Sage Journals: Discover world-class research

Abstract

Background

Intelligent wearable devices have potential for chronic obstructive pulmonary disease (COPD) monitoring, but the effectiveness of combining cough and blowing sounds for disease assessment is unclear.

Objective

The objective was to assess COPD severity via physiological parameters and audio data collected by a smartwatch.

Methods

COPD patients underwent lung function tests, electrocardiograms, blood gas analysis, and 6-min walk tests. The patients’ peripheral arterial oxygen saturation (SpO₂), heart rate variability (HRV), heart rate (HR), and respiratory rate (RR) were continuously monitored via a smartwatch for 7–14 days, and voluntary cough and forceful blowing sounds were recorded twice daily. The HR, SpO₂, and RR were categorized into all-day, sleep, and wake periods and summarized using the mean, standard deviation, median, 25th percentile, 75th percentile and percent variation. The correlations among lung function, physiological parameters, and audio data were analyzed to develop a model for predicting COPD severity.

Results

Twenty-nine stable patients, with a mean age of 67.0 ± 5.8 years, were enrolled, and 89.7% were male. HR, HRV, RR, cough sounds, and blowing sounds were significantly correlated with the Global Initiative for Chronic Obstructive Lung Disease (GOLD) grade, with cough sounds showing the highest correlation (r = 0.7617, p < .001). Cough sounds also had the strongest correlation with the mean 6-minute walking distance (r = 0.6847, p < .001), whereas blowing sounds had the strongest correlation with the Body mass index, airflow Obstruction, Dyspnea, and Exercise capacity index (r = −0.6749, p < .001). A logistic regression model using the RR and blowing sounds as key predictors achieved accuracies of 0.77–0.89 in determining the GOLD grade, with a Cohen's kappa coefficient of 0.6757.

Conclusions

Audio data were more strongly correlated with lung function in COPD patients than were physiological parameters. A smartwatch with audio collection capabilities effectively assessed COPD severity.

Trial Registration

ClinicalTrials.gov NCT05551169

Keywords

Wearable devices telehealth COPD disease diagnosis pulmonary function

Introduction

Chronic obstructive pulmonary disease (COPD) is the most common chronic respiratory disease and is currently the third leading cause of death worldwide.^1,2 In 2019, it affected approximately 391.9 million people and accounted for over 3 million deaths worldwide.^3,4 COPD imposes a large and growing burden on society in terms of direct and indirect costs.⁵ Epidemiological studies revealed that there are nearly 100 million COPD patients in China, with prevalence rates of 13.7% in adults over 40 years old and 27.4% in people over 60 years old.^6,7 These statistics underscore the urgent need for effective COPD management strategies to alleviate the disease burden, which relies heavily on accurate disease assessment.⁸

COPD is a lung disease characterized by irreversible airflow limitation, which progressively worsens and leads to a reduced quality of life.⁸ Lung function tests are the gold standard for diagnosing and assessing airflow limitation severity. However, these tests are expensive, require trained professionals, and are not always accessible, especially in low-resource settings.^9,10 Furthermore, annual testing may fail to detect deterioration promptly. Although hand-held spirometers offer a potential solution for routine monitoring at home, their low patient compliance and operational complexity limit their practicality.^11,12

Recent technological advancements in wearable devices have made the assessment of COPD faster, simpler, and cheaper. These smart devices are capable of measuring valuable metrics, such as peripheral arterial oxygen saturation (SpO₂), the respiratory rate (RR), the heart rate (HR), and heart rate variability (HRV).¹³ Studies have shown that compared to healthy individuals, COPD patients typically exhibit higher RR and HR,¹⁴ lower SpO₂, and reduced HRV.¹⁵ Several studies have already used wearable devices for diagnosing and monitoring COPD. Some studies have utilized wearable devices to measure HRV and physical activity parameters to differentiate COPD patients from healthy individuals,^16,17 while others have used wearable devices or mobile devices to monitor cough sounds, blowing sounds, and other audio data for COPD diagnosis and lung function prediction.^18–23

However, there is relatively limited research on the use of wearable devices, particularly smartwatches, for monitoring physiological data to predict the severity of COPD.^24,25 For the collection of audio data, most current studies focus on using microphones, smartphones, or wearable sensors,²⁶ with no research using smartwatches for audio data collection. Furthermore, studies combining these two types of data for disease assessment are scarce. Therefore, we are interested in exploring whether a model based on physiological parameters and audio data measured by smartwatches can effectively assess the severity of COPD and how well it performs. We hypothesize that a model combining physiological parameters and audio data measured by smartwatches could accurately evaluate the severity of COPD.

In this study, for COPD assessment, we included lung function, the COPD Assessment Test (CAT), the modified British Medical Research Council (mMRC) score, the 6-min walk test (6MWT), and the Body mass index, airflow Obstruction, Dyspnea, and Exercise capacity (BODE) index.²⁷ Our primary goal was to evaluate the ability of a smartwatch to record cough and forceful blowing sounds and physiological parameters to assess COPD severity.

Methods

Study population

Patients with COPD at the outpatient clinic of the Department of Respiratory and Critical Care Medicine, Peking University First Hospital, were screened from June to August 2022. Eligible participants who were over 18 years of age, in a stable condition, and able to operate a mobile phone independently were included. Those experiencing acute exacerbations were eligible for enrollment 3 months after resolution. Patients with other chronic respiratory diseases, a history of lobectomy and/or lung transplantation, or pleural disease were excluded. Patients with severe comorbidities and those who could not wear a watch were also excluded. The detailed inclusion and exclusion criteria are provided in supplementary data 1 (Table S1). This is a preliminary study, with a sample size of about 30 COPD participants, based on previous studies.^24,25

All the participants provided signed informed consent forms prior to study initiation. The study was in compliance with the Declaration of Helsinki. The study protocol was approved by the ethics committee of Peking University First Hospital (Approval number: 2022083) and was registered at www.ClinicalTrails.gov (NCT05551169).

Data collection and processing

Dataset collection

Demographic characteristics, including age, sex, height, and weight, were collected. COPD was assessed via the CAT and mMRC questionnaire. The participants underwent lung function tests, the 6MWT, electrocardiography (ECG), and arterial blood gas analysis. The BODE index, which assesses the prognosis of COPD patients via body mass index (BMI), the degree of airway obstruction, dyspnea status, and exercise capacity, was calculated via an established empirical model.²⁸ According to the Global Initiative for Chronic Obstructive Lung Disease (GOLD) report criteria,²⁷ patients were categorized as having mild (forced expiratory volume in 1 s (FEV1) ≥ 80% of the predicted value, GOLD grade of 1), moderate (50%≤FEV1 < 80% predicted, GOLD grade of 2), severe (30%≤FEV1 < 50% of the predicted value, GOLD grade of 3), or very severe (FEV1 < 30% of the predicted value, GOLD grade of 4) COPD.

The participants were instructed to use a smartphone app to complete a daily COPD questionnaire that distinguished acute exacerbation from daily fluctuations. Acute exacerbation of COPD (AECOPD) is defined as a sudden worsening of respiratory symptoms beyond the daily variation range, resulting in a change in the medication regimen and manifesting as the presence of purulent sputum, an increased sputum volume, and increased dyspnea.²⁷ Exacerbation was diagnosed according to the criteria modified from Anthonisen and colleagues²⁹ and validated in several studies.^30–35 Further details of the questionnaire are available in supplementary data 2.

The participants were given a smartwatch (Watch GT3/Watch 3, Huawei, China) and a smartphone (Huawei, China) with a preinstalled data collection app. The smartwatches automatically collected photoplethysmography (PPG) and acceleration (ACC) signals to calculate the RR, SpO₂, HR, and HRV. The participants were required to wear the devices for 7‒14 consecutive days, with HRV measured every 10 min during sleep (each sampling lasted for 1 min). They were also required to wear the smartwatch for at least 10 h a day, with a minimum of 7 h worn during the night. The participants were trained to correctly use the smartwatches to record cough and forceful blowing sounds twice daily. The participants were asked to collect audio data in a quiet environment and in a resting state. During data collecting, they were instructed to place the wrist-worn smartwatch at a distance of 30 cm from their mouth. Each audio collection required the patient to actively cough 3–5 times and blow forcefully at least three times for more than 6 s each time until no air was exhaled. Further details of the audio data collection procedure are available in supplementary data 2.

Feature extraction

Digital signal processing, specifically filtering, was applied to the PPG and ACC signals to extract parameters such as the HR, RR, and SpO₂ related to respiratory diseases. The physiological parameters were categorized into sleep, wake, and all-day periods. Data from monitored days during these periods were then aggregated via statistical measures such as the mean, standard deviation (SD), median, 25th percentile, 75th percentile, and rate of variation (consisting of daily and weekly variation rates). The daily variation rate was calculated by dividing the sum of the variation rates during the monitoring period by the number of monitoring days. The weekly variation rate was calculated via the highest and lowest values in the monitoring period:

Weekly variation rate = \frac{Max value - Min value}{(Max value + Min value) / 2} \times 100 %

This resulted in 17 (3 × 5 + 2) subfeatures per physiological parameter; details are provided in supplementary data 1 (Table S2). For HRV, the same statistical functions produced 105 (21 × 5) subfeatures (Table S2).

Speech signal processing technology was used to denoise and analyze audio signals, segmenting the data into 15-s intervals. Various feature extraction methods were used to derive time- and frequency-domain features that were aggregated into calculated statistics such as the mean and variance. Detailed features are available in supplementary data 1 (Tables S3 and S4). A total of 772 features were extracted from cough sounds, and 1295 features were extracted from blowing sounds. These features were aggregated over the monitoring period, resulting in 3860 (772 × 5) cough sound subfeatures and 6475 (1295 × 5) blowing sound subfeatures. Data for the monitored days were summarized via statistical functions such as the mean, SD, median, 25th percentile, and 75th percentile.

Outcome

The primary outcome was the association between COPD severity and physiological/audio data recorded via a wearable device. The secondary outcomes were (1) the correlations of lung function, the CAT score, the mMRC score, the 6-min walking distance (6MWD) and the BODE index with wearable device data, and (2) participants’ compliance with the devices. Compliance was defined as the percentage of the time required to complete data collection compared with the minimum required time (70 h, calculated as 10 h per day for 7 days, with nighttime data collection from 9 pm to 8 am for 30 h).

Statistical analysis

Count variables are presented as frequencies and percentages, normally distributed continuous variables are presented as the means and SDs, and nonnormally distributed variables are presented as medians and inter-quartile ranges. The participants were grouped by GOLD grade. Due to the small sample size in the GOLD 4 group, this group was combined with the GOLD 3 group. Data from three groups were compared via the chi-square test, ANOVA, and the Kruskal‒Wallis H test, with the exact test used for groups with fewer than five participants. The Bonferroni method was used to adjust the p-value for multiple comparisons. To investigate the correlation between smartwatch physiological data and clinical data, Pearson's test was used for continuous variables with a normal distribution, and Spearman's test was used for continuous variables with a nonnormal distribution and categorical variables.

The subfeature with the strongest correlation from the univariate analysis was selected as the most representative feature. These subfeatures were then used as independent variables in a multiple linear regression model to predict lung function. In addition, an ordered multinomial logistic regression model was formulated to predict COPD severity (GOLD grade). Internal validation was performed via leave-one-out cross-validation (LOOCV), with each participant's data used sequentially as a test set, while the remainder were used for training, and average model performance was calculated.

Statistical analysis was conducted via R version 4.3.2 (R Foundation for Statistical Computing; Vienna, Austria). A p-value below .05 was considered statistically significant for all tests.

Results

Clinical and physiological characteristics

Thirty-one participants were screened; 29 were included after one participant with a BMI less than 18 kg/m² and one participant without COPD were excluded. No arrhythmia was detected on ECG. Data from one participant with mild acute COPD exacerbation from 4 August to 8 August 2022, were excluded.

There were no significant differences in age, sex, BMI, smoking status, CAT score, or mMRC score among the three groups. The BODE index was significantly greater in the GOLD 3 + 4 group than in both the GOLD 1 and GOLD 2 groups, with no significant difference between the GOLD 1 and GOLD 2 groups. Compared with the GOLD 1 group, the GOLD 3 + 4 group had a significantly lower 6MWD; however, this difference was not observed when compared with the GOLD 2 group. The smartwatch data revealed significant differences in the RR and blowing sounds between the GOLD 3 + 4 group and both the GOLD 1 and GOLD 2 groups. In addition, compared with the GOLD 1 group, the GOLD 3 + 4 group presented significantly greater HRV and cough sounds; however, this difference was not observed when compared with the GOLD 2 group. The characteristics of the participants are shown in Table 1.

Table 1.

Characteristics of the subjects according to disease severity.

Characteristic	All(N = 29)	GOLD 1(N = 10)	GOLD 2(N = 10)	GOLD 3 + 4(N = 9)	p-value
Age (years)	67.0 ± 5.8	66.4 ± 7.2	68.2 ± 4.3	66.2 ± 6.1	0.72
Sex (male)	26(89.7)	8(80.0)	9(80.0)	9(100)	0.75
BMI (kg/m²)	23.2(22.2, 26.1)	23.7(21.8, 28.3)	24.3(22.9, 27.0)	22.1(21.9, 24.3)	0.15
Smoking history					0.29
Never smoked	4(13.8)	3(30.0)	1(10.0)	0(0)
Former /Current smoker	25(86.2)	7(70.0)	9(90.0)	9(100.0)
Cumulative smoking exposure, (pack-years)	40.9 ± 26.5	36.7 ± 26.6	48.1 ± 32.2	36.4 ± 21.0	0.60
Hypertension	12(41.4)	4(40.0)	7(70.0)	1(11.1)	0.045
Cardiovascular disease	3(10.3)	1(10.0)	1(10.0)	1(11.1)	1.00
Diabetes	4(13.8)	2(20.0)	2(20.0)	0(0.0)	0.507
Inhaled drug use					0.205
Irregular use	3(10.7)	3(30.0)	0(0.0)	0(0.0)
Current regular use	16(57.1)	4(40.0)	5(55.6)	7(77.8)
Former regular use	9(32.1)	3(30.0)	4(44.4)	2(22.2)
FEV1 (L)	1.7 ± 0.7	2.4 ± 0.5	1.7 ± 0.3^a	1.0 ± 0.2^a ^,b	＜0.001
FEV1% (%)	69.8(43.0, 83.4)	89.3(82.4, 92.0)	69.5(61.9, 73.0)^a	35.4(27.5, 43.0)^a ^,b	＜0.001
FVC (L)	3.3 ± 0.7	3.8 ± 0.8	3.1 ± 0.5	3.0 ± 0.7^a	0.03
FVC% (%)	97.0 ± 17.5	112.4 ± 10.1	94.2 ± 11.3^a	83.2 ± 16.9^a	＜0.001
FEV1/FVC	0.6(0.4, 0.6)	0.6(0.6, 0.7)	0.6(0.5, 0.6)	0.3(0.3,0.4)^a ^,b	＜0.001
Blood gas analysis
PaO₂ (mmHg)	82.9 ± 8.7	85.4 ± 10.3	83.4 ± 6.6	79.3 ± 8.7	0.35
PaCO₂ (mmHg)	40.2 ± 4.1	38.6 ± 3.1	38.9 ± 2.6	43.7 ± 4.7^a ^,b	0.01
CAT score	3.0(2.0, 5.0)	3.0(2.8, 3.0)	2.5(1.5, 4.3）	5.0(3.5,13.5)	0.19
mMRC score					0.08
0	18(62.1)	8(80.0)	7(70.0)	3(33.3)
1	6(20.7)	2(20.0)	1(10.0)	3(33.3)
≥2	5(17.2)	0(0)	2(20.0)	3(33.3)
BODE index	1.0(0, 2.5)	0(0, 1.0)	0.5(0, 1.3)	3.0(2.5, 5.0)^a^,b	＜0.001
BODE group					＜0.001
0∼2	22(75.9)	10(100.0)	10(100.0)	2(22.2)^a^,b
≥2	7(24.1)	0(0)	0(0)	7(77.8)
6MWD (m)	429.9 ± 117.6	491.1 ± 111.2	448.3 ± 57.3	341.6 ± 130.0^a	0.01
6MWD group					0.22
≥350	22 (75.9)	8(80.0)	9(90.0)	5(55.6)
≤349	7(24.1)	2(20.0)	1(10.0)	4(44.4)
Physiological and audio parameters
RR (sleepRR75^c, bpm)	16.0(15.0, 17.5)	15.0(14.8, 16.3)	16.0(15.5, 17.0)	20.0(17.0, 21.5)^a^,b	0.001
HR (HRweekv^d, bpm)	0.77 ± 0.09	0.81 ± 0.07	0.78 ± 0.08	0.73 ± 0.10	0.19
HRV (validRriNum_median^e)	48.4 ± 10.2	42.8 ± 12.2	49.2 ± 6.1	54.4 ± 8.7^a	0.049
SpO₂ (sleepspo2median^f, %)	95.3 ± 1.1	95.7 ± 1.0	95.1 ± 1.1	94.9 ± 1.2	0.30
Forceful blowing sound (RMS_15_UpperQuantile^g)	0.28 ± 0.03	0.31 ± 0.02	0.28 ± 0.02	0.25 ± 0.03^a ^,b	＜0.001
Cough sound (MFCC_83_sd^h)	0.73(0.62, 0.86)	0.61(0.56, 0.64)	0.76(0.64, 0.94)^a	0.84(0.78, 0.99)^a	＜0.001

Data were presented as numbers and percentages [n (%)], mean and standard deviation (mean± SD), median with upper and lower quartiles [median (Q1, Q3)]. FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity; FEV1/FVC: the ratio of FEV1 to FVC.

p < .05 compared to the GOLD 1 group.

p < .05 compared to the GOLD 2 group.

sleepRR75 refers to the 75th percentile of RR during sleep of the monitoring period.

HRweekv refers to the weekly variation rate of HR.

validRriNum_median refers to the median of validRriNum during the monitoring period. validRriNum is defined as the number of valid RR intervals during sleep.

sleepspo2median refers to the median of SpO₂ during sleep of the monitoring period. ^gRMS_15_UpperQuantile refers to the upper quartile of the root mean square of sound energy for blowing sounds. ^hMFCC_83_sd refers to the standard deviation of the mel-frequency cepstral coefficients for cough sounds.

HRV: heart rate variability; HR: heart rate; RR: respiratory rate; GOLD: Global Initiative for Chronic Obstructive Lung Disease; 6MWD: 6-min walking distance; BODE: Body mass index, airflow Obstruction, Dyspnea, and Exercise capacity; CAT: COPD Assessment Test; COPD: chronic obstructive pulmoanry disease; mMRC: modified British Medical Research Council; BMI: body mass index.

The correlations of airflow limitation, lung function, symptom scores, activity capacity and prognosis with smartwatch data

The HR, HRV, RR, cough sounds, and blowing sounds were significantly correlated with the GOLD grade of airflow limitation, with cough sounds showing the strongest correlation (r = 0.7617, p < .001), followed by blowing sounds (r = −0.7293, p < .001), whereas SpO₂ showed no correlation (Table 2, Figure 1). The highest correlation with the GOLD grade was observed in the following variables: the weekly variation rate (HRweekv) for HR, the upper quartile of the nighttime RR (sleepRR75) for the RR, the median number of effective RR intervals during sleep (validRriNum_median) for HRV, the upper quartile of the root mean square of blowing sound energy (RMS_15_UpperQuantile) for blowing sounds, and the SD of the mel-frequency cepstral coefficients (MFCC_83_sd) for cough sounds. Further details on the specific subfeatures of each smartwatch feature can be found in supplementary data 1 (Table S5).

Figure 1.

The correlation between smartwatch data and clinical parameters. In boxplots, the central line indicates the median. The lower and upper hinges indicate the first and third quartiles. The lower and upper whiskers extend from the hinge to the smallest and largest values no further than 1.5 * inter-quartile range from the hinge. The black dots represent outlier.

Table 2.

Matrix of correlations between smartwatch data and clinical variables.

	RR	SpO₂	HR	HRV	Cough sound	BlowingSound
GOLD Grade	0.6664^b^,^hp < .001	−0.2860^b^,^fp = .13	−0.3797^a^,^jp = .04	0.4485^fp = .02	0.7617^ep < .001	−0.7293^hp < .001
CAT score	0.3386^a^,dp = .07	−0.2646^a^,^jp = .17	−0.3315^a^,ⁱp = .08	0.5538^gp < .001	−0.6295^hp < .001	−0.6019^ep < .001
mMRC score	0.2263^c^,^fp = .24	−0.2601^c,^ep = .17	−0.3164^a^,ⁱp = .09	0.4230^gp = .02	0.5953^ep < .001	−0.5628^fp = .001
BODE index	0.6554^c,^fp < .001	−0.2811^b^,^fp = .14	0.2994^c,^hp = .11	0.4171^dp = .03	0.5964^ep < .001	−0.6749^ep < .001
BODE group	0.5909^c^,^fp < .001	−0.3072^a^,^jp = .11	−0.2601^a^,^jp = .17	0.4100^fp = .03	0.6260^ep < .001	0.5972^fp < .001
6MWD	−0.4420^c^,^fp = .02	0.3660^b^,^fp = .05	0.3676^a^,ⁱp = .05	−0.3712^gp = .05	0.6847^fp < .001	−0.5408^ep = .002
6MWD group	−0.2889^a^,ⁱp = .13	−0.4523^a,^fp = .01	−0.2264^a^,^jp = .24	0.3625^gp = .06	0.6260^dp < .001	−0.6646^dp < .001
FEV1%	−0.6129^b^,^hp < .001	0.3068^a^,^fp = .11	0.2792^a^,^jp = .14	−0.3895^fp = .04	−0.7660^ep < .001	0.7820^hp < .001
FEV1	−0.6295^c^,^fp < .001	0.2413^a^,^jp = .21	0.3281^a^,ⁱp = .08	−0.3708^dp = .05	−0.6985^ep < .001	0.6546^hp < .001
FVC%	−0.4869^c^,^fp = .007	−0.2600^b^,^ep = .17	−0.2040^c^,^gp = .29	−0.4406^hp = .02	−0.6837^ep < .001	−0.6724^gp < .001
FVC	−0.3613^c^,^fp = .05	−0.4835^c^,^fp = .008	0.3632^b^,^ep = .05	0.3950^ep = .04	−0.6298^ep < .001	−0.7636^gp < .001
FEV1/FVC	−0.6474^b^,^hp < .001	0.2972^a^,^fp = .12	−0.3376^c^,^hp = .07	0.4467^dp = .02	−0.6480^ep < .001	0.7115^hp < .001

All-day data.

Sleep period data.

Wake period data.

^–jThey represent the data calculated as the mean, standard deviation, median, lower quartile, upper quartile, daily variation rate, and weekly variation rate, respectively.

The smartwatch data were correlated with lung function, the CAT score, the mMRC score, the BODE index, and the 6MWD (Table 2, Figures S1 and 1). The presence of blowing sounds had the strongest correlation with forced vital capacity (FVC) (r = −0.7636, p < .001) and the FEV1/FVC ratio (r = 0.7115, p < .001), whereas the presence of cough sounds had the strongest correlation with FEV1 (r = −0.6985, p < .001).

The presence of cough sounds also had the strongest correlations with the CAT score (r = −0.6295, p < .001), mMRC score (r = 0.5953, p < .001), BODE group (r = 0.6260, p < .001), and 6MWD (r = 0.6847, p < .001). The presence of blowing sounds correlated best with the BODE index (r = −0.6749, p < .001) and the 6MWD group (r = −0.6646, p < .001).

Prediction models

We selected the HR, HRV, RR, blowing sounds, and cough sounds as variables to predict the GOLD grade. When the Akaike information criterion (AIC) was used for selection, SpO₂ was excluded because of an insufficient sample size. The RR and blowing sounds were retained in the optimal model. Internal validation via LOOCV revealed that the accuracy of the ordered multinomial logistic regression model for determining COPD severity was 0.77–0.89, as shown in Table 3. The Cohen's kappa coefficient of the model was 0.6757, indicating substantial agreement between the estimated severity and the true value.

Table 3.

The performance of the model to distinguish the severity of COPD.

	GOLD 1	GOLD 2	GOLD 3 + 4
Precision (positive predictive value)	0.8182	0.7000	0.8571
Recall (sensitivity)	0.9000	0.7000	0.7500
Sensitivity	0.9000	0.7000	0.7500
Specificity	0.8889	0.8333	0.9500
Positive predictive value	0.8182	0.7000	0.8571
Negative predictive value	0.9412	0.8333	0.9048
Balanced Accuracy	0.8944	0.7667	0.8500

Cohen's kappa coefficient is κ = 0.6757. COPD: chronic obstructive pulmonary disease; GOLD: Global Initiative for Chronic Obstructive Lung Disease.

The most highly correlated data for each parameter were selected from the univariate analysis to create a multiple linear regression analysis model to predict lung function (FEV1, FVC, FEV1/FVC ratio). The most relevant selected subfeatures are listed in supplementary data 1 (Table S5). The variables for the final model were selected via a bidirectional stepwise procedure to minimize the AIC, as detailed in Table 4 and Figure 2. For FEV1 prediction, the performance metrics of the model were as follows: coefficient of determination (R²): 6450, root mean square error (RMSE): 0.3990 L, and mean absolute error (MAE): 0.2998 L. For FVC prediction, the model performance metrics were as follows: R²: 0.7098, RMSE: 0.3979 L, and MAE: 0.2978 L. For FEV1/FVC prediction, the performance metrics of the regression model were as follows: R²: 0.5078, RMSE: 0.0927 L, and MAE: 0.0707 L.

Figure 2.

Regression model for lung function predictions versus true values. The gray curve is the line of equality.

Table 4.

The summary information for the lung function regression models.

FEV1 = 5.73-0.23RR + 1.18blowing sound_MFCC-0.09cough sound_MFCC
R²	p	RMSE	MAE
0.6450	<.001	0.3990	0.2998
FVC = 17.82-0.22SpO₂-1.18blowing sound_MFCC-4.36cough sound_LPC
R²	p	RMSE	MAE
0.7098	<.001	0.3979	0.2978
FEV1/FVC = 1.59-0.03RR + 0.01HRV + 0.04blowing sound_MFCC
R²	p	RMSE	MAE
0.5078	<.001	0.0927	0.0707

R²: coefficient of determination; RMSE: root mean square error; MAE: mean absolute error; HRV: heart rate variability; RR: respiratory rate; FEV1: forced expiratory volume in 1 s; FVC: forced vital capacity.

Compliance of the participants with the smartwatch

Patients wore the smartwatch for a median of 19.8 h per day (Q1: 18.3, Q3: 20.9) and 10.0 h per night (Q1: 9.9, Q3: 10.3). All patients had an average of more than 10 h of daily wear and more than 30 h of nightly data collection, achieving 100% compliance.

Discussion

Principal results

Our study identified two key findings that advance the use of wearable devices in assessing COPD. First, a smartwatch equipped with audio collection capabilities demonstrated efficacy in assessing airflow limitation, activity capacity, and prognosis in patients with COPD. Second, audio metrics outperformed traditional physiological parameters in assessing multiple dimensions of COPD.

Physiological parameters from wearable devices in chronic obstructive pulmonary disease assessment

Wearable devices have increasingly been used for assessing the severity of COPD. Tiwari et al. used data on HRV and physical activity metrics collected by smartwatches to predict the severity of COPD symptom, achieving an F1 score of 0.41 and an area under the curve of 0.64.²⁵ Odame et al. employed a respiratory chest belt to measure parameters such as the RR in patients with suspected and confirmed COPD, reporting significant correlations with spirometric variables (r = 0.274–0.329, p < .05).²⁴ They also built a model to predict COPD severity, achieving an accuracy of 96.4% for determining GOLD grade (GOLD 1+2/GOLD 3+4). In another study, Bédard et al. reported a positive correlation (r = 0.342, p = .028) between 24-h HRV from ambulatory ECG and FEV1 in COPD patients.³⁶ In our study, we revealed that the RR was more strongly correlated with the GOLD classification (r = 0.6664, p < .001) than with HRV (0.4485, p = .02). The optimal ordered logistic regression model for predicting GOLD classification identified the RR as a key predictor, with the upper quartile of the sleep RR being the most relevant subfeature. This may be because the sleep RR is less influenced by motion artifacts and better reflects the true state of the respiratory system. In conclusion, we found that the RR measured by wearable devices is an effective indicator for assessing COPD.

Audio metrics in chronic obstructive pulmonary disease assessment

In addition to physiological parameters, we found that audio-based metrics such as cough and forceful blowing sounds had a stronger correlation with clinically relevant outcomes. In previous studies, they have primarily focused on using cough sounds to detect COPD or asthma, as well as to estimate lung function.^{19,20,22,23,37–40} For instance, some studies predicted the FEV1/FVC ratio via cough sounds with MAEs ranging from 8% to 16.4%^19,20 and blowing sounds ranging from 5.1% to 7.57%.^22,39 There are also studies based on “voice” prediction of lung function, which predicted the FEV1/FVC ratio with an MAE of 7.4% to 12.5%.^41,42 These findings suggest that sounds seem to have good performance in predicting lung function. The theoretical basis for this is that both coughing and forceful blowing involve pulmonary mechanisms similar to those used during the spirometry test, which includes deep inhalation followed by forceful exhalation, thereby predicting lung function with greater accuracy. However, current methods predominantly rely on smartphones or microphones to capture audio data. The key technological innovation of this study is the use of smartwatches, which are more comfortable and convenient than other body-worn sensors, to collect these sounds, offering a promising alternative for monitoring lung function.

Clinical relevance of wearable devices data in chronic obstructive pulmonary disease assessment

Our study also revealed that smartwatch data, especially the RR and cough and blowing sounds, strongly correlated with clinical outcome indicators such as the mMRC score, CAT score, 6MWD, and the BODE index. These indicators are important for guiding treatment and rehabilitation and for predicting the prognosis of COPD.⁸ These findings suggest that our disease assessment model has the potential to complement in-hospital care by enabling continuous, objective monitoring of a patient's condition at home. This could facilitate timely adjustments to disease management plans, ultimately improving patient outcomes.

Application of prediction model in chronic obstructive pulmonary disease management

Remote monitoring is essential for future COPD management and relies on accurate predictive models of disease progression and patient adherence. Previous studies have revealed that wearable devices are valuable for early prediction of AECOPD through changes in physiological parameters, such as a decrease of 653 steps per day, an increase in the breathing rate, and poor sleep quality.^43–45 Additionally, the remote monitoring of clinical symptoms and physiological parameters, such as the HR, temperature, oxygen saturation, RR, and number of steps walked, combined with machine deep learning technology has shown high sensitivity and specificity in predicting AECOPD.⁴⁶ In this context, our study contributes further by achieving full short-term adherence to smartwatch usage and developing an assessment model with superior predictive accuracy for lung function. Our model demonstrated a lower RMSE and higher R² compared to methods used in previous studies.²⁴ Furthermore, our model has a stable recall rate of 70% to 90%. In the future, the model established in this study can be applied to remote daily home monitoring of COPD patients to assess its feasibility, cost-effectiveness, and impact on outcomes, which will have a certain application value in disease assessment. Research should also explore how wearable devices can support personalized disease management, including using machine learning models to predict exacerbations through remote monitoring.

Limitations

While this study provides valuable insights into the use of wearable devices for COPD assessment, several limitations must be considered. First, the relatively small sample size may limit the generalizability of the findings. Larger cohorts in future studies are needed to validate these results and establish more robust conclusions. Second, the study sample was predominantly male and had a narrow age range, which may limit the applicability of the findings to more diverse populations. Future studies are needed to validate the ability to assess how well the wearable devices perform across different demographic groups, such as females or individuals of different age ranges. Finally, the study did not explore potential confounders that could influence the outcomes, such as comorbidities, which should be considered in future research.

Conclusions

In conclusion, this study involved the innovative use of smartwatches for audio data collection, and a model built on the basis of cough sounds, forceful blowing sounds, and the RR can accurately assess COPD severity and predict lung function. Technological advancements have made COPD disease assessment more comfortable and convenient and offer a potential tool for COPD remote home surveillance.

Supplemental Material

sj-docx-1-dhj-10.1177_20552076251320730 - Supplemental material for Intelligent wearable devices with audio collection capabilities to assess chronic obstructive pulmonary disease severity

Supplemental material, sj-docx-1-dhj-10.1177_20552076251320730 for Intelligent wearable devices with audio collection capabilities to assess chronic obstructive pulmonary disease severity by Chunbo Zhang, Kunyao Yu, Zhe Jin, Yingcong Bao, Cheng Zhang, Jiping Liao and Guangfa Wang in DIGITAL HEALTH

Supplemental Material

sj-docx-2-dhj-10.1177_20552076251320730 - Supplemental material for Intelligent wearable devices with audio collection capabilities to assess chronic obstructive pulmonary disease severity

Supplemental material, sj-docx-2-dhj-10.1177_20552076251320730 for Intelligent wearable devices with audio collection capabilities to assess chronic obstructive pulmonary disease severity by Chunbo Zhang, Kunyao Yu, Zhe Jin, Yingcong Bao, Cheng Zhang, Jiping Liao and Guangfa Wang in DIGITAL HEALTH

Footnotes

Acknowledgements

We would like to acknowledge Huawei Technologies Co.,Ltd for the development and optimization of the device and HUAWEI Research platform to this study.

Contributorship

JL and GW conceived and designed the study. CBZ, KY, and JL wrote the manuscript. KY, ZJ, YB, and CZ contributed to the data collection. CBZ, KY, YB, JL, and CZ performed statistical analyses. CBZ and KY were responsible for drawing the figures. CBZ, KY, ZJ, JL, and GW contributed to the discussion of the findings and revised the manuscript accordingly. All authors have approved the final submission.

Data availability

The datasets used during the current study are available from the corresponding author on reasonable request.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethical approval

The ethics committee of Peking University First Hospital approved this study (Approval number: 2022083). All the participants provided signed informed consent forms.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Beijing Health Technologies Promotion Program, the Capital’s Funds for Health Improvement and Research, National High Level Hospital Clinical Research Funding (High Quality Clinical Research Project of Peking University First Hospital), National High Level Hospital Clinical Research Funding (Interdepartmental Research Project of Peking University First Hospital), (grant number BHTPP202053, CFH 2022-1G-4073, NO. 2023HQ01, NO. 2023IR49).

Guarantor

ORCID iDs

Chunbo Zhang

Guangfa Wang

Supplemental material

Supplemental material for this article is available online.

References

Lozano

Naghavi

Foreman

, et al. Global and regional mortality from 235 causes of death for 20 age groups in 1990 and 2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet 2012; 380: 2095–2128.

Mathers

Loncar

. Projections of global mortality and burden of disease from 2002 to 2030. PLoS Med 2006; 3: e442.

Prevalence and attributable health burden of chronic respiratory diseases, 1990-2017: a systematic analysis for the global burden of disease study 2017. Lancet Respir Med 2020; 8: 585–596.

Christenson

Smith

Bafadhel

, et al. Chronic obstructive pulmonary disease. Lancet 2022; 399: 2227–2242.

Viegi

Pistelli

Sherrill

, et al. Definition, epidemiology and natural history of COPD. Eur Respir J 2007; 30: 993–1013.

Wang

Yang

, et al. Prevalence and risk factors of chronic obstructive pulmonary disease in China (the China Pulmonary Health [CPH] study): a national cross-sectional study. Lancet 2018; 391: 1706–1717.

Zhong

Wang

Yao

, et al. Prevalence of chronic obstructive pulmonary disease in China: a large, population-based survey. Am J Respir Crit Care Med 2007; 176: 753–760.

Global strategy for the diagnosis, management and prevention of chronic obstructive pulmonary disease 2024 report [updated 2023-11-132024-7-9], https://goldcopd.org/2024-gold-report/.

Masekela

Zurba

Gray

. Dealing with access to spirometry in Africa: a commentary on challenges and solutions. Int J Environ Res Public Health 2019; 16: 62.

10.

Cusack

Chaudhary

, et al. Under- and over-diagnosis of COPD: a global perspective. Breathe 2019; 15: 24–35.

11.

Cruz

Brooks

Marques

. Home telemonitoring in COPD: a systematic review of methodologies and patients’ adherence. Int J Med Inf 2014; 83: 249–263.

12.

Carpenter

Jurdi

Roberts

, et al. A review of portable electronic spirometers: implications for asthma self-management. Curr Allergy Asthma Rep 2018; 18: 53.

13.

Lukas

, et al. Emerging telemedicine tools for remote COVID-19 diagnosis, monitoring, and management. ACS nano 2020; 14: 16180–16193.

14.

Warnier

Rutten

Numans

, et al. Electrocardiographic characteristics of patients with chronic obstructive pulmonary disease. Copd 2013; 10: 62–71.

15.

van Gestel

Steier

. Autonomic dysfunction in patients with chronic obstructive pulmonary disease (COPD). J Thorac Dis 2010; 2: 215–222.

16.

Rahman

Nemati

Rahman

, et al. Automated assessment of pulmonary patients using heart rate variability from everyday wearables. Smart Health 2020; 15: 100081.

17.

Rahman

Ahmed

, et al. Breatheasy: assessing respiratory diseases using mobile multimodal sensors. In: Proceedings of the 2020 international conference on multimodal interaction; Virtual event, Netherlands, 2020. pp.41–49. Association for Computing Machinery.

18.

Chatterjee

Rahman

Ahmed

, et al. Assessing severity of pulmonary obstruction from respiration phase-based wheeze-sensing using mobile sensors. In: Proceedings of the 2020 CHI conference on human factors in computing systems, Honolulu, HI, USA, 2020, pp.1–13. Association for Computing Machinery.

19.

Porter

Claxton

Brisbane

, et al. Diagnosing chronic obstructive airway disease on a smartphone using patient-reported symptoms and cough analysis: diagnostic accuracy study. JMIR Form Res 2020; 4: e24587.

20.

Pan

, et al. A forced cough sound based pulmonary function assessment method by using machine learning. Front Public Health 2022; 10: 1015876.

21.

Goel

Saba

Stiber

, , et al. Spirocall: measuring lung function over a phone call. In: Proceedings of the 2016 CHI conference on human factors in computing systems, San Jose, California, USA, 2016, pp.5675–5685. Association for Computing Machinery.

22.

Larson

Goel

Boriello

, et al. Spirosmart: using a microphone to measure lung function on a mobile phone. In: Proceedings of the 2012 ACM conference on ubiquitous computing, Pittsburgh, Pennsylvania, 2012, pp.280–289. Association for Computing Machinery.

23.

Infante

Chamberlain

Fletcher

, et al.

Use of cough sounds for diagnosis and screening of pulmonary disease

. In: 2017 IEEE Global Humanitarian Technology Conference (GHTC),

San Jose, CA, 2017, pp.1–10. IEEE.

24.

Odame

Atkins

Nyamukuru

, et al.

Inferring COPD Severity from Tidal Breathing.

In: 2021 43rd Annual international conference of the IEEE engineering in medicine & biology society (EMBC),

Mexico, 2021, pp.1392–1398. IEEE.

25.

Tiwari

Liaqat

, et al. Remote COPD severity and exacerbation detection using heart rate and activity data measured from a wearable device. In: 2021 43rd Annual international conference of the IEEE engineering in medicine & biology society (EMBC), Mexico, 2021, pp.7450–7454. IEEE.

26.

Serrurier

Neuschaefer-Rube

Röhrig

. Past and trends in cough sound acquisition, automatic detection and automatic classification: a comparative review. Sensors 2022; 22: 2896.

27.

Global strategy for the diagnosis, management and prevention of chronic obstructive pulmonary disease 2022 report [updated 2021-11-152024-4-29], https：//goldcopd.org/gold-reports/.

28.

Celli

Cote

Marin

, et al. The body-mass index, airflow obstruction, dyspnea, and exercise capacity index in chronic obstructive pulmonary disease. N Engl J Med 2004; 350: 1005–1012.

29.

Anthonisen

Manfreda

Warren

, et al. Antibiotic therapy in exacerbations of chronic obstructive pulmonary disease. Ann Intern Med 1987; 106: 196–204.

30.

Wilkinson

Donaldson

Hurst

, et al. Early therapy improves outcomes of exacerbations of chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2004; 169: 1298–1303.

31.

Seemungal

Donaldson

Paul

, et al. Effect of exacerbation on quality of life in patients with chronic obstructive pulmonary disease. Am J Respir Crit Care Med 1998; 157: 1418–1422.

32.

Donaldson

Wilkinson

Hurst

, et al. Exacerbations and time spent outdoors in chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2005; 171: 446–452.

33.

Bhowmik

Seemungal

Sapsford

, et al. Relation of sputum inflammatory markers to symptoms and lung function changes in COPD exacerbations. Thorax 2000; 55: 114–120.

34.

Donaldson

Seemungal

Bhowmik

, et al. Relationship between exacerbation frequency and lung function decline in chronic obstructive pulmonary disease. Thorax 2002; 57: 847–852.

35.

Seemungal

Harper-Owen

Bhowmik

, et al. Respiratory viruses, symptoms, and inflammatory markers in acute exacerbations and stable chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2001; 164: 1618–1623.

36.

Bédard

Marquis

Poirier

, et al. Reduced heart rate variability in patients with chronic obstructive pulmonary disease independent of anticholinergic or β-agonist medications. Copd 2010; 7: 391–397.

37.

Ramesh

Vatanparvar

Nemati

, et al. CoughGAN: generating synthetic coughs that improve respiratory disease classification. In: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 2020, pp.5682–5688. IEEE.

38.

Nemati

Rahman

Blackstock

, et al Estimation of the lung function using acoustic features of the voluntary cough. In: 2020 42nd Annual international conference of the IEEE engineering in Medicine & biology society (EMBC), Montreal, QC, Canada, 2020, pp.4491–4497. IEEE.

39.

Rahman

Ahmed

Nemati

, et al. ExhaleSense: Detecting high fidelity forced exhalations to estimate lung obstruction on smartphones. In: 2020 IEEE International conference on pervasive computing and communications (PerCom), Austin, TX, USA, 2020, pp.1–10. IEEE.

40.

Viswanath

Garrison

Patel

. Spiroconfidence: determining the validity of smartphone based spirometry using machine learning. In: 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Honolulu, HI, USA, 2018, pp.5499–5502. IEEE.

41.

Saleheen

Ahmed

Rahman

, et al. Lung function estimation from a monosyllabic voice segment captured using smartphones. In: 22nd International conference on human-computer interaction with mobile devices and services, Oldenburg, Germany, 2020, p.Article 10. Association for Computing Machinery.

42.

Chun

Nathan

Vatanparvar

, et al. Towards passive assessment of pulmonary function from natural speech recorded using a mobile phone. In: 2020 IEEE international conference on pervasive computing and communications (PerCom), Austin, TX, USA, 2020, pp.1–10. IEEE.

43.

Crook

Büsching

Keusch

, et al. The association between daily exacerbation symptoms and physical activity in patients with chronic obstructive pulmonary disease. Int J Chron Obstruct Pulmon Dis 2018; 13: 2199–2206.

44.

Shorofsky

Bourbeau

Kimoff

, et al. Impaired sleep quality in COPD is associated with exacerbations: the CanCOLD cohort study. Chest 2019; 156: 852–863.

45.

Yañez

Guerrero

Pérez de Alejo

, et al. Monitoring breathing rate at home allows early identification of COPD exacerbations. Chest 2012; 142: 1524–1529.

46.

Kouri

Gupta

Yadollahi

, et al. Addressing reduced laboratory-based pulmonary function testing during a pandemic. Chest 2020; 158: 2502–2510.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.04 MB

0.07 MB