**Authors and Affiliations**
Behrouz Rostami, Mohammad Sarraf, Eunjung Lee, Zachi Attia, Jae Oh, Kenneth Fetterly, Paul Friedman, Rajiv Gulati, and Mohamad Alkhouli.
Department of Cardiovascular Medicine, Mayo Clinic, Rochester, MN

**Background**
AI can be used to extract high level information out of medical images. Deep Learning has shown promising results when applied on medical video analysis applications. In this study, a 3D CNN was used to detect left ventricular diastolic dysfunction. The proposed algorithm classifies the patients into normal or abnormal LVDF groups.

**Methods**
The dataset included 6,114 Cath Lab angiogram studies from 6,093 patients collected between the years 2015 to 2021. Mean age of the patients is 67.96, and 39% of the patients are women. A 3D DCNN (X3D) architecture was used to classify the input angiogram video clips. Two video clips of LCA from LAO-caudal and RAO-caudal projections were selected per study as inputs during model training and evaluation. The trained model then was applied on unseen test set samples. The hypothesis is that the model would be able to capture very high-level features out of the angiograms and relate them to the LVDF target label with a high performance.

**Results**
The dataset was split into train, validation, and test set in patient level with the proportion of 70%, 15%, and 15% of the whole cohort, respectively. The testing set AUC was 0.86 with a sensitivity of 0.779 and specificity of 0.807.

**Conclusion**
This study introduces a deep learning methodology designed for the detection of left ventricular diastolic dysfunction among patients. Demonstrating robust performance in the classification of normal and abnormal Cath lab angiograms, the proposed approach holds significant promise. As a potential decision support system, this model has the capacity to enhance the efficiency of decision-making processes for cardiologists involved in treatment procedures.



Automated Assessment of Left Ventricular Diastolic Function from Coronary Angiograms with Video-Based Deep Learning Algorithms