An artificial intelligence (AI)–enabled echocardiography model demonstrated high accuracy for detecting congenital heart disease (CHD), with reduced performance in external datasets. In this model development and validation study published in Circulation, EchoFocus-CHD was developed to automate detection of 12 critical and 8 non-critical CHD lesions using echocardiographic data. The primary endpoint was detection of composite critical CHD.

The analysis used echocardiograms from Boston Children’s Hospital for internal training (80%) and testing (20%), including 3.4 million videos from 54,727 studies (median age 7.1 years; 5.8% critical CHD). External evaluation included 167,484 videos from 3,356 studies (median age 2.5 years; 29.4% critical CHD). These findings are relevant for settings with limited access to expert echocardiography interpretation.

EchoFocus-CHD achieved high internal accuracy for composite critical CHD (area under the receiver operating characteristic curve [AUROC] 0.94; positive likelihood ratio [LR+] 7.50; negative likelihood ratio [LR−] 0.14) and non-critical CHD (AUROC 0.90). Performance declined in the referral cohort (AUROC 0.77 for critical CHD), with greater expert disagreement (κ=0.72 vs 0.82 internally). Individual lesion detection ranged from AUROC 0.83 to 1.00 for critical CHD and 0.70 to 0.96 for non-critical CHD.

Explainability analyses showed consistent prioritization of clinically relevant echocardiographic views across datasets, while dimensionality reduction identified domain shift between cohorts. Retraining on expanded US datasets improved international performance (AUROC 0.87) and calibration. 

These findings indicate that automated CHD detection is feasible, although external validation and domain variability remain important considerations before clinical implementation.