Deep learning interpretation of echocardiographic images predicts incident heart failure and subtypes.

BACKGROUND: Accurate prediction of incident heart failure (HF) may help prioritize HF preventive therapies. Deep learning interpretation of echocardiograms may improve HF risk prediction beyond clinical risk models. We trained and validated a deep learning model to predict incident HF from transthoracic echocardiographic images (Echocardiogram-to-Heart Failure, "Echo2HF") METHODS: Echo2HF was developed using 4,057,664 echocardiogram videos from 70,763 patients receiving longitudinal ambulatory care at Massachusetts General Hospital (MGH). Performance for 10-year incident HF was evaluated in an internal MGH test set and an external test set of 34,802 individuals without prevalent HF from Brigham and Women's Hospital (BWH). Model performance was evaluated using area under the receiver operating characteristic curves (AUROC) and compared with the Pooled Cohorts Equations to Prevent Heart Failure (PCP-HF) and the Predicting Risk of cardiovascular disease EVENTs (PREVENT) clinical risk scores. RESULTS: Echo2HF was trained in 64,167 individuals and evaluated in a hold-out sample of 6,394 individuals from MGH (279 HF events, age 62 ± 17 years, 48% women) and 34,802 individuals from BWH (1280 events, age 62 ± 15 years, 56% women). Echo2HF discriminated incident HF, with 10-year AUROC of 0.83 (95% CI0.81-0.85] and 0.82 (95% CI 0.81-0.83) at BWH, with numerically higher discrimination versus both PCP-HF and PREVENT. CONCLUSION: Deep learning analysis of echocardiograms accurately discriminated future HF risk, with favorable performance over current clinical HF scores. Future work should assess whether broader use of artificial intelligence-enabled echocardiographic risk stratification may improve HF prevention and clinical outcomes, including among individuals who do not have a clinical indication for echocardiography.