(VP077) NEW DEEP LEARNING APPROACH FOR AUTOMATED DIAGNOSIS OF ATRIAL FIBRILLATION BY ECHOCARDIOGRAPHY WITHOUT ECG

Background: Artificial intelligence (AI) has been shown to be capable of assessing cardiac structure and function on echocardiography, yet there is limited literature on its application towards automated rhythm detection on imaging. This can be of clinical relevance in point-of-care ultrasound (POCUS) which lacks rhythm strips and may improve workflow in clinical settings. We trained an AI model proficient in detecting atrial fibrillation (AF) through 2D echocardiography imaging without requiring ECG data.

METHODS AND RESULTS: Transthoracic echocardiography studies of consecutive patients ≥18 years old at our tertiary care centre were retrospectively reviewed for AF and sinus rhythm. Apical 4-chamber cines with three cardiac cycles, interpreted by level III echocardiographers as the gold standard based on imaging assessment and rhythm strip, were also verified with a 12-lead ECG done within 30 days to form datasets of agreeing or conflicting rhythm labels between the modalities. Agreeing datasets were introduced to the deep learning model ResNet(2+1)D with an 80-10-10 train-validation-test split ratio. Precision, recall, F1-scores, and AUROC scores were calculated to assess the AI model’s performance in accurately diagnosing the rhythm. Furthermore, a qualitative analysis was conducted by assessing the model’s sensitivity to spatio-temporal features on cines, which was transformed into heatmaps and evaluated for clinical relevance in rhythm detection.

634 patient studies (1205 cines) with agreeing rhythm label datasets were included (FIGURE 1). After training, the AI model achieved high accuracy on validation for detection of both AF and sinus rhythm (mean F1-score = 0.92; AUROC = 0.95) with consistent performance on the test dataset (mean F1-score = 0.94; AUROC = 0.98). The AI model was further evaluated on a subset of same-day echocardiogram studies and ECGs with agreeing or conflicting rhythm labels (n=40). Using the echocardiographer’s assessment as the gold standard, this resulted in a mean F1-score of 0.92 and AUROC of 0.99. Generated heatmaps indicate that the AI model is perceiving differences in atrial structure and function between cines in AF compared to sinus rhythm, with a particular focus on irregularities in the atrioventricular valves (FIGURE 2). This suggests that the model is potentially using echocardiographic features of atrial remodelling to phenotype AF.

Conclusion: AF detection by AI on echocardiography without ECG appears accurate when compared to an echocardiographer’s assessment as the gold standard. This has potential clinical implications in POCUS for detection of early or paroxysmal AF missed on conventional ECG and adds to the armamentarium for stroke risk stratification.