The effects of different algorithms on the performance of a strain-based wearable device estimating respiratory rate during cycling exercise

The growing development of unobtrusive0 wearable devices measuring physiological signs is favoring breathing monitoring in applied sports settings. Respiratory frequency (fR) is attracting particular attention as a marker of physical effort and as a variable that can be easily monitored with sensors detecting chest-wall movements (e.g., strain sensors) integrated into straps or textiles. However, the algorithms used to extract fR values often do not take into account the specific needs of exercise monitoring. The purpose of this study was to investigate the performance of a commercial wearable device (i.e., Bioharness™) embedding a strain sensor measuring chest-wall movements. The performance of the Bioharness™ in estimating fR was evaluated by comparison with an airflow signal registered with a reference flowmeter. This comparison was made after having extracted fR from the raw respiratory signal using both time domain and frequency domain algorithms and by evaluating the performance of the manufacturer algorithm. Five volunteers were enrolled to perform both a ramp incremental respiratory frequency (henceforth reported as Incremental) test and a High Intensity Interval Training (henceforth reported as HIIT) test. The performance of the Bioharness™ was good when fR was extracted from the raw respiratory signal, and lower mean absolute error (MAE) values were found when using frequency domain algorithms compared to time domain algorithms. MAE and mean absolute percentage error values generally decreased with the increase in the length of the window used to estimate fR. However, the manufacturer algorithm underestimated fR in all the conditions (Incremental and HIIT), especially at fR values above 40 bpm. Hence, the algorithms estimating fR for exercise monitoring should carefully consider the specific needs of the activity recorded.