Deep neural network analyses of spirometry for structural phenotyping of chronic obstructive pulmonary disease

摘要: BACKGROUNDCurrently recommended traditional spirometry outputs do not reflect the relative contributions of emphysema and airway disease to airflow obstruction. We hypothesized that machine-learning algorithms can be trained on data identify these structural phenotypes.METHODSParticipants enrolled in a large multicenter study (COPDGene) were included. The points from expiratory flow-volume curves using deep-learning model predict phenotypes chronic obstructive pulmonary (COPD) CT, results compared with metrics an optimized random forest classifier. Area under receiver operating characteristic curve (AUC) weighted F-score used measure discriminative accuracy fully convolutional neural network, forest, phenotype CT as normal, emphysema-predominant (>5% emphysema), airway-predominant (Pi10 > median), mixed phenotypes. Similar comparisons made for detection functional small (>20% parametric response mapping).RESULTSAmong 8980 individuals, network was more accurate discriminating predominant emphysema/airway (AUC 0.80, 95%CI 0.79-0.81) measures spirometry, FEV1/FVC 0.71, 0.69-0.71), FEV1% predicted 0.70, 0.68-0.71), classifier 0.78, 0.77-0.79). also emphysema/small 0.91, 0.90-0.92) 0.78-0.82), 0.83, 0.80-0.84), comparable 0.90, 0.88-0.91).CONCLUSIONSStructural COPD identified approaches, demonstrating their potential individuals targeted therapies.TRIAL REGISTRATIONClinicalTrials.gov NCT00608764.FUNDINGThis supported by NIH grants K23 HL133438 R21EB027891 American Thoracic Foundation 2018 Unrestricted Research Grant. COPDGene is NHLBI U01 HL089897 HL089856. (NCT00608764) through Industry Advisory Committee comprising AstraZeneca, Boehringer-Ingelheim, GlaxoSmithKline, Novartis, Sunovion.