A fully automated deep learning framework for assessing left ventricular systolic function from coronary angiography with validation in acute myocardial infarction

Background Left ventricular ejection fraction (LVEF) is a key determinant of prognosis and therapeutic decision-making in patients undergoing coronary angiography (CAG). However, echocardiography may not be immediately available during catheterization, particularly in acute myocardial infarction (AMI). Whether LVEF can be reliably inferred directly from routine CAG in a fully automated manner remains uncertain. Purpose To develop and validate a fully automated end-to-end deep learning framework for predicting reduced LVEF from CAG videos, and to assess its robustness in an independent AMI cohort. Methods We retrospectively analyzed 4,256 CAG paired with transthoracic echocardiography within 7 days. Reduced systolic function was defined as LVEF <40%. A fully automated pipeline was implemented, incorporating diagnostic-content filtering, quality control, and exam-level aggregation. A SlowFast-based video model combined with adaptive data-repair optimization was trained to generate study-level predictions. Performance was evaluated on a held-out internal test set and an independent external AMI cohort (n=74). Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals. Results On the internal test set, the late-phase configuration achieved an AUC of 0.89, with a sensitivity of 0.72 and a specificity of 0.91. Performance remained stable across sex, age, and comorbidity subgroups (all AUC >0.84). Robust discrimination was preserved when inference was restricted to a subset of projections (AUC 0.87–0.88), supporting stability under heterogeneous acquisition conditions. In the external AMI cohort, overall discrimination remained consistent (AUC 0.86). Conclusion A fully automated end-to-end deep learning framework can assess left ventricular systolic dysfunction directly from routine CAG with robust performance across acquisition variability and external validation in AMI.