Affecting 40,000 neonates in the United States (US) annually, congenital heart defects (CHDs) are the largest contributor to birth defect-related mortality.¹ Population-based birth defects registries have been key data sources for describing CHD and other birth defect occurrence by population characteristics and better understanding factors associated with CHD occurrence risk, as they are large and representative of statewide populations.^2–4 Beyond etiologic studies, postnatal outcome studies of individuals with CHDs will be instrumental for understanding the impact of CHDs from birth to adulthood.

Classifying complex combinations of CHDs appropriately for outcome studies is challenging from both clinical and epidemiologic perspectives.⁵ Cases evaluated in studies of birth defect registry data are often identified using morphologically-based codes, such as British Pediatric Association (BPA) codes.^6,7 While these codes are helpful to assess the prevalence and etiology of CHDs, they do not capture complex cardiac phenotypes observed among patients with CHDs or the nuances of clinical decision-making. Alternatively, empirically-based classification tools such as the Society of Thoracic Surgeons (STS) Congenital Heart Surgery Database (CHSD) offer a diagnosis-based classification system.⁸ While tools like the STS CHSD risk models that stratify patients according to procedure type and/or diagnosis are helpful to evaluate outcomes among patients warranting surgical intervention at a tertiary center, they are less helpful for birth defects registry data that do not systematically include these procedure/surgical data. Moreover, registries also include patients without surgical intervention. Further, while databases like the CHSD are helpful to evaluate outcomes within the perioperative and/or postoperative periods specifically, they are not designed for evaluation of outcomes outside this period, which is frequently the focus of birth defect registry studies.

For studies using birth defect registry data, there are currently two common, morphologically-based methods of systematically accounting for multiple CHD codes in etiologic analyses.^9,10 The first analyzes all cases with the CHD code of interest (a traditional “broad” approach), regardless of co-occurring CHD codes. For example, all cases with the pulmonary stenosis code (746.010) would include both cases with “isolated” pulmonary stenosis (i.e., accompanied by no other major CHD code) and those with ≥1 major co-occurring CHD code in the same analytic group. This approach, however, may be irrelevant to outcomes analyses if coding results in cases with substantially differing mortality/morbidity risk profiles are included in the same group. Additionally, CHD groups may overlap, as this approach only accounts for the presence of a given CHD of interest (e.g., defect A) without accounting for any co-occurring defects (defect B), and therefore individuals possessing both defects A and B would be double counted.

A second method for CHD classification analyzes cases with a single CHD code and no other CHD codes (a traditional “exclusive” approach) and is considered to be more appropriate when the clinical impact is expected to differ between patients with 1 versus >1 CHD codes, such as the pulmonary stenosis subgroups described above. However, the exclusive approach is inadequate for many moderate to high-risk lesions as it ignores common CHD patterns. Thus, it excludes commonly or universally co-occurring CHD combinations, such as cases with interrupted aortic arch type B and a large ventricular septal defect, who comprise most cases with interrupted aortic arch type B, or children with complex lesions and co-occurring CHDs required for survival to birth.^11,12

Further, there are common multiple CHD patterns with highly differential mortality that would be ignored by either traditional broad or exclusive approaches. In the case of pulmonary atresia (PA; 746.000), a heterogenous lesion with variable mortality estimates depending on its subtypes (with ventricular septal defect, intact ventricular septum, or complex), the broad approach would evaluate PA as a whole group, despite this heterogeneity. Conversely, the exclusive approach would theoretically remove all cases with PA with ventricular septal defect or complex PA that, by definition, include co-occurring CHDs. Therefore, classifying CHDs into core common patterns of lesions may be necessary to optimally parse out heterogeneity, which is not possible using currently available approaches.

Given these aforementioned limitations, a classification system that flexibly balances the benefits of the broad and exclusive approaches is needed. Similar to that of clinical registries like STS, the use of both clinical and epidemiologic expertise to make classification decisions can better ensure that CHD classifications remain clinically relevant while ensuring adequate sample size.^13–15 Here, we report a lesion-specific system for classifying cases with co-occurring CHDs for outcome studies based on consideration of expected clinical impact, surgical/intervention pathway, and the most common combinations of CHDs based on prior work of many experts and investigators. To demonstrate its utility, we applied the proposed outcome-based classification system to data from a large state birth defects registry. Although we developed this approach for investigators studying data from the Texas Birth Defects Registry, we present it here in sufficient detail so researchers may apply it to other datasets.

After applying the C-CLOC system to 391,517 cases in the TBDR, 62,262 had an eligible CHD (Table 1). Left-sided lesions were the parent CHD category most frequently observed (13.4% of cases with CHDs), followed by right-sided lesions (9.2%), conotruncal (6.1%), endocardial (4.8%), looping defects (3.3%), and double outlet right ventricle (1.8%). The most frequently observed CHD lesion was ventricular septal defect (59.8%). A small proportion of subjects (4.9%) lacked sufficient information to classify them into any of the 59 C-CLOC groups.

This study proposes a CHD classification system tailored to outcome studies. Within population-based birth defect registries, morphology-based classification systems (broad or exclusive approaches) have been useful for conducting CHD surveillance, association studies, and descriptive epidemiologic investigations. However, the broad approach does not account for major co-occurring CHDs that increase risk for poor outcomes and subsequently may result in heterogenous CHD groups. Conversely, the exclusive approach can improve homogeneity in groups at the expense of statistical power due to low cell counts. Informed by four practicing clinicians and experts in pediatric cardiology, birth defect classification systems, and outcomes research, C-CLOC was developed to balances the benefits of the broad and exclusive approaches. Importantly, C-CLOC classifies CHDs with respect to other commonly co-occurring CHDs that are expected to modify their risk, a feature that is not otherwise available in traditional methods as these subgroups were specified a priori by clinical experts’ input when devising C-CLOC phenotyping schema. Ideally, C-CLOC groups would be based on validated mortality reports. However, prior population-based CHD mortality reports are primarily one of the following: 1) limited to patients undergoing surgery, 2) insufficiently granular, or 3) are based on the existing BPA codes which have major weakness for mortality reporting as discussed. We hope to use these codes to describe population-based mortality in future work, and may need to refine over time. Our results do suggest that, for many CHDs, C-CLOC groups are more homogeneous than those from the broad approach while retaining case counts that are larger than those from the exclusive approach. C-CLOC also classifies subjects into mutually exclusive CHDs categories, though, as expected, this generally resulted in some cases moving from lower to higher complexity lesion categories due to assignment of cases to the CHD category for the highest complexity lesion present.

Key benefits of C-CLOC include its stratification of CHDs into clinically meaningful groups that can be applied for prognosis-based clinical counseling. In the case of pulmonary atresia, a clinically heterogenous lesion that differs in risk depending on its subtype, stratification of pulmonary atresia into complex, ventricular septal defect, and intact ventricular septum subtypes allow for better comparisons with published clinical data. Similarly, neonatal mortality estimates for high-risk lesions ascertained via C-CLOC phenotyping were consistent with prior clinical literature.^21–23 Additionally, C-CLOC does not force exclusion of patients with genetic anomalies, enabling its application to study effect modification or stratification by genetic syndromic diagnoses (a notable point given that up to one-third or more of patients with CHDs have a genetic condition).^12,24,25 Importantly, designating a standard classification system created to evaluate CHDs according to their expected clinical consequences rather than morphologic homogeneity, will allow increased utility of coded population-based resources for outcomes-based research (e.g., state birth defects registries) and may be instrumental for risk stratifying individuals with CHDs over the lifespan.

Following replication and validation in other CHD cohorts, C-CLOC may be a useful tool to supplement prior classification systems used in outcome studies. Systems such as the Risk Adjustment in Congenital Heart Surgery (RACHS-1) scores, STS European Association for Thoracic Surgery (STAT) mortality scores, and Aristotle Basic Complexity (ABC) scores have been applied to study mortality risk in children with prior CHD surgery.^26–30 While these systems extrapolate comprehensive information based on available clinical and surgical data, they are limited to those with surgery and cannot characterize patients who did not undergo surgery. Additionally, there has been a call within the surgical community for diagnosis-based classification even among those undergoing surgery, as this would allow comparison of intervention approaches, since not all patients with the same condition undergo the same procedures.¹³ With respect to length of outcomes evaluated, prior outcome studies of children with CHDs often focus on perioperative and postoperative outcomes, whereas outcomes evaluating their course following hospital discharge is not well-understood.^26–31 Thus, a comprehensive diagnosis-based approach that can be applied to population-based data sources with outcomes data available like C-CLOC offers many benefits. Investigations comparing utility of C-CLOC to prior systems like RACHS-1, STAT, and ABC scores are needed for further assessment.

There are several other potential classification systems that have been used for outcomes that also do not incorporate surgical history, including those discussed by Reller et al and Glidewell et al.^32,33 Reller’s investigation has the strength of cases (n ~ 3,000) being manually reviewed by a team of pediatric cardiologists.³² This allowed for STS classifications to be assigned which added significant granularity and specificity to the dataset, as 89% of cases had only one primary STS-coded CHD. This manual classification, however, is not feasible many studies, particularly for investigations with tens or hundreds of thousands of cases or who do not have access to electronic medical record. The goal of the C-CLOC classification system is for BPA codes to stand alone and allow for automatic C-CLOC classification without manual review of records or merging with STS databases. Additionally, the Reller classification system as published was not mutually exclusive and did not seem to account for miscoded lesions. For example, a subject with both HLHS and atrioventricular canal defect (AVC) coded (which are mutually exclusive) in the Reller paper would either be counted as both, or perhaps review of the subject record would show that one code was in error. In the C-CLOC system, knowing that unbalanced right-dominant atrioventricular canal defect are often miscoded as HLHS but that the reverse is very rarely true, this would automatically only be coded as complex atrioventricular canal defect.

In the Glidewell et al 2021 study, CHD cases were classified into 4 groups: “severe”, and then three groups of “non-severe”, consisting of “shunt+valve”, “shunt”, and “valve”.³³ While Glidewell et al 2021 assessed a large population using available U.S.-based resources, the classification system was very broad, was not lesion-based, and grouped together very broad risk populations. For example, both simple transposition of the great arteries and hypoplastic left heart syndrome were in the “severe” group, but these have dramatically different mortality rates (for example 5.8% vs. 22% hospital mortality in a multicenter retrospective study.³⁴ The C-CLOC system aims to be much more granular, especially in regards to more high-risk CHD lesions. Additionally, because C-CLOC groups developmentally similar lesions together, it can be used for genetic association studies as well.

Our initial “validation” in comparison to manual clinical coding for HLHS suggests that applying C-CLOC had a higher classification accuracy than the broad and exclusive approaches, maintaining more of a balance between sensitivity and specificity, whereas the broad and exclusive approaches respectively optimized either sensitivity or specificity but not both. HLHS was well-suited for validation given its frequent misclassification, which may not necessarily be the case for other CHDs. For the present study, “gold standard” HLHS cases were identified by a pediatric cardiologist given there was no standardized tool available to identify a distinct group of HLHS, although given that HLHS gold standard cases were identified by the same team that generated the C-CLOC algorithm, HLHS validation results should be interpreted with caution. While it is likely that manual clinical review of individual records should continue to represent the gold standard for CHD classification, such as that of prior literature, automated approaches such as the C-CLOC system may be increasingly useful for researchers utilizing data from surveillance systems that lack resources to perform manual classification.³² Further, our approach offers a standardized system for efficiently classifying thousands of cases, with coding criteria that can be easily reapplied as new data emerges or modified and reimplemented as appropriate for specific research questions.