Diagnostic or prognostic study
Complete AI-Enabled Echocardiography Interpretation With Multitask Deep Learning
- Diagnostic accuracy study
- development and retrospective, multisite validation
- internal temporal validation
- 4 external validation cohorts
Authors' conclusion
In this study, an AI system that automatically interprets echocardiograms maintained high accuracy across geography and time from complete and limited studies.
Study context
- Population
- TTE studies at Yale New Haven Health System hospitals and clinics during routine care (development January 2016 to June 2022; internal validation July to December 2022), plus 4 external cohorts
- Participants
- 24,405 participants · 32,265 TTE studies
- Comparison
- Artificial intelligence (AI) system vs interpreting cardiologist's assessment
- Outcome
- AUC for diagnostic classification tasks and mean absolute error for parameter estimation tasks, AI vs interpreting cardiologist
Performance
As reported in the abstract
Diagnostic classification (AUC) and parameter estimation (normalized mean absolute error) vs interpreting cardiologist
18 diagnostic classification tasks, median (IQR) AUC
0.91 (0.88-0.93)
21 echocardiographic parameters, median (IQR) normalized mean absolute error
0.13 (0.10-0.18)
Further findings
LVEF estimation, mean absolute error
- Internal validation
- 4.2%
- External validation
- 4.5%
Moderate or worse LV systolic dysfunction, AUC
- Internal validation
- 0.98
- External validation
- 0.99
RV systolic dysfunction, AUC
- Internal validation
- 0.93
- External validation
- 0.94
Severe AS, AUC
- Internal validation
- 0.98
- External validation
- 1.00
Abbreviated TTE cohort, 15 diagnosis tasks, median (IQR) AUC
- Median (IQR) AUC
- 0.91 (0.87-0.94)
Point-of-care ultrasonography (ED acquisitions), 14 tasks, median (IQR) AUC
- Median (IQR) AUC
- 0.85 (0.77-0.87)
Original publication
Complete AI-Enabled Echocardiography Interpretation With Multitask Deep Learning. JAMA. Published 22 July 2025. doi:10.1001/jama.2025.8731
Summarised from the published abstract; every number is checked against it