Artificial Intelligence Assists in Quality Assessment of Spirometry in Clinical Trials

S. Stanojevic (Halifax (NS), Canada), P. Powell (Sheffield, United Kingdom), G. Hall (Perth, Austria), M. Topalovic (Leuven, Belgium), N. Das (Leuven, Belgium), K. Ray (Leuven, Belgium), F. Ratjen (Toronto, Canada), R. Jensen (Toronto, Canada), C. Jimenez-Ruiz (Madrid, Spain), C. Hernandez (Barcelona, Spain), F. Burgos (Barcelona, Spain), W. Janssens (Leuven, Belgium)

Background

A deep-learning based algorithm to standardize quality control review has recently been developed (Das et al., ERJ 2020). Here we externally validate this software.

Methods

Spirometry data collected in preschool children (Hospital for Sick Children, Canada) and healthy adults (Healthy Lungs for Life event; Madrid, Spain) were evaluated by ArtiQ.QC software using the 2005 ATS/ERS standards. Manoeuvre acceptability (acceptable/usable/unacceptable) by ArtiQ.QC was compared to quality
control decisions made by trained operators at the time of test.

Results

A total of 246 curves from 30 test occasions in children (healthy and cystic fibrosis) between the ages of 2.5 and 6 years of age were analyzed. Overall quality control decisions agreed for 88% (214/246) of the curves; ArtiQ.QC accepted 72% of curves compared with 67% by the operator. In a subset that was re-evaluated, there was 95% agreement with ArtiQ.QC.
A second dataset with 1231 curves from 247 test occasions in healthy adults between the ages of 18 and 60 years was also analyzed. The agreement between ArtiQ.QC and the technicians was 81%; ArtiQC accepted 58% of curves compared with 60% by the operators. Disagreement between ArtiQ.QC and technician mostly occurred
when ArtiQ.QC detected a cough, sub-maximal effort or hesitation in the curve effort.

Conclusions

ArtiQ.QC had good overall agreement with technician review for both paediatric and adult data. Using ArtiQ.QC to review spirometry curves in real-time or retrospectively can save time and evaluate data quality.