Deep-learning networks accurately detect pulmonary hypertension in patients with idiopathic pulmonary fibrosis.

Background: There is currently no widely used noninvasive tool that accurately predicts pulmonary hypertension (PH) in patients with idiopathic pulmonary fibrosis (IPF). We aimed to assess the feasibility of developing deep-learning algorithms for detecting PH in patients with IPF. Methods: A retrospective derivation cohort included patients with IPF and PH and patients with IPF without PH, matched for age, sex and forced vital capacity (FVC%). Individual echocardiographic frames were split into a training and a validation set. The network consisted of three connected two-dimensional convolutional subnetworks; each one accepted an apical four-chamber view, a parasternal short-axis view and a lung computed tomography view from the same individual, respectively. The output was then merged with serum brain natriuretic peptide levels and diffusion lung capacity of carbon monoxide. Results: Overall, 65 patients with PH and IPF (68.6±7.2 years, 75.4% female, FVC 62.7%) and 65 matched patients without PH were included in the derivation cohort. A training set of 122 patients with and without PH revealed a >99% accuracy in detecting PH. A validation set demonstrated 87.5% accuracy per patient for correctly detecting PH (area under the curve (AUC) 0.971, 95% CI 00.521-1.000; p=0.0004; sensitivity 83.3%, specificity 100%). An external testing set of six patients with IPF and PH and six matched patients without PH showed 87.5% accuracy per patient (AUC 0.844, 95% CI 0.540-0.992; p=0.03; sensitivity 100%, specificity 75.0%). Conclusions: Deep-learning algorithms can accurately predict PH in patients with IPF. Further validation in a larger external cohort is warranted.