Abstract 13261: Utility of a Novel Ensemble Based Deep Learning Network for the Automatic Detection of Pulmonary Hypertension and Right Ventricular Dilatation: A Study Based on Data From 783 Individuals From Two Tertiary Centres

Abstract

Background: Pulmonary arterial hypertension (PAH) is associated with high morbidity and mortality. Echocardiography represents the main screening tool but requires specific expertise and diagnosis is often delayed. We hypothesized that an ensemble of deep learning (EDL) networks could reliably detect PAH in routine echocardiograms and improve patient management. Methods: Consecutive patients with PAH, Tetralogy of Fallot (TOF) or atrial septal defects (ASD) and right ventricular (RV) dilation, as well as normal controls were included. An ensemble of deep convolutional networks incorporating apical and parasternal views as well as tricuspid regurgitation gradients (TRG) into one unified network was trained to classify patients by presence of (invasively confirmed) PAH. Results: Overall, 408 PAH patients (median age 59 years, 74% female; 34% idiopathic, 66% associated PAH, RV inflow diameter 4.7±0.9 cm), 308 patients with RV dilatation (201 TOF and 107 ASD patients (med. age 41 years, 81% female, RV-inflow diameter 4.6±1.0 cm) and 67 normal controls (med. age 49 years, RV-inflow 3.4±0.6 cm) were included. Overall, 84% of PAH patients had usable TRGs and the sensitivity of detecting PAH by conventional RV diameters and TRG (estimated RV systolic pressure >40 mmHg) was 68.6% (81.7% in patients with adequate TRGs). In contrast, only 35% of normal controls and 64% of RV dilatation patients had interrogable TRGs. The EDL algorithm trained on 53,832 frames and validated on 26,474 frames not used for training achieved an on-frame accuracy of 95.0% in the validation set. On a per patient basis all normal controls and PAH patients were correctly classified, while two patients with RV-dilatation, elevated RV pressure and poor image quality were assigned to the PAH group, yielding an overall accuracy of 97.6% and a sensitivity of detecting PAH of 100%. Conclusions: The study is the first to demonstrate the utility of ensemble based deep learning algorithms, trained on routine echocardiograms, to detect PAH based on large tertiary center data. These algorithms outperform by far conventional echocardiographic evaluation for detecting PAH. Due to the automated process involved, they could be deployed to the community and contribute to the earlier diagnosis of PAH.