Figure 2.

Detection of chronic obstructive pulmonary disease (COPD) by convolutional neural network (CNN) in the COPDGene cohort. (A) Receiver operating characteristic curve, C statistic, and summary table for the CNN prediction of COPD in the COPDGene testing cohort. Clinical COPD was defined based on FEV₁/FVC less than 0.7. The CNN defined COPD based on CNN predicted probability of COPD greater than 0.5. (B) The predicted probabilities are the predicted probability of the outcome (COPD) by the CNN. The observed proportions are the observed proportions of participants in that decile who had the outcome. Reference lines indicate perfect correlation (slope = 1; intercept = 0). The Hosmer-Lemeshow test is a test for evidence of poor calibration. That is, a nonsignificant P value (>0.05) indicates no evidence for poor calibration.