Abstract 14817: Machine Learning Computer Vision for Point of Care Decision Support of Echocardiographic Identification of Hypertrophic Cardiomyopathy

Sreekanth VemulapalliFanjie KongFawaz AleneziAnna GiczewskaLillian KangKaren ChiswellRicardo HenaoAndrew Wangduke university,Durham,NCDuke Clinical Rsch Institute,Durham,NCDuke Univ,Durham,NC

DOI: https://doi.org/10.1161/circ.146.suppl_1.14817

IF: 37.8

2022-11-01

Circulation

Abstract:Circulation, Volume 146, Issue Suppl_1, Page A14817-A14817, November 8, 2022. Introduction:Hypertrophic cardiomyopathy (HCM) has an estimated prevalence of 1/500 but remains underdiagnosed. Echocardiography is the primary diagnostic modality but requires recognition of varied patterns of hypertrophy. Point of care tools to facilitate improved identification of HCM are needed.Methods:We identified 400 study cases of HCM in adults meeting professional society guideline diagnostic criteria from 2000-2020 at Duke University Medical Center. 900 study controls matched on age, sex, year of echo, and ejection fraction were also identified. Multi-vendor echocardiographic examinations performed within 12 months of initial diagnosis were used to train a convolutional neural network (CNN) to identify HCM. The model consisting of a 18-layer deep R(2+1)D residual network was developed on 60% of the data (training), 20% of the data (validation) was used to select model hyperparameters and to monitor for overfitting, and the remaining 20% (test) was used for performance estimation. CNN performance was evaluated by the receiving operating characteristic (ROC) and precision-recall curve and their summaries, namely, area under the ROC (AUROC) and average precision (AP,i.e., average positive predictive value). We report performance characteristics at the study level.Results:Among 400 HCM cases and 900 controls, the median age was 65, with 48.1% male. Hypertension was present in 44.9%, with heart failure in 16.3% and Atrial Fibrillation in 12%. Median EF was 55% with 14.5% having an EF < 50%. Median diastolic dimension was 4.4 cm and moderate or severe MR was present in 8.4%. Study-level AUCROC for the CNN was 0.853 and the average precision was 0.698 (Figure 1).Conclusion:A machine learning CNN is able identify study-wise HCM with an AUC of 0.853. Further work will be necessary to ensure explainability of results and similar multi-institutional performance as a point of care tool for the echocardiographic identification of HCM on diagnostic rates.

cardiac & cardiovascular systems,peripheral vascular disease