Deep learning automates complete quality control of spirometric manoeuvre

Abstract

Background: ATS/ERS spirometric acceptability criteria includes recommendations on forced expiration (FE) only. Unstable tidal breathing (TB) and insufficient forced inspiration (FI) also affects the validity of clinical indices like FVC. Further, these recommendations require visual inspection of spirograms, which introduces large inter-technician variability.

Objectives: We developed a deep learning model called convolutional neural network (CNN) to determine spirometry acceptability based on the complete manoeuvre that included TB, FI and FE.

Methods: Three experienced technicians independently labelled the acceptability of 1,325 complete manoeuvres using ATS/ERS 2005 protocols and a majority opinion established gold standard. We processed flow-volume loops into images, calculated ATS/ERS quantifiable criteria and developed a CNN on these features. Model development was done on 1000 curves by recalibrating a CNN, previously trained on 13,000 curves with acceptability labels from a junior technician, to predict the gold standard. We used the remaining 325 curves as a test-set.

Results: Around half of 1,325 manoeuvres met ATS/ERS acceptability. On testing (N=325), CNN demonstrated an accuracy of 87% with a high sensitivity (95%) and moderate specificity (80%). Finally, excluding 20% of manoeuvres of uncertain quality improved prediction accuracy (90[WJ1] %) and sensitivity (99%) but unchanged specificity.

Conclusion: Our model combines the visual experience of skilled technicians and ATS/ERS guidelines in automating spirometry acceptability based on entire manoeuvre. This can help in standardizing spirometry quality in different settings like clinical studies, laboratory testing and primary care practice.