Chronic stress has been identified as a critical factor that influences people’s physical and mental well-being [1,2]. However, the effect of a stressor on an individual’s well-being also depends on his or her coping mechanism [3]. In addition to stress management group interventions and self-help literature, the use of stress management apps makes it possible nowadays to learn a broad range of problem-focused and emotion-focused coping methods [4-6]. Moreover, relaxation methods are also commonly integrated in apps for the management of chronic pain [7,8], chronic diseases [9], and anxiety [10]. Through interactive design and gamification, such apps can potentially increase the users’ motivation [11,12]. This could reduce the economic burden for the health care system [13]. There are first indications that some stress management apps are indeed effective [14,15], underpinning the usefulness of this prevention approach.

Deep and slow diaphragmatic breathing can lead to a state of relaxation. Therefore, it is frequently taught as a basic strategy for the management of stress, anxiety, posttraumatic stress disorder [16], and pain [17]. Traditionally, breathing trainings are guided by health professionals, but the increasing importance for technology-driven approaches such as health apps can be attributed to financial reasons [13]. A broad range of apps specially designed for breathing trainings are available, but breathing exercises are also regularly incorporated in stress and anxiety management apps [5,10,18]. Most of these apps simply pace a distinctive breathing pattern, using audio or visual instructions. Regarding the effectiveness of these instructions, it has been shown that a wave-based visualization of the desired breathing pattern can be more easily followed compared with a circle-based visualization or a traditional audio instruction [19].

Besides pacing, providing biofeedback is another approach for breathing trainings (eg, [20-24]). With biofeedback, information from 1 or multiple sensors is used to gain greater awareness of physiological functions. Besides breathing rate, current stress management apps also target skin conductance [25] and heart rate [26,27]. Most mobile biofeedback solutions, however, require additional costly devices with integrated sensors (eg, a belt [28], wearable textile sensors [21,29], or clothing-adhered biosensors [30]).

In this study, biofeedback refers to feedback about the movement of the abdomen during a breathing task. One example for this kind of biofeedback is the BellyBio Interactive Breathing app for iOS devices by RelaxLine. It uses the mobile phone’s built-in accelerometers to capture the abdominal breathing movements. For deep and slow breathing, the sound of the ocean is transformed to relaxing music. However, so far, no study has analyzed the effectiveness of such abdominal breathing feedback. Moreover, the app is recommended only for people who are already familiar with breathing exercises. Direct instructions should be used for novices instead [19].

The Breathing-Mentor app is a biofeedback breathing app that was developed to provide such direct instructions. It combines the effective wave-base visualization of the desired breathing pattern [19] with biofeedback on the actual breathing behavior, using the mobile phone’s accelerometers. This approach allows comparing the desired breathing pattern with the actual breathing behavior in real time.

To investigate the feasibility and usefulness of the additional biofeedback, a control version without biofeedback, that is, with visual instruction only, was implemented as well. For this purpose, we conducted a user study to reveal how people who are unfamiliar with breathing exercises deal with both versions of Breathing-Mentor. The focus of this study was to determine the users’ ability to follow the breathing instructions and their subjective usage experience.

The overall signal processing approach for transforming the accelerometer measurements into the breathing signal as visualized on the screen and used for data analysis is detailed below.

Accelerometers measure 3D linear acceleration, a combination of body acceleration and acceleration resulting from gravity, in the local sensor coordinate system. As the participants are stationary during breathing training, the acceleration resulting from gravity constitutes the major portion of the measurement. Moreover, in the training target pose, this component provides information about lateral and anterior and posterior tilt of the mobile phone with regard to the sagittal and transversal body plane, respectively (see Figure 2). The basic idea is that—with the frame including the mobile phone being placed on the upper abdomen—abdominal breathing results not only in small up and down movements of the mobile phone but also in a change of the mobile phone’s orientation, where the tilt change relative to the transversal body plane is dominant. This again results in acceleration measurements with major dispersion direction approximately in the sagittal plane.

These assumptions were confirmed in pretests (1 min, 3 trials) with 5 persons already trained in abdominal breathing. For these trials, the frame including the mobile phone was positioned by the investigator (instructed by the algorithm developer) with its base above the center of the upper abdomen, so that the mobile phone’s long edge was approximately leveled, and the display was facing the person. The recorded accelerometer data from these trials were then used to obtain the major dispersion direction as the first principal component. A reference range (representing deep abdominal breathing) was extracted by projecting all recorded accelerometer vectors onto this principal component and calculating the average minimum and maximum values over the test persons.

The average accelerometer vector was also used as reference vector for aiding a repeatable positioning of the custom frame on the participants of the study and thus improving the validity and reliability of the extracted breathing signal. For this, the app provided an interactive calibration procedure with traffic light feedback on the angle deviations between the currently measured accelerometer vector and the reference vector in the xy-plane and in the xz-plane (green: <5°, orange: <15°, red: otherwise; see Figure 2). The angle deviations were calculated using scalar products between the respective vectors. This is based on the assumption that the mobile phone is kept rather stationary during the procedure, and therefore, the accelerometer measures mainly acceleration resulting from gravity, as mentioned above. The angle deviations in the xy-plane and in the xz-plane can be controlled by slowly moving the custom frame including the mobile phone on the upper abdomen laterally or forward and backward, respectively. For a successful alignment, the deviation was required to be in the green area (below 5°) in both planes for 5 seconds. The breathing signal was then obtained from the live accelerometer measurements by applying an infinite impulse response filter (resistor-capacitor low-pass filter) with cut-off frequency of 0.5 Hz, projecting the filtered measurements on the major dispersion direction again using the scalar product and scaling the result so that the reference range mapped to (−1, 1) according to the target sine wave.

To investigate the feasibility and usefulness of the additional biofeedback from the user perspective, we conducted a user study to reveal how people who are unfamiliar with breathing exercises deal with Breathing-Mentor compared with those using the control version of the app without additional biofeedback.

The screening for previous experience with breathing exercises for relaxation revealed no systematic differences between groups (see Table 2 for details). Although most participants were aware that there is a difference between abdominal and thoracic breathing and that the former can be used for relaxing purposes, only few participants actually practiced breathing and meditation exercises in their daily lives.

For the baseline block, there was a main effect of frequency bands with more power for frequency bands near the normal breathing rate (0.2 Hz; see Figure 3 for details). There were no main effect of group and no interaction between group and frequency bands (see Table 3 for details). Summarizing, the baseline measurements did not reveal any systematic group differences for objective slow abdominal breathing behavior. The subjective ratings on how easy the instruction was to follow did not reveal group differences either (t₃₇=0.56, P=.58; EG: mean 4.55, SD 0.69; CG: mean 4.68, SD 0.82).

In both groups, there was an increase of the spectral power in the recommended frequency (CG: t₁₈=3.47, P=.003; EG: t₁₉=6.12, P<.001) as well as in the SNR (CG: t₁₈=5.88, P<.001; EG: t₁₉=4.16, P=.001) in the training block compared with the baseline block (see Table 4 for means and standard deviations).

To reveal changes over time, 5 time blocks of 1 min each were included as repeated measurements variable to investigate group differences in breathing performance.

For the spectral power in the recommended frequency, the analysis of variance revealed neither significant main effects nor an interaction between group and time (see Table 5 for details).

For the SNR, there was a main effect of time with decreased SNR during the first minute compared with the second minute (CG: t₁₈=−2.48, P=.02; EG: t₁₉=−2.36, P=.03). There was a main effect of group but no interaction between group and time (see Table 6 and Figure 4 for details).

Comparisons for the subjective ratings of the 2 versions of the app are provided in Table 7. There was an overall trend in favor of the app without biofeedback.

To test whether a single 5-min training session is already enough to evoke changes in the abdominal breathing patterns toward the requested breathing frequency (0.1 Hz), we compared the baseline and the postmeasurement block with regard to the spectral power of this frequency in both groups. There were no main effect of measurement block, no effect of group, and no interaction between block and group (see Table 8 for details).

Comparable results were found for the SNR. There was no main effect of measurement block, no main effect of group, and no interaction between group and block (see Table 9 for details, see also Table 4 for mean and SD).

The Breathing-Mentor app combines effective visualization of the instruction [19] with biofeedback on deep abdominal breathing, based on the mobile phone’s accelerometers. We conducted a first pilot study with 2 versions of the app to receive the user’s feedback and investigate breathing behavior during the initial 5 min of contact. To reveal possible effects of the biofeedback, both versions contained the same visual instruction, but only the full version included additional biofeedback.

The baseline block revealed that both groups were comparable before the breathing training regarding their ability to breathe deeply and slowly with the abdomen. Breathing frequencies near the normal breathing frequency (0.2 Hz) were more prominent in both groups compared with slower frequencies. This shows that participants were rather novices for slow abdominal breathing exercises. This finding agrees with the results from the questionnaire on previous experience with breathing exercises for relaxation purposes. Although most participants were aware that abdominal breathing can be used for relaxation exercises, only few participants actually reported practicing such exercises. Thus, the participants were representative of users who could benefit from a training app for diaphragmatic breathing [19].

Indeed, both versions of Breathing-Mentor (visual instruction only and visual instruction with additional biofeedback) enabled the users to realize the requested breathing frequency of 0.1 Hz more accurately compared with the baseline, as reflected by the spectral power and the SNR. This was expected, as both conditions include the wave-based visual instruction, which has already been shown to be very effective for mobile breathing training [19]. Moreover, SNR increased in both groups after the first minute and remained at a stable level. This fast adaptation of the breathing pattern toward the instructed frequency supports the effectivity of the user interface [19] and goes in line with the high subjective ratings of the ease of use (see Table 7 for details). There is, however, no further improvement within the 5-min training block. Moreover, the postmeasurement block revealed that breathing performance returns to the baseline performance in both groups, when the visual instructions are removed again. Both findings show that the 5-min training block is not enough to trigger transfer learning. Participants remain dependent on the app during the breathing exercise. However, the protection of the users’ autonomy has been identified as an important factor in a recent stress management app [11]. Therefore, additional blocks with terminal feedback without the visual instruction might be 1 possibility to counteract dependency upon the interface and to trigger transfer learning [31].

The main research question of this study was, how additional biofeedback in a mobile app, as implemented in Breathing-Mentor (see the Methods section for details), influences people’s ability to follow the visual breathing instruction and their subjective usage experience.

Although the spectral power of the desired frequency band did not result in significant group differences, the SNR was higher for the biofeedback training group (see the Results section for details). This means that abdominal breathing at the desired frequency was not more prominent compared with the CG without biofeedback, but the occurrence of undesired frequency bands was reduced for the biofeedback group, resulting in enhanced SNR values. These findings support the effectiveness of the additional biofeedback on breathing behavior.

This benefit in performance was, however, combined with lower subjective ratings regarding the effectiveness of the app to reduce stress and ease with which app instructions could be followed for the biofeedback training. This result could be a consequence of increased cognitive workload and attention resources that are required to interpret and modulate the biofeedback graph [32]. Nevertheless, ratings for ease of use and task difficulty were high in both groups. This suggests that workload was not excessive during the training. The physiological stress level and cognitive processing during the training should be addressed more deeply in future studies, as they are expected to change with proficiency level. The role of the relaxation level could be addressed by including additional objective psychophysiological parameters [25-27] to complete the subjective ratings. Cognitive measures could also be targeted with psychophysiological parameters from electroencephalography [33] or eye tracking [34].

To summarize, Breathing-Mentor seems to be a useful tool to teach specific abdominal breathing patterns. An immediate improvement of the user’s relaxation state should, however, not be expected, especially for persons who are inexperienced with breathing tasks. With further experience, tools such as the BellyBio Interactive Breathing app might be more useful, as the auditory feedback allows to close the eyes and to focus more intensively on the body, which are facilitating factors for deep relaxation [35]. Such auditory tools might also be useful for people with age-related visual impairments. A multimodal approach could be considered to extend the app to older people.

Finally, the frame that is used to hold the mobile phone at a stable position is 1 limitation factor. Although there were no user complaints concerning the usability of this approach, the correct positioning of the mobile phone was guaranteed by the calibrating procedure and the principal investigator in this study. Other fixing solutions should be considered for everyday use.

In summary, it should be noted that participants were rapidly able to adjust their breathing pattern to the instruction (within 1 min). This result supports the feasibility and usefulness of biofeedback in mobile breathing apps based on the mobile phone’s accelerometers, especially for people who are unfamiliar with breathing techniques. Immediate effects on the user’s relaxation state should, however, not be expected.