Acoustic breath-phase detection using tracheal breath sounds

Abstract

Current breathing flow estimation methods use tracheal breath sounds, but one step of the process, ‘breath phase (inspiration/expiration) detection’, is done by either assuming alternating breath phases or using a second acoustic channel of lung sounds. The alternating assumption is unreliable for long recordings, non-breathing events, such as apnea, swallow or cough change the alternating nature of the phases. Using lung sounds intensity requires the addition of a secondary channel and the associated labor. Hence, an automatic and accurate method for breath-phase detection using only tracheal sounds would be of great benefit. We present a method using several breath sound parameters to differentiate between the two respiratory phases. The proposed method is novel and independent of flow level; it requires only one prior- and one post-breath sound segment to identify the phase. The proposed method was tested on data from 93 healthy individuals, without any history of pulmonary diseases breathing at 4 different flow levels. The most prominent features were from the duration, volume and shape of the sound envelope. This method has shown an accuracy of 95.6% with 95.5% sensitivity and 95.6% specificity for breath-phase identification without assuming breath-phase-alteration and/or using any other information.

Appendices

Appendix

There are a few plausible scenarios for breathing such as normal breathing in which the breath phases simply alternates but the duration of breaths may change, breathing with a deglutition apnea for occasional swallowing (in a breathing assessment), breathing with sleep apnea, or successive inspiration/expiration. The normal alternating phases is the most common pattern of breathing; however, other breathing patterns also do occur. How our developed voting method would handle phase identification, are being described by the following synthesized examples from Fig. 4a–e.

Fig. 4

Appendix images. Synthesized plausible scenarios of breathing patterns

Scenario 1 (Fig. 4a)—alternating breathing with different durations

Figure 4a below shows a schematic diagram of a flow signal for regular breathing with ‘time’ as the x-axis. If we consider the two breath phases for any given respiratory cycle, we notice certain traits that exist for each of the 5 PIx’s. Let’s focus on one inspiration phase (c) shown in Fig. 4a. The duration of expirations are normally longer than inspiration; therefore the PI_2-vote(n) < 0 indicating an inspiration. For regular breathing the PI₄, over the expiration phase yields a higher value; thus, PI_4-vote(n) < 0 indicating an inspiration. The gradient of an inspiratory phase is higher than that of an expiratory phase. Thus, PI_5-vote(n) > 0 is indicative of an inspiration.

Thus the majority of the 3VOTE method yields ‘c’ as an inspiration phase.

Scenario 2 (Fig. 4b)—non-breath event in regular breathing (ins-exp-noise-ins-exp)

In such a situation, the onset detection method will ignore the swallowing phase as it will detect it as noise. Since the noise segment will not be considered as a phase, essentially, the algorithm will see no difference between scenario 1 and scenario 2.

Scenario 3 (Fig. 4c)—non-breath event, irregular breathing (-ins-exp-noise-exp-ins-)

In such a case that a swallowing (or noise) occur between two expirations (a common pattern of swallowing in children), the difficulty would be to determine the phases for ‘b’ and ‘c’. Using the 3VOTE method, for phase ‘b’ = n, we will have

$$ \begin{gathered} {\text{PI}}_{12} > 0,\,{\text{PI}}_{22} \approx 0^{*} \to {\text{PI}}_{2} - {\text{vote}}\left( {\text{b}} \right) > 0 = > \,{\text{expiration}} \hfill \\ {\text{PI}}_{14} > 0,\,{\text{PI}}_{24} \approx 0^{*} \to {\text{PI}}_{4} - {\text{vote}}\left( {\text{b}} \right) > 0 = > \,{\text{expiration}} \hfill \\ {\text{PI}}_{15} < 0,\,{\text{PI}}_{ 2 5} \approx 0^{*} \to {\text{PI}}_{ 5} - {\text{vote}}\left( {\text{b}} \right) < 0 = > \,{\text{expiration}} \hfill \\ \end{gathered} $$

For phase ‘c’ = n:

$$ \begin{gathered} {\text{PI}}_{12} \approx 0^{*} ,{\text{ PI}}_{22} > 0 \to {\text{PI}}_{2} - {\text{vote}}\left( {\text{c}} \right) > 0 = > \,{\text{expiration}} \hfill \\ {\text{PI}}_{14} \approx 0^{*} ,{\text{ PI}}_{24} > 0 \to {\text{PI}}_{4} - {\text{vote}}\left( {\text{c}} \right) > 0 = > \,{\text{expiration}} \hfill \\ {\text{PI}}_{15} \approx 0^{*} ,{\text{ PI}}_{ 2 5} < 0 \to {\text{PI}}_{ 5} - {\text{vote}}\left( {\text{c}} \right) < 0 = > \,{\text{expiration}} \hfill \\ \end{gathered} $$

Therefore, the 3VOTE method still yields correct phase identification.

*Regardless of the sign of these values, they will be very close to zero, causing the PI _x-vote(n) to rely more on the alternate comparison.

Scenario 4 (Fig. 4d)—non-breath event, irregular breathing (-exp-ins-noise-ins-exp-)

Similar to scenario 3, to determine the phases for ‘b’ and ‘c’ segments, using the 3VOTE method we will have

$$ \begin{gathered} {\text{PI}}_{12} < 0,{\text{ PI}}_{22} \approx 0^{*} \to {\text{PI}}_{2} - {\text{vote}}\left( {\text{b}} \right) < 0 \, = > {\text{inspiration}} \hfill \\ {\text{PI}}_{14} < 0,{\text{ PI}}_{24} \approx 0^{*} \to {\text{PI}}_{4} - {\text{vote}}\left( {\text{b}} \right) < 0 \, = > {\text{inspiration}} \hfill \\ {\text{PI}}_{15} > 0,{\text{ PI}}_{ 2 5} \approx 0^{*} \to {\text{PI}}_{ 5} - {\text{vote}}\left( {\text{b}} \right) > 0 \, = > {\text{inspiration}} \hfill \\ \end{gathered} $$

For phase ‘c’ = n:

$$ \begin{gathered} {\text{PI}}_{12} \approx 0^{*} ,\,{\text{PI}}_{22} < 0 \to {\text{PI}}_{2} - {\text{vote}}\left( {\text{c}} \right) < 0 = > {\text{inspiration}} \hfill \\ {\text{PI}}_{14} \approx 0^{*} ,\,{\text{PI}}_{24} < 0 \to {\text{PI}}_{4} - {\text{vote}}\left( {\text{c}} \right) < 0 = > {\text{inspiration}} \hfill \\ {\text{PI}}_{15} \approx 0^{*} ,\,{\text{PI}}_{ 2 5} > 0 \to {\text{PI}}_{ 5} - {\text{vote}}\left( {\text{c}} \right) > 0 = > {\text{inspiration}} \hfill \\ \end{gathered} $$

Scenario 5 (Fig. 4e)—successive inhalation or exhalation

The above pattern is rare but it may occur especially in children. In such situations, the onset detection method overlooks the local minima, and considers ‘b’ as a single breath phase. That is because in the LV signal, the local minima will be less than half the drop from the maxima and first onset of the breath phase. Thus, the algorithm will not consider the phase complete until it reaches the next minima. In such cases, the duration (PI₂) index may indicate the wrong phase; however, the volume and gradient indices (PI₄, PI₅) will still correctly identify the breath phases. Thus, using the majority vote will still yield the correct phase detection.