Machine learning of spirometry may predict CLAD onset after lung transplantation

<bold>Rationale:</bold> As Chronic Lung Allograft Dysfunction (CLAD) remains virtually unpredictable in clinical practice, we investigated the potential of routinely measured spirometry to predict CLAD onset, using machine learning. <bold>Methods:</bold> Of 1033 lung transplant recipients (transplanted 2010-2020), 383 random Controls without CLAD, and 186 CLAD patients (140 Bronchiolitis Obliterans/BOS - and 46 Restrictive Alllograft Syndrome/RAS) were included. Multiple measurements per patient were included using sliding windows, and spirometry parameters (FEV1, FEF25-75 and FVC; in Liters) were extracted from the patients’ electronic files. Time to CLAD was assessed as continuous response. Controls were censored at their last measurement. K-nearest neighbor (KNN) and random forest regression were compared, after which samples were stratified based on predicted time to CLAD onset. <bold>Results:</bold> KNN showed superior results compared to random forests, using 160 neighbors. Fig. 1a shows the fit between KNN-predicted and observed days to CLAD on our test set (R²=0.18). There were significant differences in CLAD-free survival between 5 stratified groups (Fig. 1b). <bold>Summary:</bold> Using machine learning on spirometry, lung transplant recipients may be stratified to risk groups with significant differences in subsequent time to CLAD onset. This allows to predict the risk of CLAD for each patient in real time.