This study, performed at Kobe University Hospital, was approved by the institutional review board (Clinical and Translational Research Center) (permission number: 300034). Informed consent was obtained from all subjects included in the study. All procedures performed were in accordance with the ethical standards of the institutional and national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
In this study, we used two kinds of devices that do not interfere with each other. One device is a sensitive triaxial accelerometer (WHS-3 sensorR, UNION TOOL CO., Tokyo, Japan). This accelerometer can record acceleration signals in three orthogonal directions from the area at which it is attached every 31.25 milliseconds (Fig. 1A). In this study, to detect coughs, it was attached to the epigastric region. The other device is a wearable stretchable strain sensor (C-STRETCHR, Bando Chemical Industries Ltd., Hyogo, Japan) (Fig. 1B-E). Its mechanism was described well in another study using this sensor . In brief, the detection area of this sensor can extend to almost double its size, and the capacitance of this sensor is linearly related to the strain of the sensing area. This sensor can sensitively detect the expansion of the skin. In this study, the participants wore the stretchable strain sensor around their neck. Both the accelerometer and strain sensor are small, light, and suitable for ambulatory use. These devices were worn simultaneously (Fig. 1F). Data from these devices were transferred wirelessly to tablets or personal computers. We confirmed that no coughs were induced by wearing these devices for 30 minutes among 4 healthy volunteers (2 males and 2 females).
Cough frequency measurements
In this study, from September 2019 to June 2020, 11 healthy adult volunteers with no symptoms of cough and 10 adult patients who had symptoms of cough were consecutively enrolled. For the cough frequency measurements, the participants were equipped with two devices (a triaxial accelerometer and a stretchable strain sensor) for 30 minutes, while they sat on a chair in a room. While sitting, they were allowed to talk and move their body. The healthy volunteers were asked to cough voluntarily. For the entire duration of measurement, a researcher observed each participant from the same room and manually recorded and counted the coughs. The data obtained from the healthy volunteers and patients with cough are shown in Supplementary file 1 and file 2, respectively, and the data corresponding to the coughs are marked in yellow in those files.
Waveforms by cough monitoring system
When the subjects coughed, the two different devices (the triaxial accelerometer and stretchable strain sensor) displayed specific waveforms. Typical cough waveforms are shown in Fig. 2. Cough intensity (large cough or small cough) is represented by the wave height (Fig. 3 A, B). The cough waveforms were distinguishable from those produced by speaking and laughing (Fig. 3 C, D). The cough waveforms were also distinguishable from those produced by upper body movements (Fig. 3 E, F).
Cough frequency monitoring algorithm
For the development of the automatic cough frequency monitoring algorithm, the participants’ measurement data from the triaxial accelerometer and stretchable strain sensor were divided into consecutive small “units” lasting 5 seconds each. We defined “cough units” as those corresponding to when a subject coughed within the 5-second period. We defined “non-cough units” as those corresponding to when a subject did not cough within the 5 seconds. Whether each unit corresponded to a “cough unit” or “non-cough unit” was determined by the observer who manually counted the cough records. These “labels” were used for the machine learning algorithm.
A variational autoencoder (VAE), which is a machine learning algorithm using deep learning, was used for cough feature extraction. As shown in Fig. 4 A, VAE consists of a network called an encoder and decoder. The encoder compresses the input data unit into a latent variable space, and the decoder restores the input data from the latent variable space. In other words, the VAE can automatically extract and learn multilevel features of coughs in the latent variable space. To determine whether the input data units were “cough units” or “non-cough units” from the latent variables, a k-means clustering algorithm was used. Fig. 4 B shows an example of the clustering results.
To train and evaluate the VAE and k-means, all measured units were divided into training and test datasets (Fig. 4 A). First, signal amplitude thresholds were determined to select the units that may have coughs from all the units from all participants (n=21). The thresholds were set using the datasets labeled as “cough units”. Next, 60% of the units with an amplitude greater than the threshold were included in the training dataset, and the remaining 40% of the units were used as the test dataset. The VAE built a feature extraction network using the training dataset and was clustered by the k-means algorithm. The performance of the learned network and clustering results were evaluated by the test dataset (Fig. 4 A).
Data presentation and analysis
Analyses were carried out using JMP 9.0.2 statistical software (SAS Institute Inc., NC, USA). The sensitivity (the percentage of cough units that were correctly identified by our algorithm) and specificity (the percentage of non-cough units that were correctly identified by our algorithm) were calculated among the healthy volunteers (n=11), the patients with cough (n=10), and all the participants (n=21). The sensitivity and specificity of using only an accelerometer were also calculated. Additionally, the effects of exercise (while wearing the cough monitor, one subject was asked to walk or repeatedly stand and sit) on the results of our algorithm were examined. Finally, we analyzed the generalizability of our system. A schema of the analyses conducted in this study is shown in Fig. 5. The data for the continuous variables are summarized using means (standard deviation).