Meta-regression of sulcal patterns, clinical and environmental factors on neurodevelopmental outcomes in participants with multiple CHD types

Abstract

Congenital heart disease (CHD) affects 1% of infants and is associated with impaired neurodevelopment. Right- or left-sided sulcal features correlate with Executive Function among people with Tetralogy of Fallot (ToF) or single ventricle CHD. Studies of multiple CHD types are needed to understand regional differences. Further, sulcal pattern has not been studied in people with d-transposition of the great arteries (d-TGA). Therefore, we assessed the relationship between sulcal pattern and Executive Function, General Memory, and Processing Speed in a meta-regression of 247 participants with three CHD types (114 single ventricle, 92 d-TGA, and 41 ToF) and 94 participants without CHD. Higher right hemisphere sulcal pattern similarity was associated with improved Executive Function (Pearson r=0.19, False Discovery Rate (FDR)-adjusted p=0.005), General Memory (r=0.15, FDR p=0.02), and Processing Speed (r=0.17, FDR p=0.01) scores. These positive associations remained significant in for the d-TGA and ToF cohorts only in multivariable linear regression (estimated change β=0.7, FDR p=0.004; β=4.1, FDR p=0.03; and β=5.4, FDR p=0.003, respectively). Duration of deep hypothermic circulatory arrest was also associated with outcomes in the multivariate model and regression tree analysis. This suggests that sulcal pattern may provide an early biomarker for prediction of later neurocognitive challenges among people with CHD.

Introduction

Congenital heart disease (CHD) is the most prevalent major congenital anomaly, occurring in 1% of births (A. J. Marelli et al., 2014; Pulcine & deVeber, 2021). With advances in medical and surgical care, survival of people with CHD has improved significantly in the past few decades (Green, 2004; Marino et al., 2012). This has highlighted the need to understand and treat morbidities across the lifespan including neurodevelopmental delay or disability (Blue et al., 2018; Derridj et al., 2021; Nattel et al., 2017; K. Williams et al., 2019). Different forms of CHD have varying prevalence of neurodevelopmental delay or disability, as well as disturbances of cerebral oxygenation and blood flow both in the fetal period and after birth. For example, in d-transposition of the great arteries (d-TGA), fetuses and preoperative neonates have relative cerebral hypoxia but preserved perfusion. In single ventricle CHD (SV), mixing of oxygenated and deoxygenated blood causes relative cerebral hypoxia, and in those with obstruction to systemic outflow, to hypoperfusion. The decreased delivery of oxygen during the fetal period associated with d-TGA and SV could alter brain developmental trajectory and thereby contribute to decreased brain growth and impaired cellular maturation and function (Licht et al., 2009; Mahle et al., 2002; Miller et al., 2007; C. Ortinau et al., 2012). These early changes can in turn contribute to the neurodevelopmental impairments observed to be more frequent in people with CHD, such as executive dysfunction, visual-spatial impairment, attention deficit hyperactivity disorder, language and speech disorders, developmental delay, and motor coordination difficulties (Goldberg et al., 2014; A. Marelli et al., 2016; Panigrahy et al., 2015; Watson et al., 2018; Wernovsky, 2006). Other critical congenital heart defects such as tetralogy of Fallot (ToF) typically have a milder degree of hypoxia but can alter fetal and postnatal brain perfusion. In fact, infants with ToF not associated with a known genetic syndrome generally have a lower incidence of neurodevelopmental delay or disability compared to people with d-TGA or SV (P. D. Morton et al., 2017).

These alterations in fetal cerebral hemodynamics have the potential to impact brain development. A growing number of quantitative magnetic resonance imaging (MRI) studies have provided evidence of altered brain development, growth, and maturation starting during the fetal period (Lim et al., 2016; Miller et al., 2007; Sun et al., 2015). This includes in utero volumetric reduction in structures susceptible to cerebral hypoxia, which is most pronounced in forms of CHD that have large impairments in oxygenation, such as TGA (Lawrence et al., 2019; Peyvandi et al., 2018, 2019; Rollins et al., 2021). Generally, infants with CHD exhibit lower brain volumes, delayed brain development, and decreased cortical surface area and folding, all occurring prior to corrective surgery (Lynch et al., 2018; C. Ortinau et al., 2013; Rollins et al., 2021). MRI studies have also shown that neonates with SV have decreased cortical thickness and gyrification index postoperatively, indicating disruption to white matter development (Claessens et al., 2016; Masoller et al., 2016; C. M. Ortinau & Shimony, 2020).

The development of primary cortical pits and folds, together referred to as the sulcal pattern, occurs prior to the third trimester, and their pattern is relatively unchanged during childhood (Im & Grant, 2019), indicating that sulcal patterns could be utilized as an early biomarker to predict cognitive impairments (Chi et al., 1977; S. U. Morton et al., 2021; White et al., 2010). In fetuses with CHD, there are global alterations in sulcal patterning as indicated by a change in the overall gyrification index, which is accompanied by differences in positioning of the sulcal basins (C. M. Ortinau et al., 2018). Importantly, fetal changes in brain structure on MRI are correlated with neurodevelopmental outcomes later in childhood (Sadhwani et al., 2022). A positive association between Executive Function and sulcal patterns in the left hemisphere of people with SV has previously been demonstrated (S. U. Morton et al., 2020). Right sided sulcal pattern changes were positively associated with Executive Function outcomes in people with ToF (S. U. Morton et al., 2021). Even among people with simple forms of CHD that were repaired early in childhood, sulcal patterns are different than in the general population (Asschenfeldt et al., 2021). However, it remains unclear whether the differences in location of sulcal pattern abnormalities were primarily driven by the type of CHD or other patient factors such as genetic diagnoses. Therefore, we aimed to perform a meta-regression that compares brain MRI and neurodevelopmental outcomes in cohorts with discrete forms of CHD. This study can provide new insights into the relative magnitude of impact associated with factors such as hypoxia and impaired perfusion.

Materials and Methods

Study Design

For this retrospective meta-regression, clinical and research data were obtained from three previously published single-center cohort studies of single CHD types: SV, ToF and d-TGA (Bellinger et al., 2011; Bellinger, Rivkin, et al., 2015; Bellinger, Watson, et al., 2015). Non-CHD participants from the parent studies were included for comparison.

Participants

The study was approved by the Boston Children’s Hospital institutional review board. Participants for the d-TGA cohort were recruited between 1988 and 1992 and were aged (mean ± SD) 15.7 ± 0.8 at the time of MRI evaluation. In brief, inclusion criteria for the d-TGA cohort included diagnosis of d-TGA with intact ventricular septum or ventricular septal defect, scheduled repair by 3 months of age, and coronary anatomy suitable for an arterial switch operation. Exclusion criteria were birth weight less than 2.5 kg, a recognizable syndrome of congenital anomalies, an associated extracardiac anomaly of greater than minor severity, previous cardiac surgery, or associated cardiovascular anomalies requiring aortic arch reconstruction or additional surgical procedures. Participants for the SV cohort, aged 14.8 ± 3.0, were enrolled between 2010 to 2012. Inclusion criteria included age 10–19, diagnosis of SV, and history of the Fontan procedure. Exclusion criteria included history of cardiac transplantation or cardiac surgery within 6 months of neurodevelopmental testing. Participants for the ToF cohort, aged 14.4 ± 1.0, were enrolled between 2004 and 2008. Inclusion criteria included age 13–16, diagnosis of ToF, and a minimum 3-month gap between the latest cardiac operation and neurodevelopmental testing. Patients with trisomy 21 were excluded. A referent group of healthy, typically developing controls, aged 15.1 ± 1.9, was recruited using inclusion criteria adapted from the NIH MRI study of normal brain development (Evans, 2006). Exclusion criteria for the referent group included known risk factors for brain disorders and other forms of CHD requiring surgical correction (Bellinger et al., 2011).

Patient Factors

The Hollingshead Four-Factor Index was used to calculate participant socioeconomic status (Hollingshead, 1975) with greater scores indicating higher social status (range 11–66). Race/ethnicity was self-reported by study participants and/or their parents. Intraoperative variables included duration of deep hypothermic circulatory arrest (DHCA), history of cardiopulmonary bypass, and total support time at first surgery for the d-TGA and ToF groups. Perioperative complications include perioperative arrhythmia, extracorporeal membrane oxygenation, tamponade, open chest, infection, sepsis, organ failure, reoperation/re-exploration, and postoperative intensive care unit stay. Neurological event includes stroke, seizure, choreoathetosis, and meningitis.

Neurodevelopmental Assessments

Executive Function was assessed using the Delis-Kaplan Executive Function test (DKEFS), using a composite score of six age-normalized subscores: verbal letter fluency, verbal category fluency, primary combined design fluency, combined sorting condition score, word context total consecutively correct score, and total achievement score on Tower. The mean score is 10 with a standard deviation (SD) of 3, with a lower score implying more dysfunction. General Memory was assessed using the Children’s Memory Scale (Cohen, 1997) for participants 17 years or younger and the Wechsler Memory Scale–Fourth Edition (Wechsler, 2008) participants over 17 years of age. The test yielded an average score of 100 (SD 15). Processing Speed was characterized using the Wechsler Intelligence Scale for Children, Fourth Edition (Wechsler, 2003) with an average score of 100 (SD 15).

MRI Acquisition and Sulcal Pattern Analysis

In the d-TGA and ToF studies, MRI was acquired from a 1.5 Tesla GE Twin speed magnetic resonance scanner (Bellinger et al., 2011; Bellinger, Rivkin, et al., 2015; S. U. Morton et al., 2021). For the SV study, MRI was captured on one of two systems: a 3-T General Electric system (General Electric Medical Systems) or a 1.5-T General Electric Twin speed scanner for participants with implanted devices (S. U. Morton et al., 2020). Participants were scanned without sedation using a T1-weighted, 3D spoiled gradient recalled steady state sequence with the following parameters: TR/TE=7 ms/2.8 ms, flip angle=8°, acquisition matrix=256 × 256, FOV=256 mm, and slice thickness=1 mm at 3 T; TR/TE=40s/4s, flip angle=20°, acquisition matrix=256 × 192, FOV=240 mm, and slice thickness=1.5 mm at 1.5 T. Cortical surfaces were reconstructed from MRI images and parcellated into lobar regions utilizing the FreeSurfer pipeline as previously described (S. U. Morton et al., 2020). Global pattern of primary sulcal folds was characterized using a graph structure with sulcal pits and their catchment basins as nodes (Figure 1). Per-participant sulcal pattern graph was automatically compared to the graph of each non-CHD participant using a spectral-based graph matching algorithm (Im et al., 2011), and their sulcal pattern similarity value that ranged from 0 to 1 was calculated using the geometric features of the nodes (position, depth, and area of sulcal basin) and sulcal graph topology for the left and right hemispheres and lobes. Sulcal similarity values were generated by comparing each CHD or non-CHD participant to all non-CHD participants averaged. Lower mean sulcal similarity values reflect a larger difference in sulcal features for that individual compared with the non-CHD cohort, indicating a larger deviation from the typical sulcal folding pattern.

Figure 1.

Sulcal pattern similarity values are calculated by first determining the features of each individual MRI. Each participant with CHD is then compared to each participant from the non-CHD cohort, for a single sulcal feature. The mean similarity value is then calculated for comparison between individuals. CHD: congenital heart disease.

Statistical Approach

The associations between sulcal pattern features and neurodevelopmental outcomes were assessed using unadjusted correlation, linear regression, and a secondary regression tree analysis. First, we assessed the unadjusted correlation between sulcal similarity measures (left temporal, left frontal, right frontal, and right hemisphere) and neurodevelopmental outcomes (Executive Function, General Memory, and Processing Speed) that have a previously reported association in patients with CHD. Next, a multivariate linear regression was used to further investigate the strongest associations between sulcal pattern features and neurodevelopmental outcomes, after adjustment for demographic and medical characteristics. Independent predictors were selected using stepwise Akaike information criterion (AIC) selection (Yamashita et al., 2007) with candidate predictors including those described in Table 1 and various interactions between right hemispheric sulcal similarity and diagnosis group (SV, d-TGA, and ToF). All linear regression models were re-fit excluding individuals with a known genetic diagnosis. Sulcal measures were scaled to have mean of 0 and standard deviation (SD) of 1 relative to the entire sample. All p-values were adjusted utilizing the false discovery rate (FDR) to account for multiple testing (Benjamini & Hochberg, 1995). As a secondary analysis to explore potentially nonlinear relationships and interactions between variables, we used regression trees as a machine learning technique to develop classification models that may be more clinically useful (Riccio & Gomes, 2013). To construct the regression trees, we employed the standard method of recursive partitioning. This involves splitting the data based on predictors until we reached a stopping point (minimum terminal node size=10, complexity parameter=0.025), ensuring subsets with similar dependent variable values.

Table 1.

Baseline characteristics of study participants.

	All CHD (n=247)	SV (n=114)	d-TGA (n=92)	ToF (n=41)	Non-CHD (n=94)
Variables	Mean (SD), Number (%), or Median [IQR]
Gestational age, weeks	39.3 (2.1)	38.8 (2.4)	39.9 (1.3)	39.4 (2.4)	39.6 (1.2)
Self-reported white race	221 (89)	107 (94)	81 (88)	33 (80)	75 (80)
Family social class	48.4 (12.9)	48.9 (13.8)	47.1 (12.5)	50.4 (10.9)	53.4 (9.9)
Genetic diagnosis	16 (6)	10 (9)†	0	6 (15)‡	-
Age at first operation, days	7 [4, 47]	6 [3, 27]	6 [4, 9]	163 [76, 337]	-
DHCA duration, minutes	35 (25)*	39 (29)§	36 (22)	26 (26)**	-
Total support time, minutes	129 (40)*	125 (49)§	142 (31)	96 (27)**	-
Number of perioperative complications	1 [0, 2]	1 [0, 2]	1 [0, 2]	0 [0, 1]	-
Any neurological event	51 (21)	26 (23)	20 (22)	5 (12)	-
Number of catheterizations	2 [1, 4]	4 [3, 5]	1 [0, 1]	1 [0, 2]	-
Number of operations	2 [1, 3]	3 [3, 3]	1 [1, 1]	1 [1, 2]	-
Age at MRI, years	15.0 (2.2)	14.8 (3.0)	15.7 (0.8)	14.4 (1.0)	15.1 (1.9)

CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest, d-TGA: d-transposition of the great arteries, MRI: magnetic resonance imaging, SV: single ventricle, ToF: Tetralogy of Fallot

All variables are missing <3% unless otherwise noted.

^*Data is missing for 23 out of 247 participants (9%) and not relevant for 49 participants undergoing closed surgery.

^†Previously known genetic/phenotypic syndromes included VACTERL (n=1), VATER (n=1), 22q11 (n=1), CHARGE syndrome (n=1), Kartagener syndrome (n=1), Factor V Deficiency (n=1), Factor V Leiden Deficiency (n=1), Factor VII Deficiency (n=1), balanced translocation involving chromosomes 13 and 14 (n=1) and unspecified (n=1).

^§Data missing for 14 of 65 participants (22%) undergoing open surgery and not relevant for 49 participants undergoing closed surgery.

^‡Previously known genetic/phenotypic syndromes included 22q11 (n=1), VATER (n=2), VACTERL (n=2), and Alagille syndrome (n=1).

^**Data missing for 9 out of 41 participants (22%).

To evaluate predictive performance and model fit, proportion of the variance explained by the model (R²) and the root mean squared error (RMSE) were calculated. To prevent overfitting and ensure accurate outcomes in the regression trees, we used standard machine learning evaluation metrics; namely Cross-validation RMSE and bootstrapped trees. These methods aid in evaluating regression tree performance by assessing their predictive accuracy on new, unseen data (Cross-validation) and subjecting them to diverse data subsets (bootstrapping). Computation of Cross-validation RMSE involves randomly dividing the data into ten equal-sized sets and repeatedly performing the fitting procedure, excluding 10% of the data, or one “set”, each time (Efron & Hastie, 2021; Hastie et al., 2009; James et al., 2013). The fitting procedure is performed on the remaining 90% of the data and the RMSE is computed using the excluded 10% as a testing set. We then average over the ten computed RMSEs to compute a Cross-validation RMSE; if the Cross-validation RMSE is substantially higher than the RMSE, overfitting is likely. To assess the stability of the predictors selected for the regression trees, we generated 1000 bootstrapped samples (sampling with replacement from the original sample stratified by diagnosis group) and reported which predictors occurred more than 10% of the time as either a first level or second level tree split. This helps identify consistent and influential variables because those that consistently appear early in the tree generally have a strong impact on the tree’s predictions. All statistical analyses were conducted in R, version 4.3.1 (R Foundation for Statistics and Computing), with the regression tree analyses performed using the rpart package.

Results

Participant traits

Three-hundred forty-one participants were analyzed, including 247 with CHD (114 SV, 92 d-TGA, and 41 ToF) and 94 participants without CHD (Table 1). Participants were predominantly born at term and the average participant age at MRI evaluation was 14.4–15.7 across the groups. Participants were predominantly self-reported white race (87%) and male (62%). The median age at the first operation was 6 days for both the SV and d-TGA groups, and 163 days for the ToF cohort. Across all CHD cohorts, the DHCA duration was 34.7 ± 25.3 (mean ± SD) minutes, while the total support time was 129.0 ± 40.4. DHCA duration and total support time were both lower in participants with ToF. Sulcal pattern similarity scores were lowest in the SV cohort, and Executive Function and General Memory scores were similar between the CHD groups (Table 2).

Table 2.

Sulcal and neurodevelopmental measures by cohort. Only sulcal pattern features and neurodevelopmental measures that had been previously found to have an association in one of the CHD cohorts were included.

	All (n=341)	SV (n=114)	d-TGA (n=92)	ToF (n=41)	Non-CHD (n=94)
Variables	Mean (SD)
Left temporal sulcal similarity	0.7468 (0.0143)	0.7444 (0.0140)	0.7464 (0.0153)	0.7451 (0.0145)	0.7508 (0.0128)
Left frontal sulcal similarity	0.7603 (0.0095)	0.7585 (0.0103)	0.7609 (0.0099)	0.7604 (0.0089)	0.7617 (0.0081)
Right frontal sulcal similarity	0.7614 (0.0096)	0.7593 (0.0118)	0.7623 (0.0083)	0.7609 (0.0090)	0.7631 (0.0074)
Right hemisphere sulcal similarity	0.7511 (0.0084)	0.7487 (0.0108)	0.7526 (0.0057)	0.7496 (0.0078)	0.7533 (0.0067)
Executive Function	9.4 (2.2)	8.6 (2.3)	9.4 (2.0)	9.0 (2.5)	10.8 (1.6)
General Memory	95.3 (17.7)	91.9 (15.9)	91.9 (18.0)	93.8 (21.4)	103.7 (15.1)
Processing Speed	95.9 (16.4)	88.2 (16.5)	100.1 (14.2)*	94.1 (17.7)	102.6 (13.4)

CHD: congenital heart disease, d-TGA: d-transposition of the great arteries, SV: single ventricle, ToF: Tetralogy of Fallot

All variables are missing <3% unless otherwise noted.

^*Data missing for 13 out of 92 participants (14%).

Association of sulcal pattern with neurodevelopmental outcomes via linear regression

In the unadjusted analysis, higher right hemisphere sulcal pattern similarity values were correlated with better Executive Function (Pearson r=0.19, FDR p=0.005; Figure 2, Table 3), General Memory (Pearson r=0.15, FDR p=0.02), and Processing Speed (Pearson r=0.17, FDR p=0.01). General Memory was also positively associated with right frontal sulcal pattern similarity (Pearson r=0.13, FDR p=0.05). A multivariate model (Table 4) that included participant characteristics demonstrated a persistent association of increased right hemisphere sulcal pattern similarity with better function in the three neurodevelopmental domains in the d-TGA and ToF cohorts, but not in the SV cohort. Family social class, DHCA duration, history of neurologic events such as a stroke, and number of cardiac catheterizations, were also associated with Executive Function scores. The same factors, except for number of cardiac catheterizations, were associated with General Memory, while d-TGA status, family social class, DHCA duration were associated with Processing Speed. When individuals with a genetic diagnosis were removed (Table 5), the coefficient for associations between right hemisphere sulcal similarity and the three neurodevelopmental outcomes decreased in the d-TGA and ToF cohorts, with the General Memory association losing statistical significance. Further, the number of catheterizations was no longer associated with Executive Function when individuals with genetic diagnoses were removed.

Figure 2.

Associations between right hemisphere sulcal similarity and neurodevelopmental outcomes (A: Executive Function, B: General Memory, C: Processing Speed) for SV (red), d-TGA (green), ToF (blue) and non-CHD (purple) participants in the entire sample (n=341) for individuals with (triangles) and without (dots) a genetic diagnosis. Pearson r for the respective association is written in upper left-hand corner of each plot. Linear regression line of best fit (black) with 95% confidence interval (grey) is shown for each relationship. CHD: congenital heart disease, d-TGA: d-transposition of the great arteries, SV: single ventricle, ToF: Tetralogy of Fallot.

Table 3.

Correlation between neurodevelopmental outcomes and sulcal pattern similarity in the entire sample (n=341). Only sulcal pattern features and neurodevelopmental measures that had been previously found to have an association in one of the CHD cohorts were included.

	Sulcal Pattern Similarity Measures Pearson r (FDR p)
	Left Temporal	Left Frontal	Right Frontal	Right Hemisphere
Executive Function (n=337)	0.12 (0.06)	0.06 (0.31)	0.11 (0.10)	0.19 (0.005*)
General Memory (n=338)	0.07 (0.27)	0.08 (0.21)	0.13 (0.05*)	0.15 (0.02*)
Processing Speed (n=326)	0.10 (0.11)	0.02 (0.76)	0.10 (0.11)	0.17 (0.01*)

^*FDR p < 0.05, CHD: congenital heart disease, FDR: false discovery rate, SV: single ventricle, ToF: Tetralogy of Fallot

Table 4.

Multivariable associations between neurodevelopmental outcomes and characteristics in the CHD cohort. Sulcal measures have been scaled to have mean 0 and standard deviation 1 in the entire sample. Data are presented as β ± SE (FDR p).

	Executive Function (n=171)	General Memory (n=173)	Processing Speed (n=161)
	β ± (FDR p).
Right hemisphere sulcal similarity (d-TGA and ToF)	0.7 ± 0.2 (0.004*)	4.1 ± 1.9 (0.03*)	5.4 ± 1.7 (0.003*)
Right hemisphere sulcal similarity (SV)	−0.2 ± 0.2 (0.55)	−2.5 ± 1.8 (0.24)	−0.9 ± 1.5 (0.64)
d-TGA (vs. SV)	0.2 ± 0.4 (0.64)	−0.7 ± 4.3 (0.88)	12.4 ± 3.8 (0.004*)
ToF (vs. SV)	−0.1 ± 0.5 (0.88)	1.9 ± 4.6 (0.73)	7.9 ± 4.0 (0.08)
Self-reported white race	1.0 ± 0.5 (0.06)	-	-
Family social class	0.03 ± 0.01 (0.005*)	0.41 ± 0.09 (<0.001*)	0.29 ± 0.08 (0.002*)
Age at first operation, days	-	-	−0.01 ± 0.01 (0.38)
DHCA duration, minutes	−0.02 ± 0.01 (0.002*)	−0.15 ± 0.05 (0.006*)	−0.23 ± 0.05 (<0.001*)
Number of perioperative complications	−0.19 ± 0.10 (0.08)	-	-
Any neurological event	−0.9 ± 0.4 (0.02*)	−9.4 ± 3.0 (0.006*)	−4.0 ± 2.6 (0.19)
Number of catheterizations	−0.26 ± 0.09 (0.01*)	-	-
Number of operations	-	−2.1 ± 2.0 (0.38)	−1.8 ± 1.7 (0.38)
R2	0.38	0.28	0.42
RMSE	1.7	15.2	12.7
Cross-validation RMSE	1.9	16.9	13.3
% Change in RMSE	9.2	11.5	4.7

^*FDR p < 0.05, CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest, d-TGA: d-transposition of the great arteries, FDR: false discovery rate, RMSE: root mean squared error, SE: standard error, SV: single ventricle, ToF: Tetralogy of Fallot

Table 5.

Multivariable associations between neurodevelopmental outcomes and characteristics in the CHD cohort without genetic diagnoses. Sulcal measures have been scaled to have mean 0 and standard deviation 1 in the entire sample. Data are presented as β ± SE (FDR p).

	Executive Function (n=163)	General Memory (n=164)	Processing Speed (n=152)
	β ± (FDR p).
Right hemisphere sulcal similarity (d-TGA and ToF)	0.6 ± 0.2 (0.02*)	2.7 ± 2.0 (0.14)	4.7 ± 1.8 (0.01*)
Right hemisphere sulcal similarity (SV)	−0.2 ± 0.2 (0.52)	−2.9 ± 1.8 (0.17)	−0.9 ± 1.6 (0.62)
d-TGA (vs. SV)	0.5 ± 0.4 (0.39)	3.4 ± 4.6 (0.55)	14.4 ± 4.3 (0.004*)
ToF (vs. SV)	0.2 ± 0.5 (0.70)	6.1 ± 5.0 (0.31)	9.0 ± 4.3 (0.07)
Self-reported white race	1.0 ± 0.5 (0.07)	-	-
Family social class	0.03 ± 0.01 (0.007*)	0.39 ± 0.09 (<0.001*)	0.28 ± 0.08 (0.004*)
Age at first operation, days	-	-	-0.01 ± 0.01 (0.61)
DHCA duration, minutes	−0.02 ± 0.01 (0.003*)	−0.15 ± 0.05 (0.01*)	−0.23 ± 0.05 (<0.001*)
Number of perioperative complications	−0.2 ± 0.1 (0.14)	-	-
Any neurological event	−1.0 ± 0.4 (0.02*)	−10.6 ± 3.1 (0.003*)	−4.3 ± 2.7 (0.17)
Number of catheterizations	−0.2 ± 0.1 (0.14)	-	-
Number of operations	-	0.0 ± 2.2 (0.99)	−0.4 ± 2.0 (0.88)
R2	0.35	0.28	0.40
RMSE	1.7	14.8	12.7
Cross-validation RMSE	1.9	16.2	13.7
% Change in RMSE	10.4	8.3	7.5

^*FDR p < 0.05, CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest, d-TGA: d-transposition of the great arteries, FDR: false discovery rate, RMSE: root mean squared error, SE: standard error, SV: single ventricle, ToF: Tetralogy of Fallot

Regression Tree Analysis

A regression tree analysis for Executive Function scores demonstrated DHCA duration and family Hollingshead index score as the strongest factors (Figure 3). The highest level of impairment was seen in the 18 participants with a DHCA duration of 64 minutes of longer (mean Executive Function score=6.2). By contrast, the 31 participants with a DHCA duration less than 64 minutes and a family social class score of 60 or greater had a mean score of 10. Other factors included in the regression tree included cardiac catheterizations, right hemisphere sulcal pattern similarity, and gestational age at birth. Resampling of the data for the regression tree demonstrated persistent inclusion of family social class, DHCA duration, number of catheterizations, and right hemisphere sulcal similarity in the first or second levels of the tree (Table 6).

Figure 3.

Regression tree predicting Executive Function (R²=0.38, root mean squared error (RMSE)=1.7, Cross-validation RMSE=2.1) in the CHD sample (n=168). CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest.

Table 6.

Results of 1000 bootstrap resamples for each regression tree. For first level, the number in the box is the percentage of bootstrap samples in which the variable is chosen to be in the first level of the tree. For the second level, the number in the box is the percentage of bootstrap samples in which the variable is chosen to be in the second level of the tree.

	Executive Function (n=168)		General Memory (n=170)		Processing Speed (n=158)
Variable	First level (%)	Second level (%)	First level (%)	Second level (%)	First level (%)	Second level (%)
Right hemisphere sulcal similarity	0.4**	13.4	0.4	3.4	0.7**	9.6
d-TGA (vs. non-d-TGA)	0	0.3	0	0.2	7.1	3.4
ToF (vs. non-ToF)	0	0.1	0	0.9	0	0
Gestational age	1.3**	7.3	1.1	2.5	0.2	1.9
Self-reported white race	0.1	0.2	0	0	0	0
Family social class	2.9*	26.1	28.5*	37.5	0.5*	8.7
Genetic diagnosis	0	0	0	0	0	0
Age at first operation	0	5.8	0.8	7.5	0	2.8
DHCA duration	77.8*	18.1	49.1*	19.9	22.0**	19.2
Total support time	0.2	1.7	0.9**	3.1	0	3.8
Number of perioperative complications	0.1	1.6	1.5	2.9	0	0.4
Any neurological event	1	2.6	14.5**	15.6	0	3.2
Number of catheterizations	13.3**	16.7	2.9	3.8	18.9	10.5
Number of operations	2.8	3.5	0.3	0.3	45.0*	7.7
Age at MRI	0.1	2.6	0**	2.4	5.6*	28.8

^*Included in first or second level of presented regression tree (Figures 3–6).

^**Included in third level or lower of presented regression tree (Figures 3–6).

DHCA: deep hypothermic circulatory arrest, d-TGA: d-transposition of the great arteries, MRI: magnetic resonance imaging, ToF: Tetralogy of Fallot

The regression tree for General Memory also included DHCA and family social class in the first two levels (Figure 4). The 12 participants with DHCA duration of 67 minutes or more had the highest level of impairment (mean score=71), while the 81 participants with DHCA duration less than 67 minutes and family social class score of 50 or more had the highest General Memory scores (mean score=98). History of neurologic event such as stroke, age at MRI, and total support time during the first operation were also included in the regression tree. Of these, only family social class, DHCA duration, and history of neurologic event were more likely to be included in the first two levels of the regression tree (Table 6).

Figure 4.

Regression tree predicting General Memory (R²=0.28, root mean squared error (RMSE)=15.1, Cross-validation RMSE=17.9) in the CHD sample (n=170). CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest, MRI: magnetic resonance imaging.

In contrast, the first level of the regression tree for Processing Speed scores was the number of operations (Figure 5). The 17 participants aged 16 or older with three or more operations had the lowest Processing Speed scores (mean score=74), while the 54 participants with fewer than 3 operations and family social class score of 51 or more had the highest Processing Speed scores (mean score=104). Other factors included in the regression tree were right hemisphere sulcal pattern similarity and DHCA duration. In the bootstrap resampling, number of operations, age at MRI, DHCA duration, and number of catheterization most often occurred in the first or second level (Table 6).

Figure 5.

Regression tree predicting Processing Speed (R²=0.42, root mean squared error (RMSE)=12.8, Cross-validation RMSE=15.2) in the CHD sample (n=158). CHD: congenital heart disease, DHCA: deep hypothermic circulatory arrest, MRI: magnetic resonance imaging.

Discussion

We used linear regression and machine learning techniques to assess the association of patient characteristics and sulcal pattern features with Executive Function, General Memory, and Processing Speed in participants with SV, d-TGA, and ToF. In the overall cohort, higher right hemispheric sulcal pattern features were associated with improved Executive Function, General Memory, and Processing Speed scores. This is consistent with studies of unilateral stroke among non-CHD participants, where attention and Executive Function were more severely impacted among participants with right hemisphere injury compared to those with left hemisphere injury (Shimonaga et al., 2021). However, when the right hemisphere data from the SV group was considered separately from the d-TGA and ToF groups in an adjusted multivariate linear regression, only the latter had a significant positive association with Executive Function. This indicates that confounders, such as DHCA duration, may impact the relationship between right hemispheric sulcal pattern similarity and neurocognitive outcomes within the three CHD types. Further, when individuals with a genetic diagnosis were removed from the analysis, the association between General Memory and right hemisphere sulcal pattern similarity remained positive but was no longer statistically significant. This is consistent with sulcal pattern being significantly influenced by genetic factors. Regression tree analyses, which account for potential non-linear interactions between predictors, indicated that right hemispheric sulcal pattern was positively associated with both Executive Function and Processing Speed scores. Bootstrap resampling showed that these results were stable when predicting Executive Function. The regression tree analyses may be viewed as secondary or exploratory as they exhibited a higher degree of overfitting than the multivariate linear regression models.

This analysis builds upon previous studies that have also assessed differences in brain MRI traits among the d-TGA cohort. In a previous analysis of this d-TGA cohort, greater abnormalities in the white matter structural networks, including changes to global efficiency and modularity, were associated with Executive Function scores (Panigrahy et al., 2015). Another study on this d-TGA cohort also reported that reductions in network connectivity among regional subnetworks in the left and right hemispheres were associated with increased attention-deficit hyperactive symptoms (Schmithorst et al., 2016). Our findings build on those results by localizing the impact of the sulcal pattern changes that correlate with Executive Function to the right hemisphere. These results are consistent with a study on white matter microstructure and neurodevelopment with this d-TGA cohort, where Executive Function correlated with right precentral white matter microstructure as measured by fractional anisotropy (Rollins et al., 2014). Further work in other cohorts with d-TGA would be important to assess the consistency of these findings. However, participants with ToF also demonstrated a consistent positive association of right-sided sulcal pattern features with neurodevelopmental measures. While the right hemispheric sulcal pattern correlated with all neurodevelopmental outcome measures in the entire CHD sample, the association only remained among the d-TGA and ToF participants after adjustment for multiple patient features. Further, when individuals with a known genetic diagnosis were removed, the association between right hemisphere sulcal pattern similarity and General Memory was no longer statistically significant. Given the high prevalence of genetic contributions to ToF, this could support assessment of sulcal patterns differences specifically among people who have ToF and genetic diagnoses (Althali & Hentges, 2022). This difference could be due to the relatively larger impact of factors such as DHCA duration and other postnatal influences on the sulcal patterns of SV participants, which may outweigh prenatal influences. However, compared with the d-TGA cohort, the SV cohort had similar DHCA duration, total support time, and rate of neurological events. The two groups also had relatively similar family social class and age at MRI. Given these broad similarities, there may be unmeasured factors, such as genetic risk and degree or duration of prenatal and postnatal hypoxia, that may be modifying the relationship between sulcal pattern and Executive Function. However, a higher number of operations, which could reflect a longer duration of hypoxia, was consistently selected as a predictor of Processing Speed in the regression tree analysis. One MRI study has demonstrated a 15% reduction in fetal cerebral oxygen delivery in participants with mixed types of CHD (d-TGA, ToF, and hypoplastic left heart syndrome), which was associated with a 13% decreased in brain volumes despite relative preservation of blood flow at the time of imaging (Sun et al., 2015). The shared association between sulcal pattern and neurodevelopmental disability was observed in both d-TGA and ToF cohorts (Lee et al., 2021), despite only the former having severe fetal hypoxia. This suggests that other factors in addition to severe fetal cerebral hypoxia led to alterations in sulcal pattern development and subsequent Executive Function impairment, Further, the lack of right hemisphere sulcal pattern changes in the SV cohort studied here indicates that unmeasured factors such as differences in placental and cerebral vascular resistance may be important mediators of the impact of fetal hypoxia (Campbell et al., 2019; I. A. Williams et al., 2012).

While sulcal pattern features are a potential biomarker for neurodevelopmental risk in people with CHD, they are one piece of a larger model. Other factors, such as longer duration of DHCA and lower family social class as measured by the Hollingshead index, were consistently associated with lower neurodevelopmental testing scores. In the general population, environmental factors such as poverty, poor nutritional status, and stress have been associated with cortical and subcortical volumes and surface area (McDermott et al., 2019). Maternal stress during infancy was shown to be significant predictors of differential methylation in adolescents, further supporting the potential for such factors to modify developmental health at a molecular level (Essex et al., 2013). Similarly, maternal education has been demonstrated to be positively associated with literacy and cognition in children with CHD (Gaynor et al., 2015; Mulkey et al., 2016). Other MRI measures, such as amygdala volume, are negatively correlated with Executive Function in children with SV (Sato et al., 2022). Therefore, larger cohorts are needed to be able to consider all these potential contributions to neurodevelopmental outcomes among people with CHD to best develop interventions that address these important contributors to health (Als et al., 1982; Lisanti et al., 2023). History of neurologic event such as postoperative seizure was also predictive of Executive Function in our cohort, whereas acquired injury is not consistently associated with outcomes in other cohorts. However, it may be that in our cohort with varied types of CHD and participant ages, the history of neurologic injury is also reflective of any preoperative immaturity due to brain maldevelopment that contributes to white matter injury risk (Miller et al., 2007; C. M. Ortinau & Shimony, 2020; Peyvandi et al., 2019; Peyvandi & Rollins, 2023).

Limitations to our study included differences in time periods and parent study protocols, limited diversity of self-reported race and family social class, as well as the single center design, all of which potentially decreased the generalizability of our results. A small number of individuals did have genetic diagnoses, but genetic testing was not uniformly conducted across the cohort. We did not correct for scanner strength as it was confounded by CHD type due to the different MRI protocols in each original study. Sulcal pit extraction, graph construction, and pattern analysis are highly reliable across scan sessions, sites, and magnet strengths (Im et al., 2013; S. U. Morton et al., 2020). The participants with SV are a heterogeneous group and include patients who have had Norwood procedures associated with periods of DHCA. Some clinical variables, such as whether participants with ToF had significant cyanosis prior to surgical repair, were not available. Further, surgical and medical care has changed since the three studies were conducted. DHCA durations over 60 minutes are less common in current surgical approaches, so the impact of DHCA may be different in cohorts with younger participants. Additionally, while the Hollingshead Four Factor index was uniformly collected in all our cohorts and is frequently used as a measure of socioeconomic status in research studies, there are several limitations to its use that need to be considered. The Hollingshead index relies heavily on education and occupation levels, disregarding other crucial socioeconomic indicators like income, wealth, and social capital. This may lead to misclassification of individuals and biased estimates of the relationship between socioeconomic status and health outcomes. Additionally, the Hollingshead index is self-reported, which can result in reporting bias and measurement error, but we did not have maternal education information available for these cohorts (Cirino et al., 2002; Hollingshead, 1975).

In conclusion, our study demonstrated a positive association between the right hemisphere sulcal similarity and Executive Function, General Memory, and Processing Speed among participants with d-TGA and ToF. DHCA duration was also associated with Executive Function across the entire CHD cohort. This finding has important implications for treatments and prevention measures to reduce effects of d-TGA and ToF on Executive Function and highlights the importance of these early fetal and neonatal periods for potential interventions that enhance resilience and minimize stress. The association between DHCA duration and Executive Function suggests that minimizing DHCA duration could be beneficial. However, significant technological advances would be needed to facilitate shorter DHCA times. Further, many inherent patient factors such as body weight can impact the duration of DHCA, and it can be difficult to encourage growth with unrepaired CHD. This highlights the potential for sulcal pattern features, which are generally established early in life, to serve as biomarkers for later neurodevelopmental challenges. The utility of routine childhood brain MRI to identify individuals at high risk of poorer neurodevelopmental outcomes, in order to support targeted interventions, is an important future direction. Further, advancements in sulcal pattern analysis are needed to facilitate more general consideration of sulcal pattern during clinical brain imaging. These findings also suggest the value in making effort to reduce social disparities, which could be done by increasing prenatal and postnatal parental support to families with children with CHD. Early intervention services to enhance family support of child development could be one such intervention. Addition work across multiple centers, enrolling participants before the teenage years studied here, is necessary to validate these findings and update targeted therapeutic interventions for children with complex CHD.