Chronic lung allograft dysfunction phenotype and prognosis by machine learning CT analysis

BackgroundChronic lung allograft dysfunction (CLAD) is the principal cause of graft failure in lung transplant recipients and prognosis depends on CLAD phenotype. We used machine learning computed tomography (CT) lung texture analysis tool at CLAD diagnosis for phenotyping and prognostication compared to radiologists’ scoring.MethodsThis retrospective study included all adult first double-lung transplant patients (01/2010–12/2015) with CLAD (censored 12/2019) and inspiratory CT near CLAD diagnosis. The machine learning tool quantified ground-glass opacity, reticulation, hyperlucent lung, and pulmonary vessel volume (PVV). Two radiologists scored for ground-glass opacity, reticulation, consolidation, pleural effusion, air trapping and bronchiectasis. Receiver operating characteristic curve analysis was used to evaluate the diagnostic performance of machine learning and radiologist for CLAD phenotype. Multivariable Cox proportional-hazards regression analysis for allograft survival controlled for age, sex, native lung disease, cytomegalovirus serostatus, and CLAD phenotype (bronchiolitis obliterans syndrome [BOS] and restrictive allograft syndrome [RAS]/mixed).Results88 patients were included (57 BOS, 20 RAS/mixed, and 11 unclassified/undefined) with CT a median 9.5 days from CLAD onset. Radiologist and machine learning parameters phenotyped RAS/mixed with PVV as the strongest indicator (AUC 0.85). Machine learning hyperlucent lung phenotyped BOS using only inspiratory CT (AUC=0.76). Radiologist and machine learning parameters predicted graft failure in the multivariable analysis, best with PVV (HR=1.23, 95%CI 1.05–1.44, p=0.01).ConclusionsMachine learning discriminated between CLAD phenotypes on CT. Both radiologist and machine learning scoring were associated with graft failure, independent of CLAD phenotype. PVV, unique to machine learning, was the strongest in phenotyping and prognostication.