across North American centers for the past decade, the impact of PSFs has not been studied to date. Our analysis found that incorporating PSFs into expected mortality models improved mortality discrimination for some BMOs and that certain PSFs had an important impact on mortality estimates, whereas others did not; however, PSFs contributed minimally to mortality prediction when standard risk variables were retained.

Since its inception in 1994, the STS-CHSD has progressively evolved with updates to the variables collected and risk models^1–5,8 Previous reports¹² have suggested that standard clinical variables captured across clinical registries likely explain only a portion of mortality risk, and in 2013, the STS-CHSD began collecting PSFs for the common BMOs. No previous models have been published specifically related to the BMOs.

Additional challenges included the large number of variables and relatively low mortality rates. Thus, we used advanced modeling techniques like those reported recently by our group that analyzed additional variables, such as patient diagnoses, in addition to procedures.⁶ We found variability in the impact of PSFs, with the addition of PSFs to standard clinical risk variables improving mortality discrimination for some BMOs, but not for others. In addition, individual PSFs were evaluated, some were important, but others were not retained in the final models. For instance, no PSFs were retained for VSD across any of the analytic methods, whereas the greatest proportion of PSFs were retained for higher-mortality procedures such as truncus and Norwood.

There are several implications and future directions related to our results. First, with the relatively low mortality rates for most BMOs, PSFs may have greater utility in the assessment of complications and length of stay. Future studies plan to explore these outcomes.

In addition, future efforts will quantify the magnitude of effect of specific PSFs and develop optimal models for both morbidity and mortality outcomes. Adding PSFs and allowing diagnosis to modify predicted risk for palliative operations may augment model performance for very high-risk surgical procedures. Taken together, these results will inform data collection recommendations, potentially condense the number of variables captured, and reduce burden on sites if not all PSFs are found to be useful.

We found several reasons why PSFs were not submitted for BMOs, which may involve a variety of potential solutions to minimize missing data. The first is failure to select the indicator that PSFs were required for a BMO and thus no PSFs submitted by site. A potential solution could be that PSF data entry is mandatory for BMOs and that this field (indicator) be removed from the data collection form and data entry software.

The second reason was that the center and STS primary procedure did not match, and thus, PSFs were not collected for STS-designated BMOs. Possible solutions to address this and ensure collection of relevant PSFs for all BMOs include incorporating the STS-CHSD algorithm for primary procedure selection into centers’ data collection software so that primary procedure selection is automated and at less risk for misinterpretation. Alternatively, centers could enter PSFs for BMOs regardless of whether they are designated as the primary procedure, although this may increase rather than minimize data collection burden. Further education and data audits may also be useful.

Third, certain centers opted not to submit or only partially submit PSFs despite selecting the indicator that PSFs were required for a BMO. A potential solution for this may be making PSF entry mandatory and additionally addressing this during the annual audits.

LIMITATIONS.

Our study has several limitations. Although based on an established, well-audited registry, center misclassification of primary procedures resulted in >20% of cases lacking PSFs. Although this could have an impact our results, it is important to note that we did not find important differences in observed baseline characteristics between BMOs with and without misclassified PSFs. Low mortality rates for most BMOs posed a risk of model overfitting, which we mitigated with Firth correction, sample splitting, and use of the LASSO. Finally, the analysis focused on outcomes during the episode of care captured in the registry and was not able to address the association of PSFs with longer-term outcomes. Such analyses will be important in future efforts.

CONCLUSIONS.

Incorporating PSFs into expected mortality models improved model discrimination for some BMOs, and certain PSFs had an important impact on mortality estimates, whereas others did not. Future work will involve evaluating the impact of PSFs on length of stay and complications. These efforts may aid in reducing the data collection burden, and support ongoing refinement of risk models.

Supplementary Material

The Supplemental Material can be viewed in the online version of this article [https://doi.org/10.1016/j.athoracsur.2025.09.025] on https://www.annalsthoracicsurgery.org.

Abbreviations and Acronyms

ASO

arterial switch operation

AVC

complete atrioventricular canal

BMO

benchmark operation

CHSD

Congenital Heart Surgery Database

LASSO

least absolute shrinkage and selection operator

PSFs

procedure-specific factors

STS

The Society of Thoracic Surgeons

Truncus

truncus arteriosus

VSD

ventricular septal defect