Detection of Right and Left Ventricular Dysfunction in Pediatric Patients Using Artificial Intelligence–Enabled ECGs

Abstract

Background

Early detection of left and right ventricular systolic dysfunction (LVSD and RVSD respectively) in children can lead to intervention to reduce morbidity and death. Existing artificial intelligence algorithms can identify LVSD and RVSD in adults using a 12‐lead ECG; however, its efficacy in children is uncertain. We aimed to develop novel artificial intelligence–enabled ECG algorithms for LVSD and RVSD detection in pediatric patients.

Methods and Results

We identified 10 142 unique pediatric patients (age≤18) with a 10‐second, 12‐lead surface ECG within 14 days of a transthoracic echocardiogram, performed between 2002 and 2022. LVSD was defined quantitatively by left ventricular ejection fraction (LVEF). RVSD was defined semiquantitatively. Novel pediatric models for LVEF ≤35% and LVEF <50% achieved excellent test areas under the curve of 0.93 (95% CI, 0.89–0.98) and 0.88 (95% CI, 0.83–0.94) respectively. The model to detect LVEF <50% had a sensitivity of 0.85, specificity of 0.80, positive predictive value of 0.095, and negative predictive value of 0.995. In comparison, the previously validated adult data‐derived model for LVEF <35% achieved an area under the curve of 0.87 (95% CI, 0.84–0.90) for LVEF ≤35% in children. A novel pediatric model for any RVSD detection reached a test area under the curve of 0.90 (0.87–0.94).

Conclusions

An artificial intelligence–enabled ECG demonstrates accurate detection of both LVSD and RVSD in pediatric patients. While adult‐trained models offer good performance, improvements are seen when training pediatric‐specific models.

In this article, an artificial intelligence model was developed to detect left and right ventricular systolic dysfunction in pediatric patients. The newly developed pediatric models performed well in the validation cohort and outperformed previously validated models derived from adult data. Those individuals who were falsely predicted to have low left ventricular systolic function were at increased risk of future systolic function. With additional validation, these algorithms may be a valuable tool in pediatric heart failure screening

Nonstandard Abbreviations and Acronyms

LVSD

left ventricular systolic dysfunction

RVSD

right ventricular systolic dysfunction

Clinical Perspective.

What Is New?

• From a cohort of 10 142 unique pediatric patients, we trained novel artificial intelligence models to accurately detect left ventricular and right ventricular systolic dysfunction from ECGs in the pediatric population.

• In a subgroup analysis of 613 patients with follow‐up data, individuals falsely predicted to have systolic dysfunction were at a >3‐fold increased risk of developing systolic dysfunction in the future.

What Are the Clinical Implications?

• ECG screening for left and right ventricular systolic dysfunction may provide a low‐cost, easily accessible option for evaluating pediatric patients, helping to identify individuals in need of echocardiogram for definitive diagnosis.

Cardiac dysfunction is an important cause of morbidity and death in children, affecting ≈35 000 children annually. ¹ While the pathogenesis of ventricular dysfunction can be variable, the symptoms are often nonspecific and insidious, leading to underdetection and delayed diagnosis. ² In one study, half of the pediatric patients presenting to the emergency department or primary care with new‐onset heart failure initially received a diagnosis other than heart failure. ³ Early identification of left ventricular systolic dysfunction (LVSD) and right ventricular systolic dysfunction (RVSD) can lead to early interventions and initiation of medical therapy that improves heart failure symptoms and mortality rate. ⁴

Currently, the diagnosis of LVSD and RVSD requires echocardiography, an expensive and resource intensive test requiring operator expertise with limited availability for screening purposes. ⁵ , ⁶ , ⁷ In contrast, the ECG, an inexpensive and ubiquitous test, has the potential to be an excellent screening tool. Despite considerable work, conventional ECG interpretation has limited sensitivity to detect ventricular dysfunction. A recent study in adults demonstrated the ability to highlight key ECG features, which on multivariable linear regression explained 45% of data variability. ⁸ Additional work has shown markers of ventricular repolarization and delayed ventricular activation are key in identifying systolic dysfunction. ⁹ These findings represent a step forward in understanding the electrocardiographic findings suggestive of reduced ejection fracture and provide evidence that more advanced statistical methods may be able to accurately identify systolic dysfunction from an ECG.

The latest advances in artificial intelligence (AI) have subsequently paved the way for the development of neural networks capable of diagnosing LVSD from 12‐lead ECGs in adult patients with a high degree of accuracy. ¹⁰ , ¹¹ , ¹² However, to date, there are no studies of AI interpretation of the 12‐lead ECG to detect LVSD and RVSD in children, a group with distinct cardiac physiology and pathology compared with adults. We hypothesized that AI‐enabled ECGs would have the ability to accurately detect LVSD and RVSD in a pediatric population and that a newly trained pediatric model would outperform a model developed in an adult cohort for the detection of LVSD.

Methods

Data Source and Patient Population

A cohort of 10 142 pediatric patients (age≤18 years) who underwent a transthoracic echocardiogram and had a 10‐second, 12‐lead surface ECG within 14 days of the transthoracic echocardiogram at the Mayo Clinic from January 1, 2002 to June 30, 2022 was identified. Only the first transthoracic echocardiogram–ECG pair for each patient was included in the study. Subjects were excluded if they had a previous cardiac surgery or invasive cardiac intervention before their first echocardiogram or during the time between their echocardiogram and ECG. Subjects with single‐ventricle physiology were also excluded. A study flow diagram is provided in Figure S1.

ECGs were collected using a Marquette ECG machine (GE Healthcare, Chicago, IL) with a sampling rate of 500 Hz and stored using the MUSE data management system. The echocardiograms, including reports and images, when necessary, for each subject were retrieved from the Mayo Clinic Echo Image Management System. Left ventricular systolic function was measured quantitatively using left ventricular ejection fraction (LVEF). When multiple LVEF measurements were available, a standard hierarchy was used to select the first available in the following order: 3‐dimensional echocardiography ¹³ ; the Simpson biplane method ¹⁴ ; and the 2‐dimensional method in the parasternal long‐axis view, M‐mode measurement, or visual estimate. LVEF <50% and ≤35% were defined as any and severe LVSD, respectively. Right ventricular function was classified semiquantitatively as normal, mild, moderate, and severe dysfunction. Echo reports that did not include an assessment of right ventricular function were reviewed by a pediatric cardiologist (T.N.). Clinical data including prior surgery and diagnoses were obtained from the electronic medical record using electronic query and confirmed through chart review.

This study was approved by the institutional review board at the Mayo Clinic. The requirement for informed consent was waived for this minimal‐risk study, and all data were collected retrospectively. All requests for raw and analyzed data and related materials are available upon request to the corresponding author, and will be subject to review by the Mayo Clinic legal department and Mayo Clinic ventures to verify whether the request is subject to any intellectual property or confidentiality obligations. Any data and materials that can be shared will be released via a material transfer agreement.

Primary and Secondary Outcomes

Three novel AI models were trained separately in the pediatric cohort to detect the following primary outcomes: (1) severe LVSD with LVEF ≤35%, (2) any LVSD with LVEF <50%, and (3) RVSD of any severity. A fourth model was created for the detection of the secondary end point of presence of either any LVSD (LVEF <50%) or RVSD.

The following prespecified analyses were performed in the test cohort. The ability of the model trained to identify LVEF <50% to identify heightened risk of future LVSD (defined as LVEF <50%) on follow‐up echocardiograms among patients with normal LVEF on initial testing was evaluated. The performance of the RVSD model to detect moderate or greater RVSD was tested. The performance of the RVSD model was assessed in a subanalysis of the test cohort excluding individuals with Ebstein's anomaly, a rare congenital anomaly overrepresented in our cohort due to a tertiary referral practice. Finally, the performance of a previously developed model for the detection of LVEF ≤35% in adults was evaluated in the current pediatric cohort. ¹⁰

AI Model Development

The newly trained pediatric models used the same architecture as previously published models for age estimation and sex prediction for feature extraction and processing. ¹⁵ The output layer used a single sigmoid activation for LVEF binary classification instead of 2 outputs with SoftMax activation. ¹⁵ In short, these models are convolutional neural networks implemented in the Keras Framework with a TensorFlow backend and Python programming language. The model input is solely the raw 12‐lead ECG waveform data of 10 seconds at 500 Hz (size 5000 ×12). Hyperparameters were tuned in the retraining process including adjustments of the learning rate to 0.001 and dropout of 0.5. All models used the Adam optimizer with binary cross entropy loss. For model development, the cohort was divided into training, validation, and test cohorts using a 70‐10‐20 split.

The development and architecture of the adult model for detection of LVSD has been previously published. ¹⁰ The main difference between the pediatric model architectures presented here and that of the adult LVSD model is that the adult model takes in multiple 2‐second windows (size 1000 ×12) of ECG overlapped by 1 second over the 10‐second input and averages over 9 sequential windows. The newly developed pediatric models use a single input of 5000 ×12 samples instead.

Model Evaluation and Statistical Methods

The primary assessment of model performance was area under the curve (AUC) of the receiver operating characteristic curve, with 95% CIs. Additional measures of performance were sensitivity, specificity, positive predictive value, negative predictive value, and accuracy. Continuous data are presented as mean±SD or median and interquartile range as appropriate. Categorical data are presented as counts and percentages. As appropriate, statistical comparisons of cohorts were made using χ², ANOVA, and Student's t tests.

Among those patients with a normal LVEF on initial evaluation, we created cumulative incidence curves for the subsequent development of LVEF <50%, stratified by those whom the AI model first predicted as false positives or true negatives. We compared these 2 groups using a log‐rank test. Subsequently, a Cox proportional hazards model was developed. The additional variables in this model were age and sex. The Cox proportional hazards model met all assumptions including the proportional hazards assumption, which was assessed visually using scaled Schoenfeld residuals. All analyses were performed in Python using the pandas, ¹⁶ numpy, ¹⁷ scikit‐learn, ¹⁸ scipy, ¹⁹ tensorflow, ²⁰ lifelines, ²¹ and matplotlib ²² packages.

Results

The cohort included 10 142 children with a median age of 12.5 years (interquartile range, 4.6–15.8 years) and 5302 boys (52.3%). The median time between ECG and echo was 0.5 days (interquartile range, 0.4–2.0 days). Baseline cohort characteristics are presented in Table 1. The cohort included 131 patients with LVEF ≤35%, 256 patients with ejection fraction (EF) <50%, 471 patients with any degree of RVSD, and 253 patients with moderate or greater RVSD. There were 609 patients with either LVEF <50% or any degree of RVSD and 9533 patients with normal right and left ventricular systolic function. Demographic data of the entire cohort and stratified by the presence of LVSD and RVSD are presented in Table 1. Causes of LVSD and RVSD are outlined in Figures S2 and S3 respectively.

Table 1.

Patient Demographics for Study Cohort Stratified by LVEF

Patient characteristic	Full cohort	Normal LVEF (≥50%)	LVEF <50%	P value	Normal RV function	Any RV dysfunction	P value
Patients, n	10 142	9886	256		9670	472
Age, y, median (IQR)	12.5 (4.6–15.8)	12.6 (4.8–15.8)	3.5 (0.1–15.1)	<0.001	12.7 (5.1–15.9)	4.7 (0.1–12.9)	<0.001
Male sex, n (%)	5302 (52.3)	5168 (52.4)	134 (54.5)	0.98	5053 (52.2)	249 (52.8)	0.87
Race, n (%)				<0.001			<0.001
White	8113 (80.0)	7940 (80.3)	173 (67.6)		7780 (80.4)	335 (71s
Black	395 (3.9)	376 (3.8)	19 (7.4)		368 (3.8)	29 (6.1)
Asian	237 (2.3)	231 (2.3)	6 (2.3)		224 (2.3)	14 (3.0)
American Indian/Alaska Native	86 (0.8)	83 (0.8)	3 (1.2)		83 (0.9)	4 (0.9)
Native Hawaiian/Pacific Islander	16 (0.2)	16 (0.1)	0 (0)		16 (0.2)	0 (0)
Other	472 (4.7)	450 (4.6)	22 (8.6)		445 (4.6)	28 (5.9)
Unknown	823 (8.1)	790 (8.0)	33 (12.9)		756 (7.8)	62 (13.1)
LVEF, median (IQR)	62 (60–65)	63 (60–65)	35 (23–44)	<0.001	63 (60–65)	59 (49.5–64)	<0.001
RV dysfunction, n (%)	471 (4.6)	353 (3.6)	118 (48.0)	<0.001	0 (0)	472 (100)

IQR indicates interquartile range; LVEF, left ventricular ejection fraction; and RV, right ventricular.

Detection of LVSD

The novel model trained on pediatric data to detect LVEF ≤35% had excellent performance, with AUC of the receiver operating characteristic of 0.93 (95% CI, 0.89–0.98) and accuracy of 89%. The previously published adult model to detect LVEF ≤35% when applied to the pediatric cohort also performed well, albeit wth a lower accuracy: AUC was 0.87 (95% CI, 0.84–0.90) and accuracy 79%. When stratified by age, the adult model performed well among those aged ≥8 years with an AUC of 0.96 (95% CI, 0.94–0.99). Comparatively, performance on children <8 years was significantly lower, with an AUC of 0.74 (95% CI, 0.68–0.79) (Figure S4 and Table S1).

There was also excellent performance noted for the novel pediatric model to detect LVEF <50%, with an AUC of 0.88 (95% CI, 0.83–0.94) and accuracy of 80%. When stratified by age, this model performed well in those aged ≥8 years (AUC, 0.90 [95% CI, 0.82–0.97]), with comparable performance in those aged <8 years (AUC, 0.85 [95% CI, 0.76–0.93]). The receiver operating characteristic curves for all 3 LVSD models are presented in Figure 1, and additional parameters of model performance are included in Table 2.

Figure 1. Receiver operating characteristic curves for models for detection of LVSD.

Receiver operating characteristic curves for adult cohort derived LVSD model applied to pediatric cohort (EF ≤35%), novel pediatric model for detection of severe LVSD (EF ≤35%), and novel pediatric model for detecting any LVSD (EF <50%). AUC indicates area under the curve; and EF, ejection fraction; and LVSD, left ventricular systolic dysfunction.

Table 2.

Performance of Novel Models for the Detection of Decreased LVEF and RVSD in Children

Test statistic	LVEF ≤35	LVEF <50	RVSD	RVSD or LVEF ≤35	RVSD of LVEF <50
AUC (95% CI)	0.93 (0.89–0.98)	0.88 (0.83–0.94)	0.90 (0.87–0.94)	0.91 (0.88–0.93)	0.87 (0.83–0.90)
Accuracy	0.89	0.80	0.81	0.81	0.81
Sensitivity	0.93	0.85	0.88	0.88	0.80
Specificity	0.89	0.80	0.80	0.80	0.82
PPV	0.108	0.095	0.158	0.181	0.201
NPV	0.999	0.995	0.993	0.993	0.986

AUC indicates area under the curve; LVEF, left ventricular ejection fraction; NPV, negative predictive value; PPV, positive predictive value; and RVSD, right ventricular systolic dysfunction.

A subgroup of 613 patients from the test cohort of the pediatric LVSD (EF <50%) model with normal LV function (EF ≥50%) on initial echocardiogram were identified as having at least 1 follow‐up echocardiogram. Of these, 55 (8.9%) developed LVSD (EF <50%) during mean follow‐up of 4.2 (SD±4.7) years. This group was stratified by the initial AI model assessment as either being a true negative (model predicted as having EF ≥50%) or false positive (model predicted as having EF <50%). The Kaplan–Meier curve of cumulative incidence of LVSD during 5 years of follow‐up stratified by initial AI classification is presented in Figure 2. Using the log‐rank test, individuals initially classified by the AI model as having LVSD (false positive) were at higher risk of developing LVSD during follow‐up compared with those who were initially classified as normal LV function (true negative) (P<0.005; Table S2). A Cox proportional hazards model adjusting for age and sex demonstrated a >3‐fold risk of development of LVEF <50% over a 5‐year follow‐up period in those classified by the model as abnormal (false positive) compared with those classified as normal (true negative) (hazard ratio, 3.42 [95% CI, 1.88–6.21]; P<0.005).

Figure 2. Cumulative incidence of LVSD among those with normal EF at the time of initial classification, stratified by AI model classification.

The 5‐year cumulative incidence of LVEF <50% among those with normal EF at the time of initial classification, stratified by AI model classification as false positives or true negatives. Log‐rank test demonstrated a significant difference between the 2 groups (P<0.005). In a Cox proportional hazards model, there was a 3‐fold risk of future low LVEF for those who were initially categorized as false positives (age‐ and sex‐adjusted HR, 3.34 [95% CI,1.94–5.77]; P<0.005) compared with those who were classified as true positive. AI indicates artificial intelligence; EF, ejection fraction; LVEF, left ventricular ejection fraction; and LVSD, left ventricular systolic dysfunction.

Detection of RVSD

The novel pediatric model for detection of any RVSD demonstrated an AUC of 0.90 (95% CI, 0.87–0.94) and accuracy of 81% in the test cohort (Figure 3). The model performance statistics are presented in Table 2. The model performance for the detection of moderate or greater RVSD in the test cohort was better, with AUC of 0.94 (95% CI, 0.90–0.98) and an accuracy of 88%. The cohort included a significant number of children with the Ebstein anomaly (n=230), a rare congenital heart disease. To assess whether this affected the model's generalizability, we evaluated the model's ability to discriminate RVSD after excluding individuals with the Ebstein anomaly (Table S3). The AUC for identifying any degree of RVSD in a cohort without the Ebstein anomaly was 0.87 (95% CI, 0.82–0.93), while the AUC for identifying moderate or greater RVSD in a cohort without the Ebstein anomaly was 0.90 (95% CI, 0.82–0.97).

Figure 3. Receiver operating characteristic curves for the RVSD model.

Receiver operating characteristic curves for the RVSD model. Using the same test cohort, several conditions were assessed. First, the model was evaluated on its ability to identify only moderate or greater RVSD, which improved model performance. Next, the model was evaluated on any degree of RVSD, but with patients diagnosed with the Ebstein anomaly removed as this group was overrepresented in our cohort. Finally, the model was evaluated on its ability to identify moderate or greater RVSD in the cohort without the patients with the Ebstein anomaly. In both cases, removal of the patients with the Ebstein anomaly resulted in only a small drop in the AUC. AUC indicates area under the curve; and RVSD, right ventricular systolic dysfunction.

Combined Models of LVSD and RVSD

Neural networks were built to predict the presence of either LVSD or RVSD from ECG data on pediatric patients. The output was a binary prediction for the presence of any dysfunction and did not provide information on which ventricle was predicted to have systolic dysfunction. The models performed similarly to the individual models for RVSD or LVSD. When an EF cutoff of ≤35% was used for LVSD, the resultant model had an AUC of 0.91 (95% CI, 0.88–0.93). When a cutoff of EF <50% was used for LVSD, the model had an AUC of 0.87 (95% CI, 0.83–0.90). Full model performance statistics are presented in Table 2.

Discussion

We demonstrate that artificial intelligence, applied to 12‐lead ECGs can detect LVSD and RVSD with accuracy in a pediatric population, including mild degrees of ventricular dysfunction. The models developed in children demonstrate higher accuracy in detecting significant LVSD compared with a model derived in an adult population, highlighting unique differences in cardiac physiology and ECG features in childhood as compared with adults. ¹⁰ , ¹² Moreover, children with a false‐positive result when screening for LVSD had a >3‐fold higher risk for future development of low LVEF. ¹⁰ To our knowledge, this work represents the first application of AI‐enabled ECG to the detection of RVSD and LVSD in children. ²³

We present a pediatric‐specific AI‐enabled ECG algorithm that can detect the presence of severe LVSD (LVEF ≤35%) with high accuracy. These results parallel those of an algorithm developed for the detection of LVEF ≤35% in an adult population by our group. ¹⁰ Initial publication of this adult algorithm showed an AUC of 0.93 with a sensitivity, specificity, and accuracy of 86.3%, 85.7%, and 85.7%, respectively, subsequently prospectively validated with similar results. ²⁴ When directly applied to the pediatric population, this algorithm generally appeared to perform moderately well, with a small drop in performance. However, on detailed analysis, the adult model experiences a dramatic decline in performance in young children while performing very well in older children and teenagers. Conversely, the newly developed pediatric‐trained models demonstrated improved overall performance and more consistent performance across age groups. Differences in the pathogenesis of ventricular dysfunction in children compared with adults may result in differences in ECG signatures detected by the algorithm. Furthermore, age‐ and sex‐related changes in ECG features of depolarization and repolarization are well described in children making it imperative that AI ECG algorithms specifically be trained on pediatric cohorts. ²⁵ , ²⁶

In addition, we present an algorithm to detect mild LVSD with LVEF <50%, which is an important distinction from previous work done in adults to detect more severe LVSD. This is essential since many causes of LV dysfunction in children can present with subtly progressive dysfunction often associated with worse outcomes. ²⁷ , ²⁸ , ²⁹ Moreover, monitoring for the development of mild LVSD is also required in individuals with congenital heart defects, muscular dystrophies, or those receiving chemotherapy. ³⁰ , ³¹ Following rigorous validation, this AI ECG model has potential as an inexpensive and readily available screening tool for pediatric LVSD.

We demonstrate that pediatric patients falsely predicted to have LVSD on initial evaluation are at a higher risk of developing LVSD over a 5‐year follow‐up period, which mirrors previous findings in adults. ¹⁰ , ¹² These findings suggest that the algorithm may be identifying signals indicating elevated risk for progression to ventricular dysfunction. A significant number of patients who developed future LVSD did so in the post–cardiac surgery setting. Further work will be needed to investigate whether the algorithm can predict postoperative risk of LVSD specifically.

In addition to numerous studies evaluating the use of AI to detect LVSD, a recent study has shown the ability of AI to accurately diagnose RVSD and right ventricular dilation in adults as well. ¹² We demonstrate excellent performance of an AI ECG algorithm for the detection of RVSD in pediatric patients. Because the distribution of cardiac disease in children includes a high percentage of intracardiac shunts and pulmonary hypertension, screening for RVSD is of particular clinical importance in this group. ³² , ³³ This algorithm in particular demonstrated robust performance across age groups and RVSD severity.

These algorithms require further prospective validation in other pediatric cohorts. Following validation, potential applications include the diagnosis of left and right ventricular dysfunction in children presenting with symptoms and as a screening tool for early recognition of ventricular dysfunction. The inclusion of children seen in a wide range of settings in this study, including emergency department, ³⁴ inpatient, intensive care unit, ³⁵ and primary care clinic, can broaden the applicability of the algorithm for diagnosis and screening to a variety of settings. ³⁶

Our study has several important limitations. The cohort had an uneven age distribution, with a large number of infants and adolescents and a smaller number of school‐age children. However, this distribution is consistent with the frequency of cardiac testing across the range of pediatric patients and the age most diagnoses are made. Multiple epidemiologic studies have demonstrated that infancy is the most common age for onset of heart failure symptoms, followed closely by adolescence. ³⁷ , ³⁸ It follows that training our algorithm on a cohort with large numbers of infants and adolescents should correlate with good clinical performance. Second, RVSD was assessed semiquantitatively using echocardiography, which is subject to inter‐ and intraobserver variability. Quantitative assessment of right ventricular function using echocardiographic and other cross‐sectional imaging techniques, although more accurate, was not uniformly available. However, this is reflective of common clinical practice in pediatric cardiac imaging. Third, it should be noted that the adult and pediatric models for EF <35% differ in how the ECG data are presented to the neural network (a single 10‐second segment versus averaged 2‐second segments), which may influence model performance and affect their direct comparison. Additionally, external validation is needed to assess the performance of this algorithm in other pediatric cohorts with greater geographic and racial diversity to ensure generalizability of the algorithm. Prospective studies can also be used to assess the incremental diagnostic value of the AI ECG algorithms beyond routine clinical practice.

A final limitation of this work is the poor positive predictive value, which implies that this screening test may lead to a high number of false‐positive cases and a significant number of unnecessary echocardiograms in children. This is a common dilemma when developing screening tools for conditions with relatively low prevalence. However, using this tool in the appropriate clinical context may mitigate much of this concern. Given the high negative predictive value for all the models, the algorithm is best viewed as a tool to minimize expensive imaging studies in children presenting with signs or symptoms that might have otherwise triggered an echocardiogram. This will leverage the utility of this tool in correctly identifying true‐negative cases among symptomatic patients rather than attempting to perform broad screening.

In conclusion, our work demonstrates that applying AI to 12‐lead ECGs can accurately identify both RVSD and LVSD. These novel pediatric models were able to identify patients at an increased risk of LVSD during 5‐year follow‐up. This study highlights that while direct application of an adult algorithm for LVSD yields satisfactory results in pediatric populations, improved performance was noted when a new algorithm was trained for this task using pediatric data. As an inexpensive and readily available test, the ECG may serve as a powerful screening tool for ventricular dysfunction within the field of pediatric cardiology.

Sources of Funding

Dr Anjewierden has received support from the National Institutes of Health StARR Resident Investigator Award for this project. (National Institutes of Health 5R38HL150086‐02). The National Institutes of Health had no role in the design and conduct of the study.

Disclosures

Dr Attia has an ownership interest in Xai.health and serves as an advisor for Anumana.ai and AliveCor. Dr Lopez‐Jimenez is an advisor for Anumana, Novo Nordisk, and Wiseacre. He also receives royalties/patent beneficiary for Anumana, and he is a consultant for Kento Health. Dr Friedman reports competing interests with Anumana, Eko Health, AliveCor. M. Madhavan receives research funding from Boston Scientific and is a researcher for Biotronik Inc. The remaining authors have no disclosures to report.

Supporting information

Tables S1–S3

Figures S1–S4