Spiro AI: Smartphone Based Pulmonary Function Testing

Abstract

Spirometry is a widely employed pulmonary function test used to benchmark lung health and assist in diagnosing chronic lung conditions such as chronic obstructive pulmonary disease and asthma. When used frequently, such as in a home or portable setting, spirometry results can predict pulmonary exacerbations or monitor the effectiveness of treatment. Unfortunately, portable options are expensive and not truly portable by modern standards. Prior work has shown it is possible to conveniently obtain spirometry metrics using the built-in microphone of a smartphone, requiring no accessories. This work proposes Spiro AI, an end to end sound-based smartphone spirometry system that includes automatic quality control and complete spirometry reporting, bringing smartphone spirometry closer to reality. Several machine learning models and deep learning architectures are thoroughly evaluated as potential components in the system. Models are trained and evaluated on thousands of patients sourced from a newly created dataset that is likely the largest audio based spirometry dataset to date. The results suggest the problem becomes increasingly difficult when the sample size scales from tens to thousands of subjects because the population is more diverse and the quality of recorded maneuvers becomes difficult to control. Nonetheless, the results suggest Spiro AI is capable of trend reporting and screening; however, in its current stage it may not be precise enough for FDA certification.