Artificial intelligence-enhanced electrocardiography (AI-ECG) has quickly advanced from a research novelty to a clinically promising tool. Originally developed in adult populations without congenital abnormalities [1,2], its application to congenital heart disease (CHD) - a field characterized by unique anatomical and electrophysiological challenges - has been relatively unexplored. Mayourian et al. [3] now deliver a major milestone with their multicentre study demonstrating that deep learning applied to ECGs can accurately detect and even predict left ventricular systolic dysfunction (LVSD) in paediatric and adult patients with CHD.
The foundation for AI-ECG in detecting LVSD was laid by studies like Attia et al. [1], who demonstrated its ability to screen for asymptomatic systolic dysfunction in adults with preserved ejection fraction. More recently, models tailored to paediatric cohorts without major CHD showed similar promise [4]. However, generalising these approaches to CHD has been limited due to the substantial heterogeneity of lesions, surgical histories, and ECG morphologies.

Mayourian et al. [3] address this gap by training and validating a convolutional neural network (CNN) on over 200,000 ECG-echocardiogram pairs from nearly 100,000 patients at two leading US paediatric cardiac centres. Their model achieved excellent discrimination (AUROC 0.95–0.96) and calibration for detecting LVEF ≤40%, despite the low prevalence of LVSD (<3% of the cohort). This performance held even in external validation, positioning this study as the most comprehensive AI-ECG application in the CHD field to date.

A notable strength of this work lies in its prognostic implications. AI-ECG stratification correlated strongly with future development of LVSD (HR 12.1 for high- vs low-risk patients) and all-cause mortality. This longitudinal value is consistent with recent work by Khurshid et al. [5], who showed that deep learning ECG interpretations could forecast atrial fibrillation risk years in advance. The current study extends this predictive paradigm to LVSD in a uniquely vulnerable population, supporting the concept of AI-ECG as a risk monitoring tool across the lifespan.

One frequent criticism of AI in medicine is its black-box nature. To address this, the authors used saliency mapping and median waveform analysis, revealing physiologically plausible high-risk ECG features such as deep S waves in V2 and T wave inversions in V6 - findings that align with previous works [6].
Furthermore, subgroup analyses showed consistent performance across common lesions like tetralogy of Fallot and ventricular septal defect, but reduced performance in rarer anatomies such as hypoplastic left heart syndrome (HLHS) and L-loop transposition of the great arteries (TGA). Similar limitations have been reported in other deep learning studies that rely on underrepresented subgroups [7].
The findings by Mayourian et al. compare favourably with prior work by Sangha et al. [6], who used ECG image inputs rather than digital waveforms to predict LVSD in adults. While this image-based approach facilitates deployment in lower-resource settings, waveform-based models like Mayourian's offer higher resolution and potentially better performance when data infrastructure is available.
Another recent contribution by Mayourian et al. [8] explored AI-ECG prediction of biventricular dysfunction in CHD using cardiac MRI - complementing the present study and affirming the versatility of this modelling approach across imaging modalities and clinical scenarios.

Given the increasing adult CHD population, AI-ECG presents a scalable, inexpensive tool for routine surveillance - especially where echocardiography is inaccessible or impractical. High-risk flags could prompt timely echocardiography or early initiation of guideline-directed medical therapy, while low-risk profiles could safely extend follow-up intervals. Importantly, prospective implementation studies and cost-effectiveness analyses - like those initiated by Yao et al. [9] - will be crucial to guide clinical integration.

In conclusion, this study sets a new benchmark for AI-ECG in CHD, offering a tool not only for cross-sectional screening but also for longitudinal risk prediction. Its potential to democratise access to cardiac surveillance in complex CHD, particularly in resource-constrained environments, is substantial. Future directions should include broader multicentre validation, federated learning to enhance performance in rare lesions, and prospective trials to assess impact on outcomes.