Abstract
Introduction: The heterogeneous imaging characteristics of COPD, with at least emphysema and airway types, hinder automated characterization and the development of disease prediction methods. Probabilistic models, such as Gaussian Mixture Models (GMM), successfully model disease heterogeneity, under the assumption that underlying data can be generated from a mixture of gaussians. This study aims to develop and evaluate a GMM approach to predict COPD severity, defined by the GOLD stage.
Methods: Inspiratory computed tomography scans from 496 consecutive participants of the COSYCONET cohort study (65 ± 9 y.o.) were split into training (N=249), validation (N=123) and test set (N=124). Two state-of-the-art (SoA) methods and the GMM approach were compared to predict GOLD: a) percentage of low attenuation areas under -950 HU (LAA-950%); b) convolutional neural network (CNN) method1; c) GMM fitted on the mean lung density and LAA-950% of 3D lung patches (2.53 cm3) was first used to predict the presence of the disease, followed by patient-wise score aggregation for GOLD prediction. Area under the precision-recall curve (AUPRC) and area under the receiver operating characteristic (AUROC) are reported on the test set as an evaluation metric.
Results:
Conclusion: The GMM probabilistic approach shows to be the best GOLD stage discriminator, as it seems to capture the population distribution better than SoA baselines. This supports the idea that out-of-distribution methods may be a promising next step for the identification and grading of heterogeneous diseases, such as COPD.
1González, G. et al. Am J Respir Crit Care Med. 2018; 197(2):193-203
Footnotes
Cite this article as Eur Respir J 2022; 60: Suppl. 66, 2421.
This article was presented at the 2022 ERS International Congress, in session “-”.
This is an ERS International Congress abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).
- Copyright ©the authors 2022