Deep Learning Integration for Normal Breathing Classification Using a Flexible Fiber Sensor

Abstract

Measuring respiratory parameters is crucial for clinical decision making and detecting abnormal patterns for disease prevention. While deep learning methods are commonly used in respiratory analysis, the image-based classification of abnormal breathing remains limited. This study developed a stitched sensor using silver-coated thread, optimized for the knit fabric’s course direction in a belt configuration. By applying a Continuous Wavelet Transform (CWT) and a two-dimension Convolutional Neural Network (2D-CNN), the model achieved 96% accuracy, with potential for further improvement through data expansion.

1. Introduction

In recent years, flexible strain sensors have gained considerable interest, especially in areas such as wearable electronics, intelligent fabrics [1,2], soft robotics [3], and structural health measurement [4]. Their remarkable flexibility and conductivity make them ideal for conforming to various surface shapes. Fabrics provide an optimal platform by adapting to the body’s natural shape while also offering characteristics that are like those found in sophisticated electronic devices [5,6,7,8,9,10]. Additionally, electronic textiles are highly flexible, offering a broad surface area for sensing without the risk of wires getting caught. With their blend of functionality and aesthetics, e-textile products are anticipated to experience significant growth in the future [11]. The core idea behind textile-based sensors is the use of a textile structure capable of detecting external stimuli. These sensors, along with electrodes, are constructed from conductive fibers, threads, or fabrics. The majority of these sensors are flexible and designed to avoid causing any discomfort or restriction to the wearer [12]. One of the key applications of fabrics lies in their strain-responsive properties. This enables them to function as strain sensors, translating physical deformations into electrical signals [13]. This property enables wearable sensors made from conductive threads, optical fibers, or conductive polymers to monitor vital signs, such as the respiration rate and electrocardiogram (ECG) signals [14].

Breathing is a fundamental activity essential for human survival and is widely applied as a key vital sign in rehabilitation, sports, and healthcare [15,16,17,18,19,20]. Respiratory patterns can be assessed by placing a strain gauge on the chest or abdomen [21]. Additionally, respiratory patterns can differ among individuals and during various activities [22]. The accurate measurement of the respiratory rate and the prediction of respiratory values will help to minimize alert fatigue within clinical decision support systems [23]. Although various methods exist for measuring respiration, many of them involve sensors that can cause discomfort or restrict movement. For instance, devices like nasal thermistors require sensors to be placed on the nose and mouth, making them impractical for everyday use. To address these inconveniences, various measurement methods utilizing fabric-based sensors are being developed. Among these is a capacitive respiratory sensor designed in the form of a belt made from conductive fabric [24,25]. This type of sensor is comfortable to wear, allowing for practical use in daily life.

Currently, there is extensive research on analyzing bio signals using deep learning methods, with many studies focusing on respiration analysis. In particular, various studies have utilized respiratory sounds to detect normal and abnormal breathing patterns as a means of disease prevention [26,27]. However, the research on classifying normal breathing by analyzing respiratory patterns remains limited, and studies that have converted signal data into image data are especially rare. Deep learning approaches, such as Multi-Layer Perceptron (MLP) [28] CNN, Long Short-Term Memory (LSTM) [29], and Recurrent Neural Network (RNN) [30] are commonly used, with CNNs being widely applied in signal processing tasks that involve image segmentation [31]. Also, frequency estimation techniques are vital for accurately estimating the respiratory rate (RR) [32]. Various methods can be employed for frequency estimation, including the Discrete Fourier Transform (DFT) [33], eigen-analysis approach [34], wavelet-based scalograms [35], and the Hilbert–Huang Transform (HHT) [36,37].

In this study, we measure respiration using a flexible stitch sensor made from conductive thread, which then detects and classifies normal breathing patterns through a CNN algorithm. The sensor is constructed using silver-coated thread and is designed as a belt with stitched flexible fabric. The respiratory measurements focus on normal breathing, typically performed in an upright posture, as well as abnormal patterns, such as breath cessation or hyperventilation. The sensor is capable of measuring capacitance values in response to strain. The collected respiratory data are organized using moving average filtering and bandpass techniques to facilitate the data analysis. Subsequently, a wavelet transformation is applied to convert the data into frequency images, which are then segmented into images at a rate of one per second. This process enables the CNN to learn and classify normal and abnormal breathing patterns effectively.

2. Materials and Methods

2.1. Fabrication of the Stitched Sensor

Flexible embroidered electrodes were fabricated using silver-coated conductive yarn (AMANN, Houston, TX, USA) as the primary material. The stitching was performed with a Brother Innov-is 95 sewing machine (BROTHER, Bridgewater, NJ, USA) on an interlock knit base fabric made of 100% polyester (Seoul, Republic of Korea). To form the electrodes, two stitch patterns were applied to a single surface of the fabric with a 3 mm gap between them. Each stitch had dimensions of 2 mm in width and 6 mm in height, while the electrode length was 60 mm. Both the upper and lower threads used in the stitching were composed of silver-coated conductive yarn, providing consistent electrical conductivity across the electrode areas.

The design considered the directional properties of the knit fabric, which has wale and course directions. The course direction is known for its greater stretchability, which is advantageous for respiratory measurements. The sensors were fabricated in both directions, and based on performance, the course-aligned sensor was selected for the respiratory experiment due to its superior stretch characteristics [38].

For practical respiratory measurement, the sensor was integrated into a belt design. The belt, made from a webbing strap and a plastic buckle (both from Seoul, Republic of Korea), allowed for the sensor to be securely fastened around the waist. Only the section containing the sensor utilized the stretchy knit fabric to ensure accurate measurement during respiration. Nickel-plated nuts and bolts were used as terminals to connect the sensor to the measurement device. This connection was achieved by wrapping the conductive thread around the bolt and securing it with a nut, providing a reliable and unobtrusive connection.

The operating principle of the flexible embroidered electrode is based on capacitance, as illustrated in Figure 1. The capacitance $C$ is calculated using Equation (1), where $A$ represents the electrode area, $d$ the distance between electrodes, and $\epsilon$ the permittivity. When inhaling, air enters the lungs, expanding the thoracic cavity and causing the embroidered electrodes to stretch. This expansion increases $A$ and simultaneously reduces $d$, resulting in an increase in $C$. Conversely, during exhalation, as air exits the lungs, the thoracic cavity contracts, causing the embroidered electrodes to return to their original state. This contraction decreases $A$ and increases $d$, leading to a decrease in $C$. Since the two electrodes are located on a single surface, the effect of permittivity changes is minimal.

$$C=\epsilon {\displaystyle \frac{A}{d}}$$

(1)

The entire fabrication process, from stitching the conductive yarn to assembling the belt-type sensor, is shown in Figure 2 and Figure 3, which provides images of the finished embroidered respiratory sensor, including a schematic (a) and an actual photo (b). The detailed specifications for the stitched strain sensor, belt strap, and buckle can be found in

Table 1. Specifications of stitched sensor shown in Figure 3.

	Stitched Respiratory Sensor	Strap	Buckle
Dimension (height × width)	75 × 50 (mm)	700 × 50 (mm)	80 × 55 (mm)
Thickness	0.56 (mm)	0.43 (mm)	11.8 (mm)
Material	Polyester	polypropylene	Polyoxymethylene

.

2.2. Respiration Data Acquisition Protocol

During respiration, the thoracic cavity moves, with the ribs rising and the transverse diameter of the chest expanding during inhalation, and the ribs descending as the vertical diameter of the chest increases during exhalation [39,40]. Based on this principle, a strain sensor was placed near the xiphoid process, which is the central part of the thoracic cavity. The sensor was used to measure thoracic movement during respiration. To isolate the measurements to the sensor alone, the support structure was made from a non-elastic material. The sensor consisted of two planar stitch electrodes, whose area and the distance between them changed with the expansion and contraction of the thorax.

When the respiratory sensor is worn and the user inhales, the difference in the distance between the electrodes ($∆d$) decreases, and the difference in the electrode area ($∆A$) increases, leading to an increase in the capacitance. Conversely, when exhaling, the $∆d$ increases, and the $∆A$ decreases, resulting in a decrease in the capacitance. This process can be explained in more detail through the following Equation (2). The initial area of the electrodes ($A_0$) increases with the expansion of the thoracic transverse area during inhalation, causing the strain sensor to stretch and the electrode area ($A_f$) to increase. The difference between the initial electrode area and the increased area ($∆A$) provides the measurement value. Similarly, the distance between the electrodes also changes. The initial distance ($d_0$) between the electrodes decreases as the sensor stretches during inhalation, and the difference between the initial and decreased distances ($d_f$) is used to obtain the measurement values. These measurements are not obtained separately but rather as a single value, reflecting the combined changes in both the electrode area and the distance between the electrodes. The operating principle of the sensor, in response to thoracic movements, can be observed in Figure 4.

$$∆d=d_0-d_f,∆A=A_f-A_0$$

(2)

To analyze the obtained measurement data using deep learning, noise removal is essential. Therefore, the respiratory data were first processed using a moving average filter to eliminate the noise, with a window size set to 5. Subsequently, a bandpass filter was applied to isolate the desired frequency range between 0.2 Hz and 1 Hz. Both normal and abnormal breathing were measured for 3 min each, and one respiratory cycle was determined to be approximately 8 Hz.

A bandpass filter range of 0.2 Hz to 1 Hz was selected based on typical adult respiratory rates, which are generally around 12 to 20 breaths per minute (approximately 0.2 Hz to 0.33 Hz for normal breathing). This frequency range minimizes the influence of external factors and noise, making it advantageous for accurately capturing respiratory signals. Additionally, experiments with actual respiratory data, including both normal and abnormal breathing patterns, confirmed that this range is optimal. When frequencies outside this range were tested, noise interference affected the respiratory signal, making accurate data collection challenging.

2.3. Classification of Normal Breathing

Various methods are used for in-depth data analysis. In this study, since we were working with 1D respiratory data, it could be analyzed directly using deep learning. However, the CNN used in this study is particularly effective for image-based learning, so we explored a method to leverage this capability. A CWT represents signals in the time–frequency domain, allowing for the simultaneous observation of the frequency components and their corresponding time intervals. This time–frequency transformation is advantageous for clearly visualizing changes in specific frequencies over time, enabling the CNN to learn these features in an image format and recognize patterns in breathing. The frequency images generated from the CWT are converted into RGB channels, which provide essential information for the CNN model to learn the frequency and amplitude variations in breathing patterns. The differences between normal and abnormal breathing are reflected in the changes across the frequency bands, which the CNN can effectively classify through the CWT images. Therefore, the CWT serves as a critical step in transforming the complex characteristics of 1D signals into 2D images, enabling the CNN to achieve high accuracy in classifying breathing patterns [41]. As a result, we decided to transform the 1D respiratory data into images (2D) using a CWT analysis, which visualizes frequency changes, such as by using color maps. Both the normal and abnormal breathing data were converted into images using the CWT analysis code provided in MATLAB 2024. The resulting images were segmented into 1 s intervals to classify the characteristics of different types of breathing.

A total of 248 images (214 for training, 28 for validation, and 26 for testing) were used for the 2D-CNN model, with each image sized at 288 × 288 × 3, utilizing RGB channels. The CNN architecture consisted of two 2D convolutional layers with 3 × 3 filters, batch normalization, ReLU activation functions, max-pooling layers with a stride of 3, and dropout layers with dropout rates of 0.4 and 0.3. The first convolutional layer contained 16 filters, while the second had 32, and both used the ‘same’ padding to maintain the output dimensions. A fully connected layer with 64 neurons was used, followed by a SoftMax layer for binary classification. The key training parameters included a mini-batch size of 64, a maximum of 5 epochs, and an initial learning rate of 0.001. Validation was performed every 5 mini batches to monitor the model’s performance, with a learning rate reduction factor of 0.5 applied every 5 epochs to enhance stability. The test results, including the accuracy and a confusion matrix, were generated post-training to assess the model’s performance. The CNN architecture used in this analysis is presented in Figure 5.

The results analyze whether the characteristics of normal and abnormal breathing can be distinguished. If normal breathing is established as the correct data, any data that significantly deviate from these characteristics can be assumed to be defective and classified as having different features. Thus, we conducted experiments using the 2D-CNN algorithm to assess how accurately it can distinguish between normal breathing, which is considered defect free, and abnormal breathing, which is regarded as defective.

In a 2D convolution, the operation can be described by Equation (3), where $m$ represents the horizontal (column) direction, and $n$ represents the vertical (row) direction. The convolution operation can be mathematically formulated as follows.

$$S(i,j)=(I\ast K)(i,j)=\sum _m\hspace{0.5em}\sum _{n}I(i-m,j-n)·K(m,n)$$

(3)

$S(i,j)$ represents the value at position $i,j$ in the output feature map. $I$ denotes the input image or the values from the previous layer. $K$ refers to the filter or kernel applied during the convolution. $m$ and $n$ are indices in the horizontal and vertical directions of the filter, respectively. At each location $i,j$, the corresponding values of the input $I$ and the filter $K$ are multiplied, and the resulting products are summed to obtain the output value. By iterating this process across all the positions in an image, a new feature map is generated.

3. Results

3.1. Characterization of the Stitched Sensor

3.1.1. Stretchability and Sensitivity

To enhance the ease of measurement for respiration measurement, we conducted an experiment to investigate the directional properties of the knitted base material for the stitch electrodes. A universal testing machine (UTM) was used for the experiment, where a 50% strain was applied in both the course and wale directions of the knit, and the resulting capacitance changes were compared. Although the movement of the human chest during breathing is less than 10%, a significant variation exists between individuals due to differences in body shape. Therefore, we conducted a 50% elongation test to assess the sensor’s capability to accommodate diverse body types with a single sensor.

The results of this experiment are shown in Figure 6. In Figure 6a, the capacitance values (pF) are compared according to the different knit directions. A strain was applied starting from the initial length of the stitch electrodes, 60 mm, and increased up to 90 mm. The initial capacitance was higher in the wale direction, at 0.79 pF compared to 0.6 pF in the course direction. In the wale direction, the capacitance gradually increased from 0.79 pF to a maximum of 0.89 pF. In contrast, the course direction started at 0.6 pF and increased to a maximum of 0.75 pF. The difference in the change for the wale direction was 0.1 pF, while the course direction showed a change of 0.15 pF, indicating that the course direction exhibited a larger variation. When comparing the raw data, this difference may not appear significant. However, when analyzed in terms of the change rates ($\Delta C/C$), the wale direction shows a change rate of up to 0.14, while the course direction reaches up to 0.28. This result indicates that the course direction has notably better stretchability.

These changes are more clearly illustrated in Figure 6b, which demonstrates that the capacitance change in the course direction is greater than that in the wale direction at the same level of strain. Specifically, the change in the course direction is approximately 8, while the change in the wale direction is around 6, indicating a significant difference. In contrast, the change in the longitudinal direction is similar, showing a value of about 1. This suggests that, due to the tensile strain associated with respiration, the sensor in the course direction experiences similar changes in the $\Delta d$, but exhibits a difference in the $\Delta A$, resulting in a greater capacitance change. The experimental results for the wale and course directions are detailed in

Table 2. Descriptive statistics of stitched sensor.

Direction	Mean	Standard Deviation
Wale	0.82	0.03
Course	0.65	0.05

. In the wale direction, the mean capacitance change is 0.82 pF, with a standard deviation of 0.03. In the course direction, the mean capacitance change is 0.65 pF, with a standard deviation of 0.05. Although the standard deviation in the course direction is slightly higher, this difference is minimal. Additionally, the higher standard deviation in the course direction suggests that the sensor has greater sensitivity in this orientation. Based on these results, it is evident that a stitch sensor produced in the course direction is more suitable for measuring respiratory changes across a wide range of body types.

3.1.2. Durability

The durability test for the stitch sensor in the course direction was conducted with a 20% strain applied at a speed of 10 mm/s for a total of 1000 cycles. The results of this test can be observed in Figure 7a. Since the stitch sensor is made with silver-coated yarn, it can be produced as a high-sensitivity sensor, although a slight hysteresis does occur [42]. However, the capacitance values consistently remained within the range of 0.65 to 0.68 pF throughout the test, and no significant visible changes were observed even after 1000 cycles.

Additionally, Figure 7b shows the stability of the sensor across specific intervals by plotting the average capacitance values for selected cycles. The intervals 1–100, 500–600, and 900–1000 were chosen, with average capacitance values of 0.66, 0.67, and 0.67, respectively. These values indicate a stable capacitance response from the initial to the final cycles.

3.2. Measuring Respiration Using Stitched Sensor

We measured respiration using a belt-type respiration sensor, conducting the test under two conditions. Both normal and abnormal breathing were measured for 3 min each, and one respiratory cycle was determined to be approximately 8 Hz. Normal breathing was measured in a standing position with regular breathing patterns, while abnormal breathing involved significant body movement or varied breathing speeds and rhythms while standing. The measurements were taken by connecting the cables of an LCR meter to the terminals attached to the sensor. The frequency was set to 50 kHz during the measurements. This setup can be observed in Figure 8.

In the graphs produced from the measurements, Figure 9a shows normal breathing, while Figure 9b illustrates abnormal breathing. The differences in the speed and peak amplitude of the breathing cycles can be observed. On average, normal breathing exhibits changes between 3.8 pF and 4.1 pF, whereas abnormal breathing shows a much wider range, from 2.8 pF to 4 pF. This indicates that normal breathing follows a regular pattern with consistent breath volumes, while abnormal breathing is highly irregular and difficult to analyze due to the unpredictable breath volumes.

To analyze the measured data using deep learning, noise removal is necessary. Therefore, the recorded respiration data was first processed using a moving average filter to eliminate the noise. A window size of 5 was used for this filtering. Following this, a bandpass filter was applied to isolate and output the desired frequency range of 0.2 Hz to 1 Hz. The normal breathing data, after applying the moving average filtering and bandpass filter, can be seen in Figure 9c, while the abnormal breathing data are shown in Figure 9d.

3.3. 2D CNN

The images in Figure 9c,d were transformed into visual representations using a CWT, resulting in the completed Figure 10. These data were segmented to classify normal and abnormal breathing. The graphs illustrate the frequency over time, with a total duration of 3 min, showcasing a frequency range from 0.2 Hz to 1 Hz. Additionally, the different colors represent varying magnitudes of the frequency, with the same color appearing at intervals, characteristic of each breathing pattern. Thus, breathing can be distinguished based on this color distribution, allowing for the identification of anomalies.

In Figure 10a, a consistent breathing amplitude is observed, where the high-frequency areas (yellow) are concentrated in a straight line. Conversely, Figure 10b displays multiple lines of yellow regions due to variations in the speed and amplitude of breathing. When examining Figure 11, which divides the data into 1 s intervals, the frequency ranges of the yellow regions in graphs a and b appear differently. In the normal breathing graph (a), the yellow regions are more prominent, while in graph b, the yellow areas appear thinner. This indicates that if the yellow regions do not display a significant amplitude, it may suggest an anomaly, allowing for it to be classified as abnormal breathing.

The 1 s data segments may capture only partial information about a respiratory cycle, as a complete normal breathing cycle from actual respiratory data typically occurs over about 3 s. Additionally, when arranging the CNN input images, all the data segments were organized in a shuffled order rather than in a sequential time order. This approach enabled the CNN to learn diverse respiratory patterns without relying on a specific sequence, enhancing the model’s ability to generalize across different breathing patterns.

The images in Figure 11 were each segmented into 214 parts, where 80% of the data were used for training, 15% for validation, and the final 5% for testing. The results after the testing phase can be seen in Figure 12, where (a) represents the accuracy graph, and (b) shows the loss data. The validation accuracy was recorded at 96.43%, while the test accuracy was 96.12%. These results indicate that both the training and the data were well learned.

Figure 13 displays the test results in the form of a confusion matrix. In this matrix, 1 represents normal breathing and 0 represents abnormal breathing. Normal breathing was identified with 100% accuracy, while abnormal breathing was correctly identified at 92.3%, with 7.7% incorrectly classified as normal breathing. This outcome suggests that it is possible to classify the characteristics of normal and abnormal breathing. Additionally, the occurrence of frequency ranges in abnormal breathing that are like those in normal breathing is an expected result. In cases of abnormal breathing, there are instances where certain amplitudes or patterns resemble those of normal breathing, as the data are derived from real human respiration. Despite this, the model successfully identified over 90% of the abnormalities, demonstrating that the use of a 2D CNN for detecting normal breathing is highly effective.

4. Discussion

In this study, a stitched sensor was used to develop a respiratory sensor that measures the movements of the thorax caused by breathing. The sensor was designed to detect both normal breathing patterns and various irregular breathing anomalies, allowing for a deep learning model to identify and classify defects. Silver-coated thread was utilized in the stitching to enable high-sensitivity sensing [42]. The sensor was designed as a belt for convenient and easy wear. For the measurements, terminals were attached to the sensor, enabling respiration without the need for additional devices when connected to the equipment.

The first consideration in the sensor development was the directionality of the knit fabric on which the stitch electrode was sewn. Knit fabrics have both wale and course directions, with the course direction exhibiting greater elasticity [38]. To assess the suitability of the knit direction for respiratory measurements, the fabric was stretched by 50% in each direction, and changes in the capacitance were observed. As a result, the change in the electrode area ($\Delta A$) in the course direction was approximately 8 units greater than in the wale direction, leading to a maximum change rate ($\Delta C/C$) of 0.28. In contrast, the wale direction exhibited a $\Delta C/C$ value of 0.14, indicating that the course direction’s$\Delta C/C$ value was approximately twice as high. If a course-direction sensor shows more significant capacitance changes at a 50% stretch, this suggests that the sensor has a broader measurement range than a wale-direction sensor, for greater stretch levels. Based on these results, we decided to develop the stitch sensor in the course direction.

To reduce noise during the respiratory measurements, we applied moving average filtering and a bandpass filter to obtain noise-reduced respiratory data. While this method does not eliminate all the noise and may be influenced by external factors, these effects did not significantly impact the analysis in this study. In future work, we plan to enhance the model’s robustness by incorporating various real-world conditions.

For the respiratory analysis, we utilized a 2D CNN, one of the deep learning analysis systems, to classify the breathing patterns. Since CNNs specialize in image classification, we converted the respiratory signals into frequency images using a CWT. The frequency magnitude was represented by different colors, making the features visible in the frequency domain. Using these images, we applied the 2D CNN to classify normal and abnormal breathing patterns. While the system could not achieve a perfect classification, it was observed that less than 10% of the abnormal breathing patterns resembled normal breathing. However, this is considered a successful result, as it is natural for normal breathing patterns to occasionally appear amid irregular breathing when measured manually. Additionally, since the data used were from live breathing measurements, there could be a limitation in the quantity of the data. To overcome this limitation, future studies could improve the classification by adding data from individuals with various body types, allowing the system to better account for data variability [43].

5. Conclusions

In this study, a stitched sensor was used to develop a respiratory sensor, and deep learning was applied to classify normal and abnormal breathing patterns. The stitch sensor, designed based on the course direction of the knit fabric, was shown to be more suitable for respiratory measurements. A belt-shaped design was adopted for easy and convenient use during respiration measurement. The respiratory data were processed using a moving average filter and bandpass filter to remove the noise and smooth the data curve. These data were then transformed into images using a CWT, and a 2D CNN was applied for classification. The final test results demonstrated an accuracy of 96%, although the classification of defective abnormal breathing patterns was not entirely perfect. Nevertheless, the results are significant as they show that the deep learning algorithm is effective for classifying normal and abnormal breathing, achieving a high accuracy of 96%.

This study presents three main contributions:

1. Innovative flexible stitched sensor design: utilizing silver-coated thread and a knit structure, the sensor offers enhanced comfort and flexibility, making it ideal for wearable respiratory monitoring applications.

2. Practical belt-based sensor for real-world applications: the belt-style sensor design is optimized for long-term respiratory monitoring in daily life, providing a user-friendly and practical solution not commonly seen in previous studies.

3. High-performance respiratory pattern classification using deep learning: By converting the respiratory data into frequency images and applying a CNN, this study achieved a high accuracy at distinguishing between normal and abnormal breathing patterns. This approach uniquely combines frequency-based image transformation and deep learning for effective bio signal analysis.