Automated Echocardiographic Detection of Congenital Heart Disease Using Artificial Intelligence

28 Mar 20264 min read0 viewsJournal Feed

GIST

Circulation, Ahead of Print. Background: Delayed or missed diagnosis of congenital heart disease (CHD) contributes to excess pediatric mortality worldwide.

Echocardiography (echo) is central to diagnosing and triaging CHD, yet expert interpretation remains a scarce and maldistributed global resource. Artificial intelligence (AI) offers the potential to democratize diagnostics and extend expert-level interpretation beyond large academic centers, but its application in CHD remains underexplored.Methods: We developed EchoFocus-CHD, an AI-enabled model for automated detection of 12 critical and 8 non-critical CHD lesions, individually and as composites.

The composite critical CHD outcome was the primary endpoint. The model expands on a multi-task, view-agnostic architecture (PanEcho) with a transformer encoder to improve focus on relevant echo views.

Clinical Editorial

Study design and model scope

Researchers developed EchoFocus-CHD, an AI-enabled system for automated detection of 12 critical and 8 non-critical CHD lesions, evaluated individually and as composite outcomes.
The primary endpoint was the composite critical CHD outcome.
The approach builds on a multi-task, view-agnostic framework (inheriting elements from PanEcho) and incorporates a transformer encoder to emphasize clinically relevant echo views.

Data cohorts and settings

External evaluation involved echo studies from external US and international centers.

Internal cohort: 3.4 million video segments from 54,727 echocardiograms, median patient age 7.1 years, with 5.8% having critical CHD and 23.6% non-critical CHD.
External referral cohort: 167,484 videos from 3,356 echocardiograms, median patient age 2.5 years, with 29.4% critical CHD and 45.6% non-critical CHD.
Training used 80% of the internal data; testing used the remaining 20% of the internal set.

Performance signals and limitations

Retraining with all available US patient data reduced some domain shift artifacts, improving international critical CHD detection (AUROC 0.87) and calibration.

Internal performance for the composite critical CHD outcome was strong (AUROC 0.94; LR+ 7.50; LR- 0.14).
Performance for individual critical lesions ranged from high to excellent (AUROC 0.83–1.00).
For composite non-critical CHD, AUROC was 0.90 (LR+ 5.00; LR- 0.23); individual non-critical lesions ranged from AUROC 0.70–0.96.
On the referral cohort, accuracy for detecting critical CHD declined (AUROC 0.77), matching higher expert disagreement in those cases (kappa 0.72 versus 0.82 for internal cases).
Explainability analyses indicated the model consistently prioritized key views (parasternal long-axis, parasternal short-axis, subxiphoid long-axis, apical) in both cohorts.
UMAP suggested a domain shift between internal and referral data.

Interpretation and context

EchoFocus-CHD demonstrates potential for automated CHD detection and could help extend expert-level interpretation beyond major centers.
The study emphasizes domain shift as a challenge when applying the model to external/referral datasets and notes that external validation is necessary before real-world deployment.
Improved performance after retraining on broader US data suggests some generalizability gains, yet real-world applicability hinges on addressing cross-site variability and achieving robust external validation.

Open questions and operational implications

The exact impact of domain shift on individual lesion detection in diverse clinical settings remains to be fully characterized.
External validation across broader international centers is required to confirm generalizability and calibration.
The balance between automated detection and expert review in triage workflows will need careful integration, given differences in appearance of referral cases and potential disagreement among specialists.