DATA COLLECTION PROCEDURE
To evaluate and inform the design of SpiroSmart, we creat-ed a dataset of audio samples. In all, 52 volunteers partici-pated in a 45-minute study session (Table 1). All partici-pants self-identified themselves as having none or only mild lung conditions. Our custom data collection application for the iPhone 4S recorded subjects’ exhalation sounds using the built-in microphone (at 32 kHz) and provided feedback to the user, coaching them through the spirometry maneu-ver. We also obtained measurements during the same ses-sion using an ATS certified standard clinical spirometer, the nSpire KoKo Legend, as the “gold standard.” The KoKo is a pneumotach spirometer and was calibrated with a 3L sy-ringe before each session.
Spirometry measurements are completely effort-dependent and patients are coached through this maneuver by a trained technician. While using the clinical spirometer, participants were coached both orally and with gestures. With Spi-roSmart, participants were coached with textual prompts on the screen and only with gestures—oral prompts would have interfered with the audio recording (Figure 3).
SpiroSmart also calculated a real-time estimate of flow (us-ing LPC gain, discussed in the next section), and displayed the measure as a real-time visualization. This also provided an incentive graphic; namely, a ball displaced vertically in a cylinder proportionally to the strength of the exhalations. After the initial burst, the ball dropped slowly to the bottom of the cylinder, signifying the end of the test. Like the Ko-Ko Legend Spirometer, SpiroSmart displayed an estimated Flow vs. Volume curve at the end of the effort (Figure 1). An estimate of exhaled volume was calculated by integrat-ing estimated flow with respect to time.
The forced expiratory maneuver was explained to partici-pants and they were asked to practice using the spirometer.
Figure 4. Block diagram of SpiroSmart’s feature extraction.
Once the participants were able to perform an acceptable maneuver according to ATS criteria for reproducibility, three efforts were recorded using the spirometer [22]. The raw flow and volume measurements from the KoKo were obtained using a USB connection and custom software. Next, participants were introduced to SpiroSmart.
In our pilot study, we observed that participants uninten-tionally varied the distance at which they held the phone as well as lip posture, potentially introducing unwanted varia-bility. We therefore had participants use SpiroSmart in four configurations, in random order: with a mouthpiece (to maintain lip posture), with a sling (to maintain distance), with neither attachment; and with both attachments (Fig-ure 3). Note that it was impossible to collect data from Spi-roSmart and the KoKo Legend at the same time so explicit ground truth is unknown. Instead, each effort from Spi-roSmart was associated with one randomly selected ac-ceptable curve from the KoKo device during that same ses-sion. The signals were aligned using PEF for the KoKo and the maximum amplitude in the audio stream from Spi-roSmart as reference points. The audio stream was seg-mented automatically starting one second before and ending six seconds after the maximum audio amplitude.
We also selected 10 participants to return for two more data collection sessions (2 days up to two weeks apart), allowing us to look at the consistency of measurements from Spi-roSmart over longer periods. The participants were asked back based on specific demographics—an equal number of men and women, and equal number of normal and abnor-mal subjects. We refer to “abnormal” subjects as those with abnormally shaped curves, not necessarily reduced lung function measures. In total, we collected data from 248 clinical spirometer uses and 864 SpiroSmart uses.
Interestingly, 6 subjects were found to have abnormally shaped curves from ailments that they were unaware of and 8 of the 13 subjects who reported lung ailments produced normally shaped curves—albeit with less than expected lung function measures.