Deep learning approach for probabilistic pulmonary function estimation from chest X-ray and peak expiratory flow rate
Christoph Killing1, Maximilian Wekerle2, Jayne S Sutherland3
1Institute of Infectious Diseases and Tropical Medicine, LMU University Hospital, LMU Munich, Munich, Germany. christoph.killing@med.uni-muenchen.de.
Background:
Spirometry remains the gold standard for assessing pulmonary function. Deep learning models have demonstrated potential for estimating measurements from chest X-rays (CXR). We aim to effectively address anatomical variability and integrate probabilistic reasoning to enhance estimation reliability near diagnostic thresholds.
Methods:
We developed a probabilistic machine learning framework to estimate the forced expiratory volume in the first second (FEV1) and the forced vital capacity (FVC) as measured through spirometry. Estimations use morphologically regularized CXRs and anthropometric-normalized peak expiratory flow rate (PEFR) as proxy for volumetric information unavailable in imaging. By estimating FEV1 and FVC z-scores, we decouple appearance from anatomic variability. We demonstrate our method on a multi-national cohort of pulmonary tuberculosis patients exhibiting diverse structural abnormalities and ventilatory impairments.
Results:
Using ensembles of neural networks, we analyze 982 CXR and spirometry pairs from 568 individuals. The best model achieves an area under curve (AUC) of 0.879 (FEV1; 99%CI 0.876, 0.881) and 0.853 (FVC; 99%CI 0.850, 0.856) in identifying moderate or severe lung-function impairment on a previously unseen test-set, signifying an AUC improvement of 0.144 (FEV1) and 0.118 (FVC) over previous methods. When allowing up to 10% of samples to remain unclassified due to uncertainty, AUC further rises to 0.894 (0.891, 0.896) and 0.857 (0.854, 0.860), respectively. Our method performs robustly across diverse impairment types and CXR pathologies.
Conclusions:
Our study shows that decoupling anatomical variability from functional assessment improves lung function estimation. Incorporation of probabilistic modeling improved diagnostic reliability around a decision threshold. Therefore, our system offers a promising approach to practical lung function estimation in settings where spirometry is unavailable.
