Breathing as Input Modality in a Gameful Breathing Training App: Development and Evaluation of Breeze 2 (Preprint)

Abstract

BACKGROUND Slow-paced breathing training can have beneficial effects on physiological and psychological well-being. Unfortunately, usage statistics indicate that adherence to breathing training apps is low. Recent work suggests that gameful breathing training may help to overcome this challenge. OBJECTIVE This study aims to introduce and evaluate Breeze 2, an updated version of the gameful breathing training app Breeze. It adds an improved appearance, a tutorial, the possibility to set specific training parameters, procedural generation of the visual biofeedback environment, and a novel real-time acoustic breathing detection algorithm that detects inhalation, exhalation, and non-breathing sounds (including silence). METHODS The breathing detection algorithm was developed using deep transfer learning and features an additional heuristic that prolongs detected exhalations to stabilize the algorithm's predictions. We evaluated Breeze 2 with thirty participants (14 females, Mage=29.77, SDage=7.33). Each participant performed two breathing training sessions with Breeze 2. In one session, the breathing detection used audio captured from headphones, and in the other session, it used audio from the smartphone's microphone. Participants answered questions regarding user engagement (UES-SF), perceived effectiveness (PE), perceived relaxation effectiveness (PRE), and perceived breathing detection accuracy (PDA). We used Wilcoxon-signed rank tests to compare UES-SF, PE, and PRE against neutral scores. To assess whether participants under- or overestimated the actual detection performance, we correlated PDA with the actual multi-class balanced accuracy and examined difference plots. We conducted a repeated-measures analysis of variance (ANOVA) to investigate the model's differences in balanced accuracy with and without the heuristic and when classifying data captured from headphones and smartphone microphones. To account for potential breathing sound differences between men and women, we controlled for between-subject effects of participants' gender. RESULTS Our results show scores significantly higher than neutral scores for UES-SF (W=459, P<.001), PE (W=465, P<.001), and PRE (W=358, P<.001). PDA correlated significantly with the multi-class balanced accuracy of the model (r=0.51, P<.001). Difference plots indicated that participants overestimated the model's performance when it performed poorly and underestimated it when it performed well. Furthermore, we found that the heuristic improved the model's balanced accuracy significantly (F(1,25)=10.25, P=.004) and that the model performed better on data captured from smartphone microphones than from headphones (F(1,25)=16.77, P<.001). We did not observe any significant between-subject effects of gender. The model alone reached a multi-class balanced accuracy of 74%. CONCLUSIONS Most participants perceived Breeze 2 as engaging and effective. Furthermore, the breathing detection worked well for most participants, as indicated by the perceived and true detection accuracy. In future work, we aim to use the collected breathing sounds to improve the breathing detection regarding its stability and performance. We also plan to employ Breeze 2 as an intervention tool in various studies targeting the prevention and management of NCDs.