Machine learning to infer neurocognitive testing scores among adolescents and young adults with congenital heart disease

Abstract

Background

Congenital heart disease (CHD) affects about 1% of births and is linked to differences in thinking and learning. Understanding how birth, genetic, clinical, and environmental factors together explain cognitive variability can inform monitoring and care. This study builds a multivariate model predicting cognition across multiple domains in adolescents and young adults with CHD.

Methods

We studied 89 adolescents and young adults (AYAs; mean age 16 years) with CHD who completed structural and diffusion MRI and fifteen neurocognitive tests across seven domains. Using an enhanced forward-inclusion and backward-elimination strategy with cross-validation, we built multivariate models incorporating biological, socioeconomic, clinical, genetic, and brain imaging features. Performance was evaluated using Pearson correlation ($r$) between observed and inferred scores, mean absolute error (MAE), and inverse inferability score (IIS).

Results

Here we show that models infer scores with moderate accuracy ($r$ = 0.245–0.648; MAE = 1.6–12.0 points; mean MAE = 6.3). Highest correlations include Digit Span ($r$ = 0.65; p < 0.001), Verbal Comprehension Index ($r$ = 0.594; p < 0.001), and Matrix Reasoning ($r$ = 0.574; p < 0.001). Domain ranking by IIS shows the best (lowest) scores for general intelligence (0.0886), followed by working memory (0.7100), and a higher (worse) score for perceptual reasoning (1.9199).

Conclusions

A multivariate approach combining brain imaging with genetic, clinical, and environmental factors provides clinically meaningful inference of individual cognitive performance in AYAs with CHD. These findings suggest complementary roles of brain, genetic, and contextual factors in shaping cognitive variability and motivate validation in larger cohorts.

Plain Language Summary

Children born with congenital heart disease can have differences in thinking and learning. We aimed to learn which factors, brain structure, genetics, medical history, and family environment, explain these differences in teens and young adults. We combined brain scans with health, family, and genetic information and used computer models to estimate scores on standard tests. The models reasonably estimated individual scores; some abilities, such as general thinking skills, working memory, and processing speed, were estimated more accurately than others. These findings suggest that brain scans, together with medical and family information, can provide a clearer picture of thinking skills. With further testing in larger, diverse groups, this approach could help guide follow‑up care tailored to each person.

Hussain et al. build multivariate machine-learning models integrating brain MRI with genetic, clinical and socioeconomic data to infer 15 neurocognitive test scores in adolescents and young adults with congenital heart disease. The models achieve moderate accuracy and highlight joint brain, genetic, and contextual contributions.

Introduction

Congenital heart disease (CHD) is the most prevalent major congenital anomaly, affecting approximately 1% of infants^1,2. With advances in medical and surgical care, around 85% of individuals with CHD now survive into adulthood^3–7. In the United States, an estimated 30,000 infants with CHD are born each year and have a normal life expectancy. Despite these improvements in survival rates, nearly 50% of these individuals experience neurocognitive impairments that manifest by adolescence^8–13 and young adulthood^14–17. Recent studies have raised concerns about the potential for accelerated brain aging in this population (adolescents and young adults [AYAs]; ages 8–30)¹⁸, with an elevated risk of early-onset dementia in adults with CHD (ages 30–80 + )^19,20.

Therapeutic strategies such as providing psychoeducation^21,22, engaging community healthcare providers²³, and implementing home-based cognitive training programs^24,25 have shown promise in promoting neuroresiliency. However, a thorough understanding of various factors contributing to neurodevelopmental impairments remains lacking^26–29. Existing studies found contributions from clinical variables such as the number of cardiac interventions³⁰, CHD type^31,32, genetic variables such as the presence of predicted damaging de novo variants or loss-of-function (LoF) variants in chromatin-modifying genes³³, or conventional MRI metrics such as volumetric measures^14,34. However, models that consider the combinatorial effects of these factors are needed^35–41, as current models explain only a fraction of the variability in individual neurocognitive testing scores^42–44.

Here, we use comprehensive brain MRI features, demographics, socioeconomics, and clinical and genetic factors, aiming to build a multivariate machine learning model to infer neurocognitive testing scores in AYAs with CHD from the comprehensive 7-site Pediatric Cardiac Genomics Consortium (PCGC) dataset (NIH UG1HL135685 and U01HL098147, PI: Newburger, 2009–2020). The tasks are twofold: (1) to identify the subset of variables that jointly contribute most to the inference of fifteen neurocognitive test scores, and (2) to rank functions across seven cognitive domains by inference accuracy. In this study, the models infer individual test scores with moderate accuracy (correlations up to 0.65; mean absolute error in the range of a few to a dozen points) and consistently select brain, genetic, clinical, and contextual variables as joint contributors. Cognitive domains differ in inferability, with general intellectual ability and working memory estimated more accurately than language and perceptual reasoning. These findings indicate that integrating brain imaging with clinical and environmental factors provides clinically meaningful estimates of cognitive performance and motivates validation in larger and more diverse cohorts.

Methods

Data

Brain MRI, demographic, socioeconomic, and genetic data were available for 89 participants from the PCGC CHD Brain and Genes study (clinicaltrials.gov NCT03070197). This study was conducted as a multicenter, cross-sectional investigation to examine associations between damaging de novo variants and neurodevelopmental as well as brain MRI outcomes in individuals with CHD. The study design and protocol were developed by investigators from the PCGC and received institutional review board approval at all participating sites and the administrative coordinating center, with reliance agreements in place across centers. Written informed consent was obtained from adult participants or from parents or legal guardians of minors, and written assent was obtained from older children when appropriate. Oversight of the study was provided by an independent data and safety monitoring committee appointed by the National Heart, Lung, and Blood Institute. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines⁴⁵ for reporting observational studies and was carried out between September 2017 and June 2020 across eight centers in the United States. The Boston Children’s Hospital institutional review board further approved this study (IRB P00039087 and P00023574). In Table 1, we present the collection site, demographic, socioeconomic, genetic, parental, and diagnosis details of this dataset.

Table 1.

Participant characteristics (N = 89)

Feature Type	Features/Attributes		Number of Subjects/Range/Mean
Demographics	Age at MRI (years)		7.94–29.91 (Mean: 15.93 $\mathit{\pm }$6.41)
	Age at Neurocognitive Assessment (years)*		7.93–29.90 (Mean: 15.92 $\mathit{\pm }$6.41)
	Sex		Male: 52 (59%); Female: 37 (41%)
	Height (cm)		118–198 (Mean: 152.44 $\mathit{\pm }$18.91)
	Weight (kg)		21.3–141.7 (Mean: 55.07 $\mathit{\pm }$25.84)
	Body Mass Index (BMI; kg/m2)		13.63–61.33 (Mean: 22.60 $\mathit{\pm }$7.86)
Diagnosis & Treatment (Clinical)	Single Ventricles with Arch Obstruction		4 (4%)
	Single Ventricles without Arch Obstruction		17 (19%)
	Bi-ventricle with Arch Obstruction		9 (7%)
	Bi-ventricle without Arch Obstruction		62 (70%)
	Total Number of Cardiac Surgeries	0	12 (13%)
		1	25 (28%)
		2	16 (18%)
		3	24 (27%)
		4	11 (12%)
		5	1 (1%)
Genetics	De novo variant (case) / No de novo variant (control)		46/43
	Presence of the LoF variant in a gene expressed in the upper quartile of genes in the developing brain		Yes: 53, No: 33, N/A: 3
	Presence of the LoF variant in chromatin-modifying gene		Yes: 8, No: 78, N/A: 3
	Presence of the LoF variant in known neurodevelopmental disorder gene		Yes: 8, No: 78, N/A: 3
	Presence of the LoF variant in constrained gene, defined as having a probability of LoF intolerance score > 0.5		Yes: 34, No: 52, N/A: 3
	Presence of ApoE Alleles*		ε3/ε3: 33, non-ε3/ε3: 56
Socioeconomics	Mother’s Education	Kindergarten-6th Grade	1 (1%)
		High School Graduate	8 (9%)
		Partial College, 2-Year Diploma, or Trade School	18 (20%)
		3- or 4-Year College/University Graduate	34 (38%)
		Post Graduate Degree	25 (28%)
		Others	1 (1%)
	Father’s Education	Kindergarten-6th Grade	1 (1%)
		High School Graduate	17 (19%)
		Partial College, 2-Year Diploma, or Trade School	16 (18%)
		3- or 4-Year College/University Graduate	28 (31%)
		Post Graduate Degree	24 (27%)
		Others	1 (1%)
	Total Family Income	< $24,999	3 (3%)
		$25,000–$49,999	7 (8%)
		$50,000–$74,999	8 (9%)
		$75,000–$99,999	9 (10%)
		$100,000–$149,999	15 (17%)
		> $149,999	23 (26%)
		N/A	24 (27%)
MRI-derived Covariates	AI–predicted brain age from structural T1-weigheted brain MRI18 (years)		6.53–35.7 (Mean: 17.30 $\mathit{\pm }$7.35)
	Accelerated brain age [deltaAGE18 = AI–predicted age - actual age] (years)		Mean: 1.37$\mathit{\pm }$3.46

IQR interquartile range, p25 25th percentile, p75 75th percentile, N/A Not available, LoF loss of function, AI artificial intelligence. (*Not used as a feature in the present study).

Inclusion/exclusion criteria

PCGC CHD cohort had N = 197 participants. They had a confirmed CHD diagnosis and were 8 years old and up. They completed study documents in English or Spanish and had available genetic sequencing data. Additionally, informed consent or parental permission and child assent were acquired for participation. They did not have a history of cardiac transplant, cardiac surgery within six months before enrollment, a known clinical genetic syndrome, or a brain injury deemed by a center investigator to confound the impact of genetic variants on neurodevelopmental outcomes significantly. Participants were excluded if they could not undergo MRI scanning due to an inability to remain still or intolerance of the confined space. This resulted in a sample size of N = 100. As MRI completion required participants to tolerate the scan without sedation, this could exclude individuals with anxiety or difficulty with state regulation (see Fig. 1b). While this specific comparison was not conducted in the present analysis, prior observations suggest that those who completed the MRI tended to have stronger neurodevelopmental performance. After an MRI quality check, some additional participants were excluded due to the low quality of images, resulting in an analysis sample size of N = 89.

Fig. 1. Study Overview and Participant Selection Process.

a Graphical abstract showing the study methodology from data collection through machine learning analysis to accuracy analysis. b Participant selection flow chart showing exclusion criteria that led the initial PCGC cohort (N = 197) to the final analysis sample (N = 89).

Non-MRI and MRI-derived covariates

17 different non-MRI variables were recorded and used in this study (see Table 1): (1) presence or absence of de novo variants in genes not previously associated with neurodevelopmental disability, defined by exclusion of individuals with known pathogenic variants in neurodevelopmental disorder- or CHD-associated genes per curated gene lists and literature (see Ref. ³³), (2) age at neurocognitive testing, (3) sex, (4) height, (5) weight, (6) Body Mass Index (BMI), (7) CHD type, (8) the total number of cardiac surgeries the participant underwent, (9) artificial intelligence (AI)–predicted brain age from structural T1-weigheted brain MRI¹⁸, by our deep learning algorithms^46–48, (10) accelerated brain age [deltaAGE¹⁸ = AI–predicted age - actual age], (11) the LoF variant in a gene expressed in the upper quartile of genes in the developing brain³³, (12) the LoF variant in chromatin-modifying gene, (13) the LoF variant in known neurodevelopmental disorder gene, (14) the LoF variant in the constrained gene, defined as having a probability of LoF intolerance (pLI) score > 0.5³³, (15) mother’s highest education level, (16) father’s highest education level, and (17) total family income.

MRI data acquisition

Neuroimaging data were collected from seven institutions using either Siemens Prisma 3 T, Siemens Prisma Fit 3 T, or General Electric (GE) Discovery 3 T scanners. Structural MRI (sMRI) was acquired using a T1-weighted inversion-prepared RF-spoiled gradient echo sequence, while diffusion MRI (dMRI) data were obtained through a multi-shell, multi-band echo-planar imaging sequence with a slice acceleration factor of 3. The dMRI acquisition included [7, 6, 15, 15, 60] directions with b-values of [0, 500, 1000, 2000, 3000] s/mm². Image acquisition protocols used in Siemens and GE scanners are shown in Table 2. Additionally, two field maps with opposite phase-encoding directions were acquired to compensate for image distortions in the dMRI data caused by magnetic field inhomogeneities.

Table 2.

MRI acquisition protocols for the T1-weighted and diffusion MRI data acquired on the Siemens and GE scanners

MRI Sequence	Scanner	TR (ms)	TE (ms)	TI (ms)	FA (◦)	FOV (mm)	Slices	Voxel Size (mm)	Scan Time (min)	Diffusion Directions (Total for b=0)	b-values (s/mm2)
T1-weighted	Siemens 3T	2500	2.88	1060	8.0	256×256	176	1.0×1.0×1.0	7:12	N/A	N/A
	GE 3T	2500	2	1060	8.0	256×256	208	1.0×1.0×1.0	6:09	N/A	N/A
T2-weighted	Siemens 3T	3200	565	N/A	Variable	256×256	176	1.0×1.0×1.0	6:35	N/A	N/A
	GE 3T	3200	60	N/A	Variable	256×256	208	1.0×1.0×1.0	5:50	N/A	N/A
Diffusion	Siemens 3T	4100	88	N/A	90.0	240×240	81	1.7×1.7×1.7	7:31	[7, 6, 15, 15, 60] (96)	[0, 500, 1000, 2000, 3000]
	GE 3T	4100	81.90	N/A	90.0	240×240	81	1.7×1.7×1.7	7:30	[7, 6, 15, 15, 60] (96)	[0, 500, 1000, 2000, 3000]

GE General Electric, TR repetition time, TE echo time, TI inversion time, FA flip angle, FOV field of view.

Motion handling and quality assurance

Motion artifacts were addressed through multiple approaches. For structural MRI (T1-weighted [T1w] and T2-weighted [T2w]), prospective motion correction was implemented using established methods^49,50. Manual quality control was performed before full image processing to exclude poor-quality structural scans or select the best scan when multiple acceptable scans were available. For cortical surface reconstruction, trained technicians reviewed the accuracy of FreeSurfer reconstructions, specifically evaluating motion artifacts among five categories of image quality issues. Each reconstruction received an overall QC score indicating whether it was recommended for use (1) or exclusion (0), with exclusion recommended if any category, including motion, was rated as severe. While prospective motion correction was not implemented for diffusion MRI or fMRI acquisitions, the multiband echo planar imaging (EPI) sequences used for these acquisitions are inherently more motion-robust than conventional single-band sequences.

Diffusion MRI preprocessing

Preprocessing of diffusion MRI series included comprehensive corrections for common artifacts and distortions. Eddy current distortions were corrected using a model-based approach based on diffusion gradient orientations and amplitudes⁵¹, with corrections limited to displacement along the phase-encode direction⁵². Head motion correction was performed by rigid-body-registering each frame to the corresponding volume synthesized from a robust tensor fit, accounting for image contrast variation between frames. Dark slices caused by abrupt head motion were replaced with values synthesized from the robust tensor fit, and the diffusion gradient matrix was adjusted for head rotation. Spatial and intensity distortions caused by B0 field inhomogeneity were minimized using FSL’s TOPUP, which relies on reversing phase-encode polarities to robustly correct EPI distortions. Gradient nonlinearity distortions were then corrected for each frame⁵³. The b = 0 diffusion images were registered to T1w structural images using mutual information⁵⁴, after coarse pre-alignment via within-modality registration to atlas brains. Diffusion images were resampled in a standard orientation with 1.7 mm isotropic resolution (equal to the acquisition resolution).

Multi-site MRI harmonization

ComBat harmonization⁵⁵ was applied to adjust for site-related variations, while age at imaging (both linear and quadratic terms), sex, and intracranial volume were regressed out from each measure. To enhance robustness, bootstrapping was performed using 63.2% of the sample across 100,000 iterations. The resulting residuals were then z-scored for clustering analysis. For further details, refer to⁵⁶.

We acknowledge that applying ComBat harmonization to the entire dataset before cross-validation creates a potential data leakage concern, as harmonization parameters may be influenced by test set data. However, our small sample size (N = 89) and multi-site design (7 sites with varying participant counts) would result in unstable harmonization parameters when applied within individual cross-validation folds. ComBat requires a sufficient sample size per site for reliable parameter estimation, which would be compromised in our small fold sizes.

Extraction of MRI features

Using FreeSurfer⁵⁷ tools, structural and diffusion tensor imaging (DTI) features were also extracted from the brain MRIs of this dataset. A total of 2693 different structural MRI features were extracted, including cortical and subcortical thickness, sulcal depth, area, volume, T1- and T2-weighted image intensity, and contrasts, etc. of various brain region-of-interests (ROIs). In addition, a total of 2994 different DTI features were also extracted, including fractional anisotropy (FA), mean diffusivity (MD), longitudinal diffusivity (LD), and transverse diffusivity (TD) measures at different DTI fiber locations and segments (see Table 3 for details).

Table 3.

List of MRI feature types and the prefixes used to refer to different features

MRI Sequence	Extracted Feature
T1- and T2-weighted (2693 features)	Cortical thickness in 257 regions Cortical sulcal depth in 258 regions Cortical surface area in 258 regions Cortical volume in 258 regions T1-weighted white matter intensity in 258 cortical regions T1-weighted gray matter intensity in 258 cortical regions T1-weighted contrast in 258 cortical regions T2-weighted white matter intensity in 258 cortical regions T2-weighted gray matter intensity in 258 cortical regions T2-weighted contrast in 258 cortical regions Volume in 42 subcortical regions T1-weighted intensity in 36 subcortical regions T2-weighted intensity in 36 subcortical regions
DTI (2994 features)	Fractional anisotropy (FA) of white matter fibers Mean diffusivity (MD) of 42 white matter fibers Longitudinal diffusivity (LD) of 42 white matter fibers Transverse diffusivity (TD) of 42 white matter fibers Volume of 42 white matter fiber tracts FA of 30 subcortical structures MD of 30 subcortical structures LD of 30 subcortical structures TD of 30 subcortical structures FA of 222 white matter regions at the gray-white matter surface MD of 222 white matter regions at the gray-white matter surface LD of 222 white matter regions at the gray-white matter surface TD of 222 white matter regions at the gray-white matter surface FA of 222 gray matter regions at the gray-white matter surface MD of 222 gray matter regions at the gray-white matter surface LD of 222 gray matter regions at the gray-white matter surface TD of 222 gray matter regions at the gray-white matter surface FA of 222 gray-white contrast regions at the gray-white matter surface MD of 222 gray-white contrast regions at the gray-white matter surface LD of 222 gray-white contrast regions at the gray-white matter surface TD of 222 gray-white contrast regions at the gray-white matter surface

DTI diffusion tensor imaging, FA Fractional anisotropy, MD Mean diffusivity, LD Longitudinal diffusivity, TD Transverse diffusivity.

Neurocognitive Outcome Measurements

We primarily included 20 different test scores. Five tests were performed using the Wide Range Achievement Test – Version 4 (WRAT-IV; see Table 4). Intelligence was assessed for 46 participants with ages ranging from 7.9 to 15.9 years using the Wechsler Intelligence Scale for Children – Fifth Edition (WISC-V), while the remaining 43 participants with ages ranging from 16.1 to 29.9 years were assessed using the Wechsler Adult Intelligence Scale – Fourth Edition (WAIS-IV). All test scores were age-standardized and normalized to the same scale, ensuring that the results are comparable across the two instruments despite their structural differences⁵⁸. The neurocognitive assessments occurred within 6 months of brain MRI acquisition: within 2 days for 81 participants, within 1–6 weeks for 6 participants, and 3–6 months for 2 participants. The age at neurocognitive tests was 7.9–29.9 years (15.9$\pm$6.4). The neurocognitive assessment was assumed to be relatively stable for AYAs within this short period⁵⁹. However, 5 scores were missing for some participants. To ensure consistency, we included only the 15 test scores available for all 89 participants as targets for machine learning inference in this study (see Table 4 for rows with gray background colors). These 15 scores belong to 7 neurocognitive domains: academic achievement, verbal comprehension, perceptual reasoning, working memory, processing speed, fluid reasoning, and general intelligence. Although the WISC-V often uses a second-order five-factor model⁶⁰ that distinguishes between fluid reasoning and visual-spatial (perceptual reasoning) abilities, the WAIS-IV follows a four-factor structure⁶¹ in which these two domains are combined. Despite this structural difference, we retained the five-factor distinction in our analysis to enable more granular exploration of domain-specific brain–behavior relationships. This approach allows for better interpretability of the domain inferability across the full age span, even though the test structure varied between younger (WISC-V) and older (WAIS-IV) participants. In addition, while academic achievement is not traditionally considered a core cognitive domain like the others, it is often treated as a functional outcome of cognitive ability. In this study, we include it as a separate domain to capture real-world, school-related performance, which is especially relevant in evaluating neurodevelopmental impact in youth and young adults.

Table 4.

Fifteen neurocognitive testing scores in 7 domains are available for all 89 participants (shaded cells)

Neurocognitive Test Name	Domain Accessed	Test Method	# Subjects	[Min, Max]	Mean$\pm$Std	IQR [p25, p75]
Word Reading	Academic Achievement	WRAT-IV	89	[72, 145]	106.37 $\pm$13.27	[97, 114]
Sentence Comprehension			89	[10, 145]	103.77 $\pm$19.10	[95, 114.5]
Spelling			89	[9, 139]	100.92$\pm$16.64	[95, 112.5]
Reading Composite			89	[66, 137]	106.04$\pm$13.02	[96, 115]
Math Computation			89	[55, 139]	100.92$\pm$16.22	[90.5, 112.5]
Verbal Comprehension Index	Verbal Comprehension	WISC-V, WAIS-IV	89 (46/43)	[68, 144]	106.96$\pm$16.30	[95.5, 119.5]
Vocabulary			89 (46/43)	[5, 19]	11.78$\pm$2.94	[10, 14]
Similarity			89 (46/43)	[2, 19]	10.61 $\pm$3.23	[8, 13]
Perceptual Reasoning Index*	Perceptual Reasoning/ Visual Spatial	WAIS-IV	43	[84, 127]	100.37$\pm$11.00	[92, 109]
Block Design		WISC-V, WAIS-IV	89 (46/43)	[3, 16]	8.98$\pm$2.80	[7, 11]
Visual Puzzles*		WAIS-IV	43	[7, 15]	10.11$\pm$2.13	[9, 12]
Working Memory Index*	Working Memory	WAIS-IV	43	[74, 131]	99.09 $\pm$12.42	[92, 105]
Digit Span		WISC-V, WAIS-IV	89 (46/43)	[5, 17]	9.86$\pm$2.88	[8, 11]
Processing Speed Index	Processing Speed	WISC-V, WAIS-IV	89 (46/43)	[62, 132]	96.83 $\pm$12.97	[89, 105]
Coding			89 (46/43)	[2, 16]	9.29$\pm$2.56	[8, 11]
Symbol Search			89 (46/43)	[4, 17]	9.53 ± 2.51	[8, 11]
Fluid Reasoning Index*	Fluid Reasoning	WISC-V	46	[74, 134]	100.93 $\pm$15.95	[89.5, 115]
Figure Weights*			46	[5, 18]	10.52 $\pm$3.03	[8.5, 13]
Matrix Reasoning		WISC-V, WAIS-IV	89 (46/43)	[5, 16]	10.46$\pm$2.71	[8, 12]
Full-Scale IQ**	General Intelligence	WISC-V, WAIS-IV	89 (46/43)	[73, 130]	101.12 $\pm$14.27	[93, 111.5]

WRAT-IV Wide Range Achievement Test - Version 4, WISC-V Wechsler Intelligence Scale for Children – Version 5, WAIS-IV Wechsler Adult Intelligence Scale – Version 4, IQR interquartile range, p25 25th percentile, p75 75th percentile. *Not used as a target score due to small sample size for inference in the present study, **Full-Scale IQ is a composite of Similarities, Vocabulary, Block Design, Matrix Reasoning, Figure Weights, Digit Span, and Coding.

Machine Learning for Inferring Neurocognitive Testing Scores

We used all the variables shown in Table 1 (except for the presence of ApoE alleles) and MRI metrics as input variables. ApoE allele information was available for 56 participants (63% of the cohort), but was missing for 33 participants (37% of the cohort). To preserve our full sample size of N = 89 and maintain statistical power for the machine learning analysis, we excluded ApoE alleles from the primary analysis rather than reducing the sample size to N = 56. We acknowledge this as a limitation, as ApoE alleles are known to influence neurocognitive outcomes and brain structure. The difference in age between MRI and neurocognitive testing was small, with the interquartile range of 0.5–1.5 days. We also standardized all continuous variables by subtracting the mean and dividing by the standard deviation before training and validating our prediction models. To streamline feature selection, we initially reduced the total of 5,687 MRI features (2,693 structural and 2994 DTI, in Section 0 and Table 3) by selecting the top 500 features with the smallest p-values in Pearson correlation with neurocognitive testing scores after false discovery rate (FDR) correction. Additionally, 17 non-MRI and MRI-derived covariates (Section 0 and Table 1), including genetic and demographic factors, were included, resulting in a total feature set of 517.

Multivariate Inference Model

We built a machine learning inference/regression model for each of the 15 neurocognitive testing scores. The inference models were based on the average inferred scores from Random Forest (RF), linear support vector regressor (SVR), and Gaussian-kernel SVR, because of their ability to account for linear and non-linear relationships between the inferred and target neurocognitive testing scores. We did not include models such as Elastic Net or LASSO, which rely on transformed or embedded feature representations, as such approaches hinder direct interpretability by obscuring the contribution of original input features to specific neurocognitive functions. Feature selection served as a major means to reduce the risk of over-fitting and find a subset of features that can jointly infer the target variable with the highest accuracy. We used an enhanced forward inclusion and backward elimination (FIBE)⁶² approach (https://github.com/i3-research/fibe). Our approach started from the single most informative feature, among all 517 features, based on the accuracy of the average inferred scores from RF, linear SVR, and Gaussian SVR, in a 5-fold cross-validation. During the forward inclusion phase, our algorithm added one feature at a time, such that adding this feature led to the greatest increase in the 5-fold cross-validated inference accuracy compared to adding any other feature. Features were added until no features could be added to further improve accuracy. This finished one round of forward inclusion. Then the backward elimination phase started to eliminate features, one at a time, among the selected features at the time. A feature was eliminated such that removing this feature led to the greatest increase in the 5-fold cross-validated inference accuracy compared to removing any other features in the tentatively selected feature subset, until no feature could be removed for further improvement of inference accuracy. This finished one round of backward elimination. The algorithm iterated between forward inclusion and backward elimination, until no feature could be added or removed to improve inference accuracy in the cross-validation further. Note that our FIBE algorithm addresses potential redundancy by only including features that demonstrably improve cross-validated inference performance as measured by Mean Absolute Error (MAE), rather than simply selecting features based on individual correlations. The algorithm iteratively evaluates whether each new feature reduces MAE beyond what is already achieved by the existing feature set, ensuring that redundant features are eliminated during the backward elimination phase if they do not contribute to improved inference accuracy. Details of this pipeline can be found in the Supplementary Information. The final inference was based on the selected subset of features, not the entire 517 features, on the 89 participants in the testing data (5-fold cross-validation). The number of finally selected features being 1/8–1/4 of the sample size helped reduce the risk of over-fitting in this study. In similar studies, this FIBE approach usually selected 5–15 features out of hundreds or thousands of clinical and MRI features, hence reducing the risk of over-fitting in the context of dozens or a lower hundreds of participants^63–65.

Accuracy evaluation and statistical test

We built an inference model for each of the 15 neurocognitive test scores. The data was split into 80% for training and feature selection, and 20% for testing, 5 times (i.e., 5-fold cross-validation) to ensure that each participant was used as a testing sample once and only once. Pearson correlation ($r$) measures (see Supplementary Equation 3) how closely our inferred scores matched the actual scores, ranging from −1 to +1. Values closer to +1 indicate better inference. In addition, Mean Absolute Percentage Error (MAPE) represents the average percentage difference between our inferred and actual scores (see Supplementary Equation 4). Lower values mean better accuracy. An inference met statistical significance of accuracy if the Pearson correlation coefficient ($r$) between the inferred and actual neurocognitive testing scores achieved a p-value < 0.05, for each of the 15 neurocognitive testing scores.

Uncertainty analysis and validation approach

Given the small sample size (N = 89), we implemented a comprehensive validation strategy to ensure robust and reliable results. Our approach addressed potential variability in feature importance and inferred estimates that may arise from the specific individuals included and the order of feature selection in the FIBE algorithm.

Bootstrap application

The 100,000 bootstrap iterations were applied specifically to the ComBat harmonization process to ensure robust site harmonization across our 7-site dataset. This approach was chosen to address site-related variations while maintaining computational feasibility.

Cross-validation strategy

For model validation and uncertainty estimation, we employed 5-fold cross-validation rather than comprehensive bootstrap uncertainty reporting. This approach is more appropriate for our sample size, as bootstrap-based uncertainty estimates can be unstable with small samples. Cross-validation ensures that each participant serves as a test case exactly once, providing robust validation without requiring extensive computational resources.

Uncertainty quantification

To address the uncertainty of our prediction models, we calculated 95% confidence intervals for all correlation coefficients using Fisher’s z-transformation. This approach provides uncertainty bounds for our inference accuracy while remaining computationally feasible and scientifically rigorous.

The rationale for this approach was that cross-validation provides more appropriate validation for our sample size than bootstrap-based uncertainty estimates.

Modeling strategy rationale

Our modeling strategy employed simultaneous feature selection across all available variables (demographic, genetic, socioeconomic, and MRI features) rather than a hierarchical baseline approach. This strategy was chosen for several reasons: (1) it allows identification of the optimal combination of features across all domains simultaneously, (2) it enables direct ranking of neurocognitive functions by inferability, and (3) it provides a comprehensive view of which factors contribute most to inference across the full feature space.

To address potential concerns about the added value of neuroimaging features, we conducted a supplementary baseline comparison using only non-MRI features (demographic, genetic, socioeconomic, and clinical). This baseline analysis, presented in the Supplementary Table 21, demonstrates the substantial incremental contribution of neuroimaging data beyond conventional clinical covariates. The proposed MRI-based approach achieved superior performance across all cognitive domains, with mean correlation improving from 0.197 (baseline) to 0.483 (proposed) and mean absolute error decreasing by 1.7 IQ points.

Our exploratory goal was to identify which neurocognitive functions are most inferable from the complete feature set while ensuring that MRI features provide meaningful added value over standard clinical variables. The MRI features were residualized for age, sex, and scanner site, which partially addresses concerns about treating all features equally. The baseline comparison confirms that the neuroimaging features capture clinically relevant information not available through demographic and clinical variables alone.

Inferability score and ranking

To determine which neurocognitive functions were easiest to infer, we created an “Inverse Inferability Score” (IIS) for each function (see Supplementary Section 1.5 and Supplementary Equations 5-6). This score combines both accuracy measures, Pearson correlation ($r$) and mean absolute percentage error (MAPE). Lower IIS values indicate better inferability. We ranked all 15 neurocognitive test scores and 7 broad domains from most inferable (rank 1) to least inferable. This helps identify which cognitive abilities are most strongly associated with brain imaging and other factors in adolescents with congenital heart disease.

Statistics and reproducibility

All analyses were performed in Python (v3.9) on an Intel Xeon E5-2650 v4 (Broadwell, 2.2 GHz) CPU with 16 GB of RAM. Preprocessing and harmonization were identical across participants; imaging features were site-harmonized (ComBat with 100,000 bootstraps), then adjusted for age, sex, and scanner. Prediction models were trained with a consensus ensemble of three regressors from scikit-learn: (1) Random Forest Regressor RandomForestRegressor (n_estimators = 100, max_depth = ;5, random_state = 42), (2) Support Vector Regressor (SVR) with Gaussian kernel SVR(kernel = 'rbf', C = 1.0, gamma = 'scale'), and SVR with linear kernel SVR(kernel = 'linear', C = 1.0, epsilon = 0.2). Model performance was evaluated using two-sided Pearson correlation ($r$), mean absolute error (MAE, in test-score units), mean absolute percentage error (MAPE), and the inverse inferability score (IIS). Ninety-five percent confidence intervals for $r$ were computed via Fisher’s z-transformation. Statistical significance was assessed at p < 0.05 with Bonferroni correction across tests. Internal validation used 5-fold cross-validation with a fixed random state; each participant (N = 89) appeared exactly once in a held-out fold per test, constituting independent replicates. Only scores available for all participants (15 tests across 7 domains) were analyzed (complete-case). All hyperparameters were fixed a priori; no adaptive tuning was performed. Reproducibility was supported by deterministic seeds (random_state = 42), identical preprocessing across participants, and a standardized training/validation protocol applied uniformly to every neurocognitive outcome.

Results

It is important to clarify that this study examines contemporaneous prediction, that is, inferring neurocognitive test scores from brain MRI features collected within the same time period (within 6 months, with 91% collected within 2 days). This cross-sectional approach is valuable for understanding brain-behavior relationships and identifying which neurocognitive functions are most inferable from brain structure. However, it has important limitations for clinical translation, as contemporaneous inference has limited clinical utility compared to longitudinal prediction, which would enable early identification and intervention for at-risk individuals. True clinical prediction would require longitudinal data where brain MRI is collected at one time point, and neurocognitive outcomes are assessed at a later time point.

Inferability of 15 neurocognitive and 7 broad domains

Among the 15 neurocognitive testing scores in the 7 domains (Fig. 2a), the top 3 neurocognitive testing scores with the lowest IIS (i.e., highest inferability) were: (1) reading composite (IIS = 0), in the academic achievement domain; (2) processing speed index (IIS = 0.0129), in the processing speed domain; and (3) word reading (IIS = 0.0165) in the academic achievement domain. These are more inferable neurocognitive testing scores in our data. In contrast, the 3 least inferable neurocognitive testing scores were: (1) sentence comprehension (IIS = 3.5164), in the academic achievement domain; (2) similarities (IIS = 2.0833), in the verbal comprehension domain; and (3) block design (IIS = 1.9199), in the perceptual reasoning domain.

Fig. 2. Pearson correlation between the actual and inferred scores.

a Ranking of neurocognitive test scores and domains based on inverse inferability score (IIS). b Number of features selected for each neurocognitive test score and domain. c–q Correlation plots are also shown between the actual and inferred scores of individual neurocognitive test scores in the ascending order of inverse inferability score (IIS). Associated adjusted p-values and MAPE values are shown within the plots.

If we rank the 7 domains by the average IIS of all neurocognitive testing scores in that domain (Fig. 2a), then the most inferable domain was general intelligence (i.e., full-scale IQ; Average IIS = 0.0886). The next most inferable domain with a much higher average IIS was working memory (Average IIS = 0.7100). The least inferable domains were (1) perceptual reasoning (Average IIS = 1.9199) and (2) academic achievement domains (Average IIS = 1.1582).

Selected features across neurocognitive domains

Out of the 517 total features, our FIBE method selected 6 to 14 features for each neurocognitive test score, with an overall mean of 11.4 (Fig. 2b). Supplementary Fig. 2 summarizes all the features (rows) selected by our FIBE algorithm (columns). Our FIBE algorithm identified 89 key features across multiple brain regions and measurement types. Structural MRI features (cortical thickness, area, volume, and sulcal depth) were predominantly located in frontal regions (superior frontal gyrus, middle frontal gyrus, orbital sulcus), temporal regions (superior temporal gyrus, transverse temporal gyrus, temporal pole), and parietal regions (postcentral gyrus, supramarginal gyrus). Diffusion MRI features (fractional anisotropy, mean diffusivity, longitudinal diffusivity) were concentrated in white matter tracts, including the cingulum bundle, inferior fronto-occipital fasciculus, and thalamic radiation, as well as gray-white matter interfaces in frontal, temporal, and occipital regions. Subcortical structures included the thalamus, hippocampus, accumbens area, and cerebellum. Genetic factors (loss-of-function variants in neurodevelopmental and chromatin-modifying genes, as well as de novo mutations) and clinical variables (including cardiac surgeries and CHD diagnosis) were also considered. Socioeconomic factors (father’s education level) contributed to predictions. The predominance of frontal and temporal regions aligns with their established roles in executive function, language, and working memory. At the same time, the inclusion of white matter tracts suggests the importance of brain connectivity in neurocognitive outcomes among adolescents with congenital heart disease. We also examine the sign of the Pearson correlation coefficient between each neurocognitive test score and each feature in the selected feature set (see Supplementary Tables 1–15) to interpret the direction of association. A positive correlation suggests that higher feature values are generally associated with increased neurocognitive test scores, whereas a negative correlation indicates that greater feature values are typically linked to lower neurocognitive test scores.

Inference accuracy and uncertainty estimates

Figure 2a shows that the Pearson correlation ($r$) between actual and inferred test scores ranged from 0.245 to 0.648. Supplementary Table 16 further presents the inference accuracy and uncertainty estimates for all 15 neurocognitive test scores. Thirteen out of fifteen correlations achieved statistical significance (p < 0.05).

The MAE in IQ points provides clinically meaningful measures of inference accuracy, ranging from 1.6 to 12.0 points (mean MAE = 6.3 points). Tests with the highest correlations included Digit Span ($r$ = 0.648, 95% CI: 0.508–0.754, MAE = 1.7 points), Verbal Comprehension Index ($r$ = 0.594, 95% CI: 0.440–0.714, MAE = 10.9 points), and Matrix Reasoning ($r$ = 0.574, 95% CI: 0.416–0.699, MAE = 1.7 points).

The confidence intervals (CIs) demonstrate that our predictions are statistically significant and provide meaningful uncertainty bounds for clinical interpretation. While the lower bound of some confidence intervals approaches zero (e.g., Sentence Comprehension: 0.039-0.432), all correlations achieved statistical significance, indicating that our predictions are significantly better than chance.

Clinical Range Distribution

To assess the clinical significance of our findings, we analyzed the distribution of participants across clinical ranges using standard neuropsychological cutoffs ( < 85 for IQ/Index scores [approximately 1 standard deviation below the standardized test mean of 100], and < 7 for scaled scores [approximately 1 standard deviation below the standardized test mean of 10]). Our cohort demonstrated a relatively normative distribution, with full-scale IQ showing a mean of 101.1$\pm$14.27 and clinically concerning rates ranging from 4.5% (Reading Composite) to 15.7% (Processing Speed Index) across IQ/Index scores. The distribution closely matched normative population expectations, with 49.4% of participants in the Average range (90-109) compared to the expected 50.0%. Detailed clinical range analysis is provided in Supplementary Information (see Supplementary Section 2.5 and Supplementary Table 19).

Key features inferring neurocognitive testing scores

In general intelligence, the most inferable broad domain, full-scale IQ was positively associated with the father’s education level, cortical area in the postcentral sulcus (BAs 2, 3), and fiber volume of the right inferior fronto-occipital was negatively associated with the cortical thickness in the superior frontal gyrus (BA 9). The mother’s education level was not selected by the FIBE approach. The correlation between the father’s and mother’s education levels is modest ($r$ = 0.14) in our data, suggesting that their contributions to cognitive outcomes may be at least partially independent.

Working memory, the second most inferable broad domain, was linked to regions supporting numerical processing and executive function. Digit span was found to be positively involved in the transverse diffusivity of gray matter in the inferior frontal gyrus (BA 47) and the cortical volume in the supramarginal gyrus (BA 40), supporting sensory integration and phonological storage.

Processing speed, the third most inferable broad domain, was associated with regions supporting attention and visual scanning. Symbol search was positively associated with the sulcal depth in the precuneus (BAs 7, 31) and nucleus accumbens volume. In contrast, coding was negatively associated with the transverse diffusivity of the gray-white matter surface in the middle occipital gyrus (BAs 18, 19) and cortical thickness in the superior temporal gyrus (BAs 22, 42). The processing speed index was also negatively associated with the cortical thickness in the insula (BAs 13, 47) and cortical volume in the superior temporal gyrus (BAs 22, 42), along with the presence of the LoF variant in chromatin-modifying genes.

Fluid reasoning, the fourth most inferable domain in our study, was primarily linked to white matter microstructural integrity and frontolimbic regions involved in abstract reasoning and problem solving. Matrix reasoning was negatively associated with the mean diffusivity in the pars orbitalis (BA 47) but positively associated with the fractional anisotropy and fiber volume in the cingulum bundle (left).

Similarities in verbal comprehension, the fifth most inferable broad domain, were associated with cognitive processing and conceptual integration regions. It was negatively associated with the T2-weighted white matter signal intensity in the inferior temporal gyrus (BAs 20, 21, 37) but positively associated with the mean diffusivity in the caudal middle frontal gyrus (BAs 8, 9). Vocabulary was negatively associated with the presence of the LoF variant in chromatin-modifying genes but positively associated with areas supporting memory and executive functions, such as the volume in the hippocampus and FA, mean and lateral diffusivity in the superior temporal gyrus (BAs 41, 42, 22). The verbal comprehension Index was negatively associated with the total number of cardiac surgeries and a LoF variant in known neurodevelopmental disorder genes, but positively associated with the CHD diagnosis. In addition, it was negatively associated with the estimated matrices in regions crucial for languages and executive function, such as the cortical thickness in the superior frontal gyrus (BAs 9, 10) and precentral sulcus (BAs 6, 4).

In the academic achievement domain, the sixth most inferable broad domain, word reading was negatively associated with the genetic marker, the presence of a LoF variant in a known neurodevelopmental disorder gene. Furthermore, word reading was found to be linked to brain regions involved in sensory and auditory processing, including a positive association with the T1-weighted white matter signal in the lingual gyrus (BAs 17, 18, 19) and fractional anisotropy in the superior temporal gyrus (BAs 22, 41, 42) but negatively associated with the cortical sulcal depth in the middle temporal gyrus (BA 21). Reading composite was positively associated with areas important for language comprehension and memory, such as the sulcal depth in the inferior temporal gyrus (BAs 20, 21) and the volume of the right hippocampus. Sentence comprehension positively involved brain structures related to language processing and connectivity, including the sulcal depth in the inferior temporal gyrus (BAs 20, 21) and the fractional anisotropy in the anterior thalamic radiation. Math computation was found to be linked to sex, where male subjects tend to score higher in math computation. In addition, it is negatively associated with the presence of the LoF variant in known neurodevelopmental disorder genes and the sulcal depth in the posterior cingulate gyrus (BA 23) and inferior frontal gyrus (BA 47). Furthermore, spelling was positively linked to a de novo genetic marker case and regions supporting visual and motor coordination, such as the fractional anisotropy in the occipitotemporal fusiform gyrus (BAs 37, 20) and the lateral diffusivity in the left cerebellum cortex.

Perceptual reasoning, the least inferable broad domain, was linked to brain regions involved in visual-spatial reasoning and executive function. Block design was positively associated with the father’s education level and the lateral diffusivity in the supramarginal gyrus (BA 40) but negatively associated with the presence of a LoF variant in a chromatin-modifying gene and the fractional anisotropy in the thalamus.

Overall, we have the following observations: (1) Father’s education level, a socioeconomic feature, was necessary for the inference of general intelligence and perceptual organization broad domains; (2) Clinical features are essential, in that the total number of cardiac surgeries and congenital heart disease diagnosis were crucial in the inference of verbal comprehension broad domain; (3) Sex, a demographic feature, was linked to the inference of the math computation, a subdomain of academic achievement domain; and (4) structural and diffusion MRI metrics contributed to inferring all seven broad domains, but by different MRI metrics in different brain structures (see Fig. 3 for cortical structures).

Fig. 3. Salient cortical regions associated with neurocognitive function across broad cognitive domains.

FIBE-selected feature-corresponding salient cortical regions for 15 neurocognitive testing scores (a–o) in 7 broad domains (7 big boxes of distinct colors). Broad domains are organized in the ascending order of average IIS (i.e., from most to least inferable broad domain).

Discussion

Thoroughly understanding the various factors influencing the risk of neurocognitive impairment is essential for timely and targeted intervention in AYAs with CHD. This study used 17 non-MRI and MRI-derived data elements (Table 1) and 5000+ structure and diffusion MRI features (Table 3) from 89 participants in 7 sites. We found that neuroimaging, socioeconomic, demographic, clinical, and genetic features interact and jointly influence various neurocognitive tasks. We have identified the most and least inferable neurocognitive testing scores among 15, in 7 broad domains (Fig. 2a), each being inferable by 6-14 (median 11) MRI and non-MRI features (Fig. 2b). Among MRI features, the key brain regions inferring neurocognitive testing scores have also been elucidated (Supplementary Fig. 2). Quantifying which neurocognitive test scores could be best inferred by which set of variables marks the first step toward (1) stratifying individual factors that impact neurocognitive outcomes, and (2) offering insights for key factors that should be targeted for future solutions aimed at improving neurocognitive outcomes.

Our simultaneous feature selection approach, while appropriate for our exploratory goals, represents one methodological choice among several valid approaches. To address this limitation, we conducted a supplementary baseline analysis comparing our proposed MRI-based model against a model using only demographic, socioeconomic, genetic, and clinical variables (Supplementary Table 21). This baseline comparison demonstrates that MRI features provide substantial incremental predictive value beyond what can be achieved with basic covariates alone, with mean correlation improving from 0.197 to 0.483 and mean absolute error decreasing by 1.7 IQ points across all neurocognitive tests. The baseline analysis confirms that neuroimaging features capture clinically meaningful information not available through conventional predictors, with 13 of 15 tests showing significantly improved correlations (p < 0.001) when brain imaging data are included. The most substantial improvements were observed for full-scale IQ and processing speed measures, suggesting that these cognitive domains are particularly dependent on the neuroanatomical features captured in our MRI protocol. While our simultaneous approach was designed to identify the optimal combination of all available features and rank neurocognitive functions by overall inferability, the baseline comparison provides crucial evidence for the clinical utility of brain imaging in this population. Future studies should continue to implement both hierarchical and simultaneous modeling approaches to provide comprehensive insights into the relative contributions of different feature types and optimize the clinical translation of neuroimaging-based cognitive assessment tools.

At least four categories of non-MRI factors were found to be important in inferring neurocognitive test scores in AYAs with CHD (see Supplementary Fig. 2). One category is genetics. The presence of predicted functionally deleterious de novo variants in genes not previously associated with neurodevelopmental disability, defined by excluding individuals with known pathogenic variants in curated neurodevelopmental disorder or CHD-associated genes, was positively associated with spelling inference, extending recent work linking DMN-related disruptions to genes involved in neural connectivity⁶⁶. In contrast, LoF variants in known neurodevelopmental genes showed widespread negative associations across multiple domains (reading, math, verbal), reinforcing the role of specific subsets of rare deleterious variants in shaping cognitive outcomes⁶⁷. It is important to interpret these findings with two caveats: (1) parental cognitive function, for which we used father’s educational attainment as a proxy, captures a substantial portion of the variance in proband cognitive ability and may confound associations attributed to de novo variants; and (2) our use of feature selection, while beneficial for model generalization, may exclude relevant features that are highly correlated with more predictive ones, complicating inferences about the absence of an association. Nonetheless, individuals with de novo variants, as well as LoF in some gene sets, such as chromatin-modifying genes, are at a higher risk³³.

The second category of non-MRI factors is socioeconomic status. Paternal education emerged as the strongest positive factor of full-scale IQ in CHD adolescents, along with neuroimaging biomarkers (see Supplementary Fig. 2). This aligns with evidence linking the father’s education to enriched cognitive environments and stress resilience⁶⁸. Neuroimaging features (e.g., right superior frontal cortical thickness, parietal diffusivity) complement socioeconomic factors, reflecting frontoparietal network efficiency modulated by environmental input⁶⁹. Block design performance is also positively associated with paternal education, highlighting the intersection of environmental support and visuospatial reasoning⁷⁰. These findings suggest that CHD cognitive outcomes are shaped by an additive influence of genetic, environmental, and neural factors, rather than a single dominant contributor. While our study was cross-sectional and observational in design, these correlations underscore the importance of multilevel models for understanding risk and resilience. If validated in longitudinal studies, such findings could inform dual-pronged strategies that integrate cognitive enrichment efforts⁷¹ with neurodevelopmental monitoring^70,72.

Clinical variables are the third category of factors crucial for neurodevelopment after socioeconomic status. CHD diagnosis, which we encoded in an ordinal scale reflecting increasing physiological and anatomical favorability (from single ventricle with arch obstruction to bi-ventricle without arch obstruction), was positively associated with verbal comprehension, while the number of cardiac surgeries negatively impacts verbal comprehension, but has no other neurodevelopmental functions. This selective effect may stem from the vulnerability of language networks to cumulative surgical stress and chronic hypoxia^73,74, emphasizing the need for targeted interventions and long-term monitoring of verbal skills for people with CHD, especially those undergoing multiple surgeries, to support verbal skills in people with CHD⁷⁵.

Findings on the role of demographics, the fourth category of non-MRI factors, on neurodevelopment in AYAs with CHD warrant more debate. Earlier literature reported an increased mortality rate in females with CHD^76,77 and increased risks of neurodevelopmental deficits in males with CHD requiring surgery⁴². However, a just-published meta-analysis study found no sex differences in neurodevelopmental outcomes among participants with CHD⁷⁸. Our findings align with this meta-analysis, that sex was not a selected feature in inferring neurocognitive test scores, except for math computation (females worse), for reasons unclear to us. Besides sex, our study did not find age as a strong factor in neurodevelopment in people with CHD aged 8-30 years, likely due to the adjustment of age of WISC-V and WAIS-IV scores. One study found that brain volume deficits in CHD correlate more with disease severity than age⁷⁹. Another study suggested that the cumulative effects of treatments and compensatory mechanisms may overshadow age-related changes⁸⁰. Despite these, future large-scale studies are needed to elucidate the effect of sex and age, especially combined with other comprehensive non-MRI and MRI factors, in shaping the neurodevelopmental trajectory after CHD.

On the MRI side, our study identified neuroanatomical regions that were associated with neurocognitive test scores in AYAs with CHD (Fig. 3). While our findings align with existing literature on the neural correlates of cognition, we also identified additional or divergent regions that may be relevant to this population. However, we acknowledge that these associations from our predictive models do not establish causal relationships or definitive neuroanatomical substrates of cognition. In our study, key cortical areas, such as Brodmann Areas (BAs) 18, 19, 21, 22, 41, and 42 for oral language⁸¹; BAs 20, 21, 41, and 42 for Semantic Processing⁸²; BA 20 for sentence comprehension⁸³; BA 37 for spelling⁸⁴; BAs 20 and 21 for similarities task⁸⁵; BAs 18, 19, 27 and 47 for vocabulary^86,87; BAs 10 and 40 for verbal comprehension^85–87; BAs 7 and 20 for processing speed^88,89; BAs 18, 19, and 37 for matrix reasoning⁸⁵; BAs 17 and 40 for block design^87,90; BAs 19, 46, and 47 for math computation⁹¹; BAs 2, 20, 40 and 47 for digit span^91,92; as well as BAs 9, 10, and 40 for full scale IQ⁹³, were consistently implicated, supporting their broad role in cognitive functions across populations. However, some discrepancies emerged when comparing our findings with those commonly reported in the general population. For instance, regions such as BAs 44 and 45, typically associated with verbal comprehension, and BAs 7, 8, 30, and 39 for processing speed, were not among the most predictive in our CHD cohort. Similarly, BAs 10, 13, and 46, often linked to perceptual reasoning, were not prominent in our results. One possible explanation is that chronic cerebral hypoperfusion, early-life cardiac surgeries, or disrupted developmental timing in individuals with CHD may lead to functional reorganization or reduced engagement of canonical cognitive circuits. In such cases, typical regions may be structurally or functionally compromised, resulting in compensatory recruitment of alternate brain areas. Conversely, some regions we identified, such as BAs 1, 2, and 3 for oral language; the insula and cerebellum for spelling; and the hippocampus and anterior thalamic radiation for memory, are less frequently reported in typically developing cohorts. These may reflect unique adaptations or vulnerabilities in the CHD brain, potentially arising from neurodevelopmental perturbations or neuroplastic compensation mechanisms. Further large-scale, longitudinal studies are needed to determine whether these atypical neural correlates represent CHD-specific patterns or general inter-individual variability.

Findings on MRI metrics (see Fig. 3 and Supplementary Fig. 2) also exhibit the relative importance between morphometry (from structural MRI) and diffusivity (from diffusion MRI). Cortical thickness, volume, area, and sulcal depth were key covariates for language-related tasks (word reading, vocabulary, and spelling), reinforcing their role in cortical development and general intelligence⁹⁴. Diffusion features, including FA, LD, and fiber bundle volumes, were particularly explanatory for visuospatial tasks (matrix reasoning) and language-related tasks, reflecting white matter integrity crucial for network efficiency. Additionally, gray-white matter interface features (FA, MD, TD, and LD) emerged as robust predictors across multiple tasks, suggesting their critical role in cognitive processing speed and structural connectivity. These findings underscore the interplay between cortical morphology, white matter pathways, and task-specific cognitive demands.

Our sensitivity analysis revealed that eliminating the most predictive features resulted in only a minor reduction in inference accuracy, suggesting a high degree of information redundancy among the selected neuroimaging and clinical features. This redundancy likely reflects the global, rather than strictly localized, impact of CHD on brain development. Physiological perturbations inherent to CHD, such as chronic hypoperfusion or hemodynamic instability, often affect the brain systemically, leading to widespread alterations in white matter microstructure and cortical development that are correlated across multiple regions. Consequently, no single localized feature is exclusively predictive; rather, multiple disparate features may serve as effective proxies for one another, capturing shared variance related to the global “brain phenotype” of CHD. This highlights the importance of multivariate models that can aggregate these distributed but redundant signals to infer cognitive outcomes.

When it comes to inferability, our study found that full-scale IQ could be inferred with the highest accuracy (9.8% cross-validation error), followed by working memory and processing speed. Lower inferability was observed for verbal comprehension, oral language, and perceptual organization (∼20% of inference errors). These findings demonstrate that different neurocognitive functions show varying degrees of predictability from brain imaging and clinical data. While these patterns are consistent with some existing literature on neural correlates of cognition, the specific mechanisms underlying these differential predictability patterns remain to be elucidated. Note that our findings should be interpreted as exploratory and hypothesis-generating rather than definitive. The MAE values in IQ points (ranging from 1.6 to 12.0 points; Supplementary Table 16) provide clinically meaningful measures of inference accuracy, but future validation in larger, independent cohorts is essential before clinical application.

Our analysis of WAIS-IV vs WISC-V differences (see Supplementary Section 2.6 and Supplementary Table 20 for details) revealed that WISC-V generally demonstrates higher standard errors of measurement than WAIS-IV (average difference = 0.456), consistent with the pattern that children’s tests typically have higher measurement error than adult tests. However, we found no significant relationship between standard error and our inference correlation ($r$ = −0.237, p = 0.510), suggesting that measurement error differences between tests do not significantly bias our prediction results. The strong relationship between standard error and Mean Absolute Error ($r$ = 0.862, p < 0.001) indicates that our models are appropriately sensitive to test psychometric properties, with higher measurement error associated with higher inference errors. While we acknowledge the limitation of combining WAIS-IV and WISC-V results due to our age-based test assignment, our findings suggest that measurement error differences between these tests do not substantially impact the validity of our brain-behavior relationship findings.

We also performed clinical stratification analysis of full-scale IQ prediction (see Supplementary Section 2.5 and Supplementary Table 18), which revealed that the model achieves 49.4% exact agreement with actual IQ classifications using standard clinical ranges, with perfect accuracy (100%) for the average IQ range (90-109 points). While the model shows a conservative bias, tending to predict participants as having average IQ, this performance represents a clinically meaningful improvement over chance and demonstrates that brain imaging can provide useful information for IQ classification. The high accuracy for the average range is particularly relevant for clinical practice, as this range encompasses the majority of the population and is most important for clinical decision-making.

While our cohort demonstrates relatively normative cognitive performance overall, the consistent selection of CHD-specific factors in our predictive models reflects their importance in explaining individual differences rather than group-level deficits. The model’s objective is to infer individual cognitive scores across the full performance spectrum, not merely to identify clinical impairment. Even within a normative distribution (mean full-scale IQ = 101.1), CHD-related variables help explain why one participant achieves a score of 88 (low average) while another reaches 114 (high average). This individual-level prediction capability has important clinical implications for personalized assessment and intervention planning. The alternative approach of applying normative population models to CHD patients would face significant limitations. Beyond the practical challenge of identifying large-scale normative datasets with comprehensive neurocognitive assessment across 15 domains, such models would inherently lack CHD-specific variables (cardiac complexity, genetic variants, CHD-related neuroanatomical alterations) that our analysis demonstrates contribute meaningfully to individual outcome prediction. Our findings suggest that CHD-specific mechanisms influence cognitive variability even within apparently normative ranges, supporting the clinical value of population-specific predictive modeling approaches.

Our paper has several limitations. First, our sample size of 89 participants, while adequate for this exploratory analysis, limits the generalizability of our findings. We acknowledge that feature importance and predicted estimates may vary depending on the specific individuals included and the order of feature selection in the FIBE algorithm. To address this limitation, we implemented 5-fold cross-validation and provided confidence intervals for all correlation coefficients (see Supplementary Table 16). While we cannot provide comprehensive distributions for all features and individuals due to practical constraints, our validation approach provides robust evidence for the reliability of our findings. The confidence intervals demonstrate the uncertainty bounds for our prediction accuracy, and our cross-validation approach ensures that each participant contributes to both training and testing phases. Future studies with larger sample sizes should implement comprehensive uncertainty reporting, including bootstrap confidence intervals for individual predictions and feature importance distributions. Such approaches would provide more detailed uncertainty quantification and further validate our findings. Second, our analysis exclusively relied on a single dataset (although from 7 sites), which may limit the applicability of our findings to other populations or imaging protocols. Third, we excluded ApoE allele information from our analysis due to the unavailability of complete data. While ApoE allele data were available for 56 participants (63% of our cohort), definitive ApoE allele information was not available for the remaining 33 participants (37% of our cohort). Including ApoE as a covariate would have required either: (1) excluding 35 participants from the analysis, which would have reduced our already limited sample size from N = 89 to N = 56, or (2) using imputation methods that could introduce bias given the substantial proportion of missing data (37%). Given our small sample size and the need to maintain statistical power for our machine learning analysis, we chose to exclude ApoE alleles from the primary analysis to preserve the full cohort of 89 participants. This decision prioritizes statistical power and generalizability over the inclusion of a potentially important genetic covariate with substantial missing data. Fourth, the absence of an independently held-out test set. Model performance was evaluated using cross-validation, which provides robust estimates of generalizability but may still be perceived as less rigorous than a strict train/validation/test split by some reviewers. Future work will benefit from validating the findings on an independent external dataset to further confirm model generalizability and reproducibility. Fifth, we used binary variables to represent genetic variants (presence/absence) rather than more granular measures such as burden scores or variant-level intolerance metrics using LOFTEE⁹⁵ (Loss-Of-Function Transcript Effect Estimator). This binary approach may not capture the full spectrum of genetic effects, particularly for genes with multiple variants or varying degrees of functional impact. The heterogeneity within our genetic categories, where a single binary variable represents genes with wide ranges of clinical outcomes and severity, may explain why genetic features appeared less frequently in our models compared to MRI and demographic variables. For example, loss-of-function variants in neurodevelopmental genes encompass a broad spectrum from mild functional impacts to severe pathogenic effects, yet our analysis treated all such variants equally. Future studies would benefit from implementing more sophisticated genetic modeling approaches, including: (1) burden scores that accumulate the number and severity of mutations within gene sets, (2) gene-level annotations using LOFTEE scores to better capture functional impact, and (3) pathway-based analyses that consider the biological context of genetic variants. Such approaches would provide a more nuanced understanding of how genetic factors contribute to neurocognitive outcomes in adolescents and young adults with congenital heart disease. Sixth, our application of ComBat harmonization to the entire dataset before cross-validation creates a potential data leakage issue that could lead to overly optimistic performance estimates. While we acknowledge this limitation, within-fold harmonization was not feasible given our small sample size and multi-site design. Additionally, we cannot provide comprehensive measures of ComBat’s success in regressing confounds and mitigating site effects due to our limited sample size. These limitations underscore the exploratory nature of our findings and the need for external validation in larger, independent cohorts. Future studies with larger sample sizes should implement within-fold harmonization or alternative approaches such as leave-one-site-out validation. Seventh, while comprehensive motion handling procedures were implemented, including prospective motion correction for structural MRI and manual quality control, our study has limitations in motion assessment. We did not quantify motion parameters (e.g., framewise displacement) for diffusion MRI acquisitions, and we did not explicitly test for correlations between motion and age or functional ability. Given that motion may correlate with age or neurodevelopmental concerns in pediatric populations, future studies should implement quantitative motion assessment and examine potential motion-related biases in the context of neurodevelopmental outcomes. Eighth, the participants in this study generally demonstrated cognitive performance within the normative range (mean full-scale IQ = 101.1 ± 14.2), with relatively few individuals showing severe cognitive impairments. While this allowed us to demonstrate meaningful prediction of individual differences within the normal range, our findings may not generalize to CHD patients with more significant cognitive deficits or those scoring substantially outside the normative distribution. The predictive relationships identified in our models may differ in populations with greater cognitive impairment, where different neuroanatomical factors or clinical variables might become more prominent predictors. Future validation studies should include participants with broader cognitive performance distributions, including those with more severe CHD-related cognitive impairments, to establish the generalizability of our neuroimaging-based prediction approach across the full spectrum of cognitive outcomes in this population. Lastly, this is a cross-sectional association study (i.e., an inference study) rather than a longitudinal prediction study, as neurocognitive tests were administered within a short window, typically within two weeks of brain MRI acquisition for nearly all participants (81 out of 89 within 2 days). Therefore, our findings reflect concurrent brain-behavior associations rather than the ability to predict future neurocognitive outcomes. Future work with longitudinal data will be necessary to assess predictive validity and developmental trajectories in AYAs with CHD.

Conclusion

Despite limitations, this study found that genetics, demographics, socioeconomics, clinical, and neuroimaging data all contribute to neurodevelopmental outcomes in adolescents and young adults (AYAs) with congenital heart disease (CHD), and the combinatorial mechanisms among 15 neurocognitive testing scores in 7 broad domains. It is the first step toward targeted intervention and differential medicine for the improvement of neurodevelopmental outcomes in AYAs with CHD. The findings need to be verified by future larger-scale, multi-site, more comprehensive, and ideally longitudinal data.

Author contributions

S.U.M. and Y.O. jointly supervised the research. M.A.H. and Y.O. conceptualized and designed the study. M.A.H. developed the machine learning algorithm. M.A.H., S.U.M., and Y.O. conducted the data analysis and interpreted the results. D.C.B., W.K.C., D.X., Y.S., J.W.N., and P.E.G. offered senior mentorship and insightful comments that shaped the study design and interpretation. S.H. generated the brain age deviation data. H.R.A. assisted in organizing the neurocognitive functions into their respective broad cognitive domains. E.A., M.B., J.C., B.D.G., E.G., D.J.H., H.H., P.M.Q., T.A.M., A.N.B., G.A.P., N.T., M.E.T., and J.W.N. contributed to the collection and structured organization of data from participating centers. M.A.H., S.U.M., and Y.O. contributed to drafting the initial manuscript. All authors reviewed, revised, and approved the final version of the manuscript.

Peer review

Peer review information

Communications Medicine thanks John Jairo Araujo, Jitse S. Amelink and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The source data underlying Fig. 2 are provided in Supplementary Data 1–3. Specifically, Fig. 2a data are available in Supplementary Data 1, Fig. 2b in Supplementary Data 2, and Fig. 2c-q in Supplementary Data 3. The datasets used for model training and validation contain controlled-access human subject data, including brain MRI, genomic (whole-exome/genome–derived variants), demographic, and socioeconomic information from participants enrolled in the PCGC CHD Brain and Genes study (ClinicalTrials.gov: NCT03070197). These data are not publicly available due to ethical, legal, and privacy considerations associated with identifiable and potentially re-identifiable participant information, as governed by the informed consent, institutional review board (IRB) approvals, and NIH data-sharing policies. Conditions of access. Access to the controlled datasets may be granted to qualified investigators for legitimate academic research purposes, subject to (i) approval by the data-holding institution(s), (ii) verification of IRB or equivalent ethics approval at the requesting institution (or determination of exemption, as applicable), and (iii) execution of an appropriate Data Use Agreement (DUA). Timeframe for response. Data access requests will be acknowledged within 2 weeks of receipt, and a decision will typically be communicated within 4–8 weeks, depending on the completion of institutional review and DUA processing. Restrictions on data use. Approved data may be used only for the purposes described in the approved request and in compliance with the DUA. Restrictions include but are not limited to: (a) no attempts at participant re-identification; (b) no redistribution or sharing of the data with third parties; (c) secure data storage and access limited to authorized personnel; and (d) destruction or return of the data upon completion of the approved research, as specified in the DUA. Request process. Access requests should be directed to Prof. Yangming Ou, PhD (yangming.ou@childrens.harvard.edu) or Prof. Sarah U. Morton, MD, PhD (sarah.morton@childrens.harvard.edu), who are responsible for coordinating institutional review and responding to data access requests.

Code availability

The source codes of our enhanced forward inclusion and backward elimination (FIBE)⁶² approach used in this manuscript are available at https://github.com/i3-research/fibe and at Zenodo⁶².

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01417-9.