The Effect of Measurement Trends in Belt Breathing Sensors ^†

Abstract

Sensors for respiratory monitoring can be classified into wearable and non-wearable systems. Wearable sensors can be worn in several positions, the chest being one of the most effective. In this paper, we have studied the performance of a new piezoresistive breathing sensing system to be worn on the chest with a belt. One of the main problems of belt-attached sensing systems is that they present trends in measurements due to subject movements or differences in subject constitution. These trends affect sensor performance. To mitigate them, it is possible to post-process the data to remove trends in measurements, but relevant data from the respiration signal may be lost. In this study, two different detrending methods are applied to respiration signals. After conducting an experimental study with 21 subjects who breathed in different positions with a chest piezoresistive sensor attached to a belt, detrending method 2 proved to be better at improving the quality of respiration signals.

1. Introduction

Continuous monitoring of physiological signals is frequently used by healthcare professionals, mostly with hospital patients. For instance, monitoring of respiration rate is commonly used to diagnose overall health of the patients [1], and several diseases can be detected through respiration rate, such as asthma or sleep apnea [2]. Other common applications for respiration rate monitoring are in sports to analyze the performance of athletes [3], the monitoring of the health state of drivers and commercial builders [4], and emotion recognition [5].

When designing a system for respiration rate monitoring, two different approaches are used by researchers: wearable and environmental systems, each of which has different techniques to achieve its purpose. For wearable systems, the most common technique is to measure chest diameter variations due to inhalation and exhalation of the user. This is commonly achieved by placing a flexible strap around the chest of the user. Through this strap, chest movements are transferred to the sensor of the system. Such a sensor could be a piezoresistive sensor [6], capacitive sensor [3], piezoelectric sensor [7], optical fiber sensors [8], or inertial measurement units [9], among others [10].

When measuring physiological signals through wearable devices for a relatively prolonged time, signals show a trend, which is a systematic increase or decrease in the obtained signal due to movements of the system or subject. However, linear fit can be applied offline to the registered respiration signals to mitigate this effect, as these trends may hinder data analysis [11]. A disadvantage of doing so is that relevant information from the original signal may be lost.

In this study, we show that the effectiveness of eliminating trends in signals depends on the length of the segmentation window of the measurements. To do so, different segmentation windows (periods of time analyzed by the algorithms) are proposed, and two different algorithms are used to analyze the respiration signals from 21 volunteers. Respiration signals were obtained through a chest-strap sensing system from a previous work [12].

2. Results

To analyze the respiration signals from all 21 volunteers, two different algorithms are used, whose basic principle consists of detecting all zero-crossings from the obtained signals by determining the zero-axis of such signals. Algorithm 1 measures the time difference between consecutive zero-crossings (breath cycles), whilst Algorithm 2 counts the number of crosses by zero. In this way, both algorithms are capable of calculating the respiration rate from the signal in breaths per minute (bpm). Both algorithms perform a segmentation of the signal (from 6 to 30 s); from each segmented window, the respiration rate is calculated, and then the mean from all the segments is taken as the average respiration rate. From each volunteer, a total of 30 respiration signals are analyzed (six different respiration rates set by a metronome, and five different activities) for a total of 630 respiration signals analyzed. In a previous work [12], a complete description of the algorithms and the database is included.

When detrending the respiration signals, two different methods were used. In the first method, the whole signal was detrended before calculating the respiration rate from each segment, whilst in the second method, each segmentation window was detrended individually. A depiction of both methods is shown in Figure 1. Relative error from each signal and each segmentation window was obtained from the original signals and each detrending method; then, a comparison in the increase or reduction in relative error was performed.

For detrending method 1 and Algorithm 1, the 6 s segmentation window shows an increase in relative error by 5.46%, whilst for the rest of the windows, the relative error decreases. By detrending the signals and calculating respiration rate with Algorithm 1, there is an average improvement (decrease in relative error) of 1.08%, with a maximum improvement of 2.52% (11 s window) and a maximum deterioration (increase in relative error) of 5.46% (6 s window). For Algorithm 2, only windows from 12 to 15 s show a deterioration of predicted respiration rate. There is an average improvement of 0.48%, with a maximum improvement of 1.78% (30 s window) and a maximum deterioration of 0.55% (13 s window).

When detrending original signals with method 2 for Algorithm 1, an average improvement of 1.54% is achieved, with a maximum improvement of 3.19% (11 s window) and a maximum deterioration of 3.58% (6 s window). For Algorithm 2, only windows of 6–8 s show a deterioration in the predicted respiration rate. There is an average improvement of 1.28%, a maximum improvement of 2.88% (30 s window), and a maximum deterioration of 2.35% (6 s window). A summary of these results is shown in

Table 1. Improvement in respiration signal after detrending the original signal with both methods, for each algorithm.

		Algorithm 1			Algorithm 2
	Average	MI 1	MD 2	Average	MI 1	MD 2
Method 1	1.08	2.52	5.46	0.48	1.78	0.55
Method 2	1.54	3.19	3.58	1.28	2.88	2.35

¹ Maximum improvement. ² Maximum deterioration.

. The improvement and deterioration obtained from each segmentation window are show in Figure 2 and Figure 3 for detrending method 1 and detrending method 2, respectively.

3. Conclusions

Both detrending methods show an improvement when calculating respiration rate; however, detrending method 2 shows a better performance, as the average improvement is higher for both algorithms. From detrending method 2 and Algorithm 1, only the 6 s window shows deterioration, and all other windows show improvement when calculating respiration rate. However, this improvement is not consistent when moving along the different windows. On the other hand, for Algorithm 2 it is noticed that, as the segmentation window increases, the improvement increases accordingly. Therefore, even though improvement in respiration rate prediction is moderate when detrending the signals, both detrending methods shows improvement when compared with the original signals with trends. For this specific application, detrending method 2 proves to be better.