Cough recognition in pneumoconiosis patients based on a flexible patch with an embedded ACC sensor for remote monitoring

Abstract

Background

Many respiratory diseases such as pneumoconiosis require to close monitor the symptoms such as abnormal respiration and cough. This study introduces an automated, nonintrusive method for detecting cough events in clinical settings using a flexible chest patch with tri-axial acceleration sensors.

Methods

Twenty-five young healthy persons (hereinafter referred to as healthy adults) and twenty-five clinically diagnosed pneumoconiosis patients (hereinafter referred to as patients) participated in the experiment by wearing a flexible chest patch with an embedded ACC sensor. The top 56% of the highest scoring features were then combined using several feature selection algorithms to perform the cough classification task. The multicriteria decision making (MCDM) method was used to select the classifier with the highest scores.

Results

The optimized classifier proposed in this paper achieved an accuracy of 87.1%, precision of 95%, recall of 79.1%, F1 score of 86.4%, and AUC of 95.4% for recognizing coughs in healthy adults; an accuracy of 96.1%, precision of 95%, recall of 97.4%, F1 score of 96.2%, and AUC of 98.7% for recognizing coughs in patients; and an overall accuracy of 92% for distinguishing coughs in the combined group of healthy adults and patients.

Conclusions

Our study demonstrated the effectiveness of an automated cough recognition system in both pneumoconiosis patients and healthy adults. This approach facilitates daily remote monitoring of cough occurrence in individuals with pneumoconiosis, potentially enhancing the ability of physicians to evaluate clinical status.

Background

Cough is a prominent symptom and the major reason for patients with respiratory diseases to seek medical care [1]. Additionally, certain digestive conditions, such as gastroesophageal reflux disease, may present with coughing as a symptom. The etiology of cough is multifaceted and extensive, particularly in cases of chronic cough without detectable anomalies on chest imaging, and accurate clinical diagnosis is the only way to correctly administer treatment [2]. Physicians frequently encounter inaccuracies in patients’ recall of the frequency and intensity of coughing. The timing, frequency, and severity of a patient’s cough are crucial not only for accurately assessing disease progression but also for guiding the development of subsequent diagnostic and therapeutic plans [3]. A 20–30% reduction in cough frequency is considered clinically significant in the context of refractory chronic cough (RCC) [4]. While pneumoconiosis cough may not be distinctive, it often complicates with other respiratory illnesses such as pneumonia, tuberculosis, and various pathogenic microbial infections. These comorbidities can lead to alterations in the cough’s characteristics, frequency, and intensity. Therefore, it is of clinical benefit to continuously remote-monitor cough occurrences in patients with pneumoconiosis if there is such a tool provided.

Pneumoconiosis is a collective term for occupational lung diseases caused by diffuse fibrosis of lung tissue due to long-term inhalation of different types of pathogenic dusts and their retention in the lungs during work. Symptoms predominantly include respiratory-related cough, sputum production, chest discomfort, shortness of breath, wheezing, hemoptysis, and systemic manifestations [5]. Beyond the imperative of industrial safety measures, health assessments and epidemiological tests are crucial. Coughing and sputum production are the most prevalent clinical manifestations of pneumoconiosis [6]. A study of 1,709 workers revealed a correlation between the incidence of pneumoconiosis and the level of dust exposure in the workplace, as well as a link between cough frequency and dust concentration [7].

Extensive research has focused on the analysis of cough-related audio signals. Mel-scale frequency cepstral coefficient (MFCC) has been validated as one of the most superior features [8]. In the early diagnosis of COVID-19, MFCC features extracted from cough audio samples via a deep convolutional neural network have achieved a recognition accuracy of 82.7% [9]. Several studies have employed audio and motion data to develop a CNN-based model for detecting cough episodes in the field, with a precision of 82% and a recall rate of 55% [10].

The use of acceleration signals for cough recognition has been widely studied. In certain studies, accelerometers have been ranked second only to audio microphones in importance and have been widely used in sensors such as electrocardiograms, thermistors, contact microphones, and chest straps [11]. Doddabasappla et al. [12] utilized the spectrum of acceleration signals collected by smartphones combined with machine learning to improve the accuracy of low- and medium-intensity cough detection to between 95.2% and 98.2%, respectively. Previous studies using acceleration signals for cough recognition have not employed acceleration signals alone as a dataset; instead, they have combined acceleration signals with other sensor signals, which enhances the accuracy of cough recognition to some extent but is less convenient for patients to wear daily.

Therefore, this study collected tri-axial chest acceleration signals from both healthy adults and patients using a wearable, flexible chest patch that is suitable for remote monitoring. Traditional classifiers were initially employed for cough recognition, followed by the selection of optimal classifiers using the MCDM method for reclassification.

Methods

Participants

In the pretest phase, a 90% cough recognition accuracy was obtained by selecting the healthy subjects as test subjects and enrolling 10 males and 5 females as subjects. The experimental design phase of this study employed hypothesis testing. Based on a pretest cough recognition accuracy of 90%, it was assumed that the ECG patch achieved 90% cough recognition in pneumoconiosis patients with an α of 0.05 and an efficacy of 80% (β of 0.2, p of 0.9, p0 of 0.72). The sample size required for the test was calculated based on the reference of [13], and at a 10% rejection rate, a total sample size of 25 was needed.

Twenty-five pneumoconiosis patients were recruited from the Shanghai Pulmonary Hospital affiliated with Tongji University, and 25 healthy young adults were recruited from Hangzhou Dianzi University. The inclusion criteria for healthy young adults were as follows: no history of smoking and no history of chronic or acute respiratory diseases (e.g., asthma, bronchitis, pneumonia). This study was conducted in accordance with the guidelines of the Helsinki Declaration of the World Medical Association (2000) and was approved and supervised by the Ethics Committee of Shanghai Pulmonary Hospital (approval no. L23-226). After receiving a detailed description of the experiment, all participants signed informed consent forms. The patients’ demographic information is listed in Table 1. Pneumoconiosis was staged based on satisfactory technical quality X-ray chest radiographs as well as standard films for the diagnosis of pneumoconiosis. Nineteen of the patients had stage I disease, 1 had stage II disease, and 5 had stage III disease.

Table 1 Demographic information of the participants

Devices

The experimental data were collected using a flexible chest patch that integrates a triaxial accelerometer and a single-lead electrocardiogram (ECG) provided by Vivalink, Inc., California, USA. During the experiment, the ECG patch was worn on the patient’s upper and middle left chest using disposable medical backing. The ECG patch has an embedded battery that can support continuous function for as long as three weeks, which makes it suitable for the remote monitoring of vital signs. The data collected by the patch were transferred to a matching smartphone through the Bluetooth tool and automatically uploaded to the Cloud platform for easy downloading and analysis after the experiment. The acceleration signal was sampled at 125 Hz.

In the healthy adult group, the subject was required to take mocking coughs with low, medium or high intensities. A noise meter (SmartSensor, China) was used for the measurement of the intensity and classification. According to the noise meter, the decibel ranges of low-intensity cough, medium-intensity cough and high-intensity cough in healthy people are 30–50 dB, 50–70 dB and 70–90 dB, respectively. In the pneumoconiosis patient group, a vital signs monitor (NT1D, China) together with a chest patch was used to monitor the respiratory rate and heart rate for 10 min to determine whether the patient’s condition was suitable for the experiment. Since audio signals are considered to be the gold standard for cough recognition, the ASR software on the SAMSUNG mobile phone was used for recording the audio information in the experiment. The different activities were identified by retrieving raw data from the ASR software together with the record timing, and the identification was done by a trained individual. During the experimental phase, when the patient was engaged in spontaneous coughing, the key mapper software on another SAMSUNG mobile phone was used to record the start time and the end time of each coughing session of the patient. The audio was recorded by the observer who gives command. Although the onset and termination time of each cough may not be accurate, the occurrence of each cough event is authentically recorded. Figure 1 shows a schematic of the experimental setup with all the devices used in the experiment, as well as the location where the ECG patch was pasted on the subject.

Fig. 1

The experimental system for cough recognition by a flexible patch with an embedded tri-axial ACC sensor

Experimental procedure

The experiment was divided into two stages. Healthy adults participated in the first phase of the experiment, and the behaviors collected were breathing, deep breathing, rapid breathing change, low-intensity coughing, medium-intensity coughing, high-intensity coughing, silence and coughing, whispering, normal talking, loud talking, yawning and sighing. The experiment was performed in an office environment. Each volunteer was required to complete two rounds of collection in the resting state (sitting in a chair) and one round of collection in the moving state (slow walking).

In the second phase, the participants were patients with pneumoconiosis, and the behaviors collected were breathing, deep breathing, shortness of breath, talking, yawning, sighing and coughing. The experiment was performed in the ward of Shanghai Pulmonary Hospital, and each patient was required to complete two rounds of collection in a stationary state (sitting on a chair) and one round of collection in a moving state (slow walking), and each round required the patient to complete 12 behaviors.

An acquisition time of 2 min was adopted for each behavior. When the subject was given a command for a specific behavior, he (she) needed to complete it in 5 s and then rested for 5 s before giving the next command. Figure 2 shows the entire experimental flow. During the cough collection process for patients, five patients failed to adhere to the prescribed protocol, exhibiting actions such as yawning, sneezing, or failing to respond. As a result, 15 rounds of data were excluded from the analysis. Additionally, there were 7 instances of spontaneous coughing during the collection process, resulting in a total of 367 coughs recorded for the patient group. In the cough collection process for healthy group, three participants performed non-protocol actions such as sneezing or failing to respond during the experiment. Consequently, 3 rounds of data were excluded, leading to a total of 288 coughs recorded for the young healthy group.

Fig. 2

Flowchart of experimental data collection from participants. (a): The healthy adult group; (b): The pneumoconiosis patient group

When the subject coughs, the tri-axial accelerometer can measure displacement along the X, Y, and Z axes, resulting in a distinct waveform for each direction. Figure 3 presents typical cough waveforms for healthy adults and patients with pneumoconiosis. Specifically, Fig. 3(a) and (b) show the waveforms of low-intensity and high-intensity coughs in a healthy adult, respectively, revealing the distinct fluctuation ranges of intensity. Figure 3(c) and (d) present the waveforms of voluntary and spontaneous coughs in a patient, respectively, revealing a significant difference in fluctuation intensity.

Fig. 3

Triaxial acceleration plots of the cough behavior of typical volunteers. (a) & (b): a healthy adult; (c) & (d): a patient

Data filtration and processing

Figure 4 illustrates the structural framework of the data processing. Data filtration was applied to ensure that the recorded acceleration signals matched the recorded audio signals on the ARS software. The verification was made by comparison with concurrency of timing during the experiment.

Fig. 4

Flowchart of data processing and cough classification

Feature selection

Fifty distinct features were extracted from the ACC signals to construct a dataset. Kurtosis is useful as it indicates the prevalence of higher amplitudes. The centre of gravity frequency represents the average spectral position of the signal. The average amplitude provides a measure of the signal’s overall magnitude. The extreme difference expressing the breadth of the data distribution. Sample entropy reflects the uniformity and complexity of the data’s distribution in multi-dimensional space. We also separately calculated the variance and standard deviation along the X, Y, and Z axes of the ACC data, covering features of data variability. Feature selection was performed using three algorithms: the Kruskal‒Wallis, XGBoost, and Random forest algorithms. The features were ranked based on the metrics derived from each algorithm. Twenty-eight of the original 50 features were selected and ranked by their composite scores, By ranking the 50 features using these three methods, we ultimately selected 28 features(as depicted in Fig. 5), which provided the most information about discriminating coughs. Effective features were then introduced into various classifiers for classification tasks.

Fig. 5

Selected features and their composite scores

Classification and evaluation metrics

K-fold cross-validation was employed for all recognition tasks. K-fold cross-validation can effectively utilize limited data and make the evaluation results as close as possible to the model’s performance on the dataset, reducing the impact of data partitioning on model evaluation. The final selected model is the one with the smallest average error from k times of modeling, which helps to mitigate the possibility of final recognition results being attributable to chance rather than being relatively stable. In this study, we chose 10-fold cross-validation to evaluate the classification models. The metric values were normalized to expedite computation. Additionally, normalization confines the preprocessed data within a defined range, thereby mitigating the adverse effects arising from outlier data.

Classifier scoring and ranking

The method of MCDM was employed to select the best model, which takes into account several evaluation criteria that have a significant impact on the classifier, as well as those with less impact. A decision matrix was constructed from the metrics, and two MCDM methods, namely, the technique for order preference by similarity to ideal solution (TOPSIS) and VIseKriterijumska Optimizacija I Kompromisno Resenje (VIKOR), were employed to score and rank the classifiers. Finally, the classification task was executed using the features selected through feature selection, and the top-ranked classifier was identified by the MCDM method. Figure 6 presents the performance evaluation criteria for the classifier and the MCDM method employed in the proposed approach. The correctness of the three subcriteria (accuracy, precision, F1 score) verifies the accuracy of the classification model. The robustness metric (recall) measures the classifier’s ability to accurately identify positive samples. Subsequently, a high detection rate and accuracy were achieved by applying the MCDM method with appropriate weighting of each performance evaluation metric. To determine these weights, we referred to the methodology in [14] and determined the following weights: accuracy, 0.12; precision, 0.32; recall, 0.32; F1 score, 0.12; and AUC, 0.12.

Fig. 6

Performance evaluation criteria of the classifier and the MCDM method used in the proposed approach

Results

In our investigation, we trained 13 different classification algorithms on three distinct datasets, comprising patients, healthy adults, and the combined group of healthy adults and patients. The algorithms implemented include random forest, extra trees, CatBoost, light GBM, gradient boosting, KNN, XGBoost, bagging, SVM, AdaBoost, logistic regression, decision tree, and naive Bayes (brief descriptions of these classifiers are given in the Supplemental Material). The accuracy, precision, recall, F1-score and AUC of the different classifiers for cough recognition in the three datasets are listed in Table 2 (P: patients; H: healthy adults; C: the combined group of healthy adults and patients). Table 2 shows that, except for the poor performance of some individual classifiers, most classifiers perform well for cough recognition in patients and healthy adults. The precisions of various classifiers (random forest, extra trees, CatBoost, KNN, gradient boosting, and AdaBoost) for cough recognition are consistent among healthy adults, all exceeding 90%. Within the patient group, random forest, extra trees, CatBoost, LightGBM, gradient boosting, KNN, XGBoost, and bagging demonstrated comparable precision levels for cough recognition, all surpassing 94%. According to the combined data from both the healthy adult and patient groups, the accuracy of cough recognition generally surpassed that of the healthy adult group but remained inferior to that of the patient group.

Table 2 The classification results with different standard classification algorithms

Following the construction of various classification models, it is necessary to select the algorithms that have achieved better results. This paper compares and ranks the constructed classifiers using the MCDM method detailed in Fig. 6. The results of the ranking are presented in Table 3. According to Table 3, the top five algorithms in terms of performance are random forest, CatBoost, bagging, gradient boosting, and KNN.

Table 3 Ranking of classification models for cough recognition using the MCDM approach

Figure 7 compares the criterion values for the top five algorithms per target variable label. As depicted in Fig. 7, the RF classifier outperformed the other classifiers in the patient group, achieving the best accuracy, precision, recall, F1 score and AUC, whereas the CatBoost classifier achieved the best indices in the healthy adult group. In the combined group of healthy adults and patients, the random forest model still had the best performance in terms of the five indices. Overall, the random forest algorithm was ranked by MCDM as the best classifier for cough recognition.

Fig. 7

Values of the evaluation criteria of the top 5 models by target variable labels. (a) Patients; (b) healthy adults; (c) the combined group of healthy adults and patients

Discussion

Principal findings

This study investigated cough recognition in clinical practice based on a flexible patch with an embedded ACC sensor suitable for remote monitoring. We developed and evaluated thirteen distinct classifier algorithms for cough recognition utilizing the MCDM method for systematic ranking. The most effective classifier algorithm was the random forest algorithm, which was applied to a dataset of pneumoconiosis patients and achieved remarkable cough recognition results, with an accuracy of 96.1%, precision of 95%, recall of 97.4%, F1 score of 96.2%, and AUC of 98.7%.

Many previous studies frequently integrate a variety of signals, including acceleration and audio, for cough detection, which can introduce complexities in patient portability and data processing. In this study, classifier models trained solely on acceleration signals obtained from a flexible chest patch exhibited superior classification outcomes. For example, Daniyal et al. [10] utilized audio and motion data for cough detection, achieving a precision of 82% and a recall of 55% in real-world settings among 16 patients with chronic lung diseases, and a precision of 83.8% with a recall of 71.7% in controlled lab environments with 13 patients with chronic lung diseases. Our model significantly outperforms these metrics. Similarly, Otoshi et al. [11] used tri-axial accelerometers and stretchable strain sensors, achieving a sensitivity of 92% and a specificity of 96% under laboratory conditions with 11 healthy adults and 10 adult cough patients. While our model matches the recall of Otoshi et al., it surpasses their accuracy. Furthermore, this study investigated the application of acceleration sensors in cough monitoring and explored the potential of flexible patches to enhance the accuracy of recognition. This approach not only streamlines the monitoring process but also holds promise for improving the precision of cough detection, contributing significantly to the management and study of respiratory conditions.

Limitations

The selected classifier of the random forest algorithm did not perform as well as CatBoost did in the healthy adults. A comprehensive review revealed that the classification efficacy for healthy adults lagged behind that of both the patients and the combined group of healthy adults and patients. This disparity may be attributed to the relatively sparse collection of cough data among healthy adults and the greater variety of mocking activities required from them, such as loud speaking and whispering, which potentially compromised the overall classification accuracy for healthy adults.

This research has limitations, including difficulties in handling low-intensity coughs due to weak chest vibrations, which may be confused with signals from activities [15]. Moreover, given the infrequent occurrence and minor clinical significance of low-intensity coughs in daily life, additional sensors may be necessary for classification support, or alternative features and classifier models more suited to this task may need to be developed. During the experiment, the subjects were only allowed to perform the predetermined behaviors. The other activities such as laughter or throat clearing were not designed in the experimental procedure, and may affect the overall accuracy of cough detection by the algorithm in real clinical practice. Additionally, the study may be limited by its sample size, with only 25 healthy adults and 25 patients with pneumoconiosis participating, necessitating larger-scale studies to validate the generalizability of the findings.

Conclusions

In this study, we used a flexible chest patch with an embedded tri-axial acceleration sensor to perform cough recognition tasks on both healthy adults and patients with pneumoconiosis using a proposed method of optimizing both the data features and classifiers. We found that the results of cough recognition tasks using the best classifier selected by the proposed method reached high accuracies (96.1% in pneumoconiosis patients, 87.1% in healthy adults, and 93.5% in the mixed group). The results suggest that flexible chest patches with the ACC and data processing methods can be used for the clinical application of cough recognition. The current study may help to provide clinicians with a convenient and objective tool for remote recording and recognizing changes in patients’ cough patterns.