A prospective observational study was conducted to recruit case and control individuals from January 2021 to November 2022. The case group comprised of patients with confirmed COPD, while the control group consisted of individuals without any definite diagnosis, and both these groups were recruited from the Guangdong Provincial Technology Research Center of COPD Rehabilitation, the Third Affiliated Hospital of Sun Yat-sen University. Specifically, the case group was derived from the outpatient, emergency, and inpatient departments of this hospital, and the control group consisted of individuals recruited by the Guangdong Provincial Technology Research Center of COPD Rehabilitation. Pulmonary function examination was performed either at or prior to the enrollment of the case group. According to the 2022 Global Initiative for Chronic Obstructive Lung Disease (GOLD) diagnostic criteria, if the ratio of the forced expiratory volume in the first one second to the forced vital capacity of the lungs is less than 70% after inhalation of the bronchodilator, one will be considered to have persistent airflow limitation [21]. This is the gold standard for COPD diagnosis and the only inclusion criterion for case groups. The exclusion criteria for the case and control groups were as follows:

This study obtained the written informed consent from all participants, and the study protocol was reviewed by the medical ethics committee of the Third Affiliated Hospital of Sun Yat-sen University (approval [2021]02-019-01).

We obtained the pertinent data of individuals in the case group through the hospital information management system, including demographic characteristics, disease history, pulmonary function test data, and acute exacerbations. We also recorded the demographic characteristics and the disease history of participants in the control group. Piezoelectric sensor–based HR and respiratory recorders (model: WSM-LN-03, medical device; China Food and Drug Administration; 20182071130) were distributed to the participants and placed under the pillow to continuously record the HR, respiration, and HRV during night sleep from 10 PM to 8 AM the next day. Prior to the examination, all participants passed the standardized training on how to use the vital sign monitoring device. The recorded vital signs were manually uploaded to the medical cloud server for further data analysis. The architecture of the proposed study is shown in Figure 1. Due to the long duration of the experiment, there may be cases of inevitable data loss. Therefore, specific exclusion criteria with regarding to the data on the medical cloud server were as follows:

The distributed vital signs monitoring device was deployed under the pillow. With piezoelectric sensors, the analogy signals representing cardiac and respiratory activities are recorded with a sampling frequency of 1 kHz. Specifically designed filter banks can detect the heartbeat interval and respiration per minute during night sleep, and the averaged HR (beat/min) and respiration (times/min) of a whole night can be obtained [22-24]. In addition, by taking advantage of the heartbeat interval, both time and frequency domain HRV parameters can be calculated, as illustrated in Table 1 [25]. Specifically, HRV contains 7 times/frequency domain features, including standard deviation of heartbeat interval (SDNN), total power of heartbeat interval (TP), low frequency of TP (LF), high frequency of TP (HF), low frequency/high frequency (LF/HF), very low frequency of TP (VLF), and ultralow frequency of TP (ULF). The features of respiration also contain the cumulative duration of RR faster than 21 times/min (RRF).

Since the bed sensors continuously monitor during night, HRV, HR, and respiration can be continuously monitored over weeks without affecting the lifestyle of the recruited patients. Considering that the features of HRV, HR, and respiration obtained in a single night sleep monitoring could be influenced by the daytime emotions and health status of the patients, we propose to use the statistical average of the continuously monitored features (HRV, HR, and respiration) over weeks instead of that recorded in a single day for further analysis. Numerically, the statistical average of the features HRV, HR, and respiration over t consecutive days are calculated by the following method. We first calculated the sum of the values of HRV, HR, and respiration for t consecutive days; then subtracted the maximum and minimum values in the values of HRV, HR, and respiration for t consecutive days; and then obtained the statistical means by dividing by t–2. In this study, we set t=1, 7, 14, and 30 days as the cutoff observation time for analysis.

The purpose of this study was to evaluate the diagnostic value of continuous HRV and respiration monitoring in patients with COPD and control groups by using pulmonary function examination as the diagnostic gold standard. We expected continuous HRV and respiration monitoring to diagnose COPD with sensitivity of 90%, tolerance error of 10% of sensitivity, specificity of 90%, tolerance error of 10% of specificity, and the test criteria α=.01. The sample size for patients with COPD and control groups was set at 1:2, and the dropout rate was set as 10%. Finally, we used PASS 11.0 software (NCSS Statistical Software) to calculate the sample size, and we found that there should be at least 37 patients with COPD and 74 controls, for a total of 111 participants.

Statistical analysis was conducted using SPSS software (version 20.0; IBM Corp). The main statistical methods included descriptive analysis, χ² test, nonparametric test, binary logistic regression, and receiver operating characteristic (ROC) curves. For statistical description, counting data were described by frequency and rate, and metrological data were described by median and quartile. For statistical tests, the χ² test was employed to compare the 2 groups in terms of the counting data, with the χ² value serving as the test statistic. The metrological data were analyzed using a nonparametric test, specifically the Mann-Whitney U test. The z score was employed as the test statistic. In the multifactor analysis, the classical binary logistic regression was applied for classification, and the ROC curve was used to evaluate the diagnosis of each variable. The determination of the cutoff values relied on the maximum Youden index as the tangent point. The test criterion was set as α=.05, and P<.05 was considered statistically significant. The classification cutoff point was set as 0.5.

The experimental process of this study is shown in Figure 2. We recruited 155 participants corresponding to a total number of 21,352 night-sleep monitored samples (person-time) of patients with COPD and healthy participants in the Third Affiliated Hospital of Sun Yat-sen University, in accordance with the abovementioned inclusion criteria for the population. After data cleaning based on the abovementioned exclusion criteria, 146 patients corresponding to a final number of 13,260 samples (person-time) were included, corresponding to case and control groups of 4608 and 8652 samples, respectively. Subsequently, the statistically averaged features of long-term HRV, HR, and respiration over the observation time of 7, 14, and 30 days were constructed for further analysis.

A total of 146 individuals who met the inclusion and exclusion criteria were enrolled in this study, comprising 37 cases and 109 controls. There were 12 mild to moderate cases and 25 severe cases. The overall compliance rate among the participants was 0.73 (IQR 0.50-0.86), with no statistically significant difference between the 2 groups (z score=–1.5; P=.15). Moreover, the average number of continuous night-sleep monitoring days per person was found to be 56.5 (IQR 32.0-113.0) days, showing no significant difference between the 2 groups again (z score=0.5; P=.65). Therefore, compliance and monitoring duration between the 2 groups were comparable. Significant differences were observed between the 2 groups in terms of age, gender, and BMI (P<.001). Specifically, the case group exhibited a higher proportion of older individuals, males, and those with low BMI in comparison to the control group (as depicted in Table 2).

This study consists of 13,260 samples (person-time), corresponding to 8652 in the control group and 4608 in the case group. During the monitoring period of 7, 14, and 30 days, there was a total of 12,396, 11,422, and 9404 person-times recorded, respectively. In 4 cutoffs of the observation time of 1, 7, 14, and 30 days, the distributions of HRV, HR, and respiration features, that is, SDNN, TP, LF, HF, VLF, ULF, HR, RR, and RRF, averaged over different timescales were of significant statistical difference, as shown in Tables S1-S4 of Multimedia Appendix 1. Specifically, the values of all the averaged features in the different timescales in the case group were significantly higher than those of the control group (P<.001). Interestingly, no significant difference was observed in the distribution of the features LF/HF between the 2 groups (P>.05). Consequently, when performing a binary logistic regression analysis, all features of LF/HF averaged over different timescales (t= 1, 7, 14, and 30 days), that is, LF/HF₁, LF/HF₇, LF/HF₁₄, and LF/HF₃₀ were excluded, while the rest 36 features were included.

HRV, HR, and respiratory features at different timescales are represented by features and timescales, where timescales are in subscripts. The overall classification accuracy of HRV, HR, and respiration at 1, 7, 14, and 30 days achieved a high level of 0.958, with an impressive Youden index of 0.904 and an excellent area under the ROC curve (AUC) of 0.989 (as presented in Table 3 and Figure 3A).

We combined the vital signs features based on 4 different timescales (t= 1, 7, 14, and 30 days) and subsequently classified the participants into case and control groups. Figure 4 demonstrates that the classification performance of HRV, HR, and respiration improved as the monitoring duration increased, with HRV, HR, and respiration at 30 days exhibiting superior performance (accuracy 0.956, Youden index 0.899, AUC 0.989). Specifically, the ranking of the performance was presented as follows: HRV, HR, and respiration in 30 days > HRV, HR, and respiration in 14 days > HRV, HR, and respiration in 7 days > HRV, HR, and respiration in 1 day.

We combined them again based on the vital signs category and subsequently classified those participants into the case and control groups. Figure 4 demonstrates an improving trend in the classification efficacy of respiration, HRV, and HR, surpassing that of single-day monitoring. On the one hand, when evaluating the classification effect using accuracy and Youden index, we observed an improvement in the classification effect of HRV with the increasing duration of monitoring, where HRV₃₀ demonstrated the highest performance (accuracy 0.866, Youden index 0.717). Similarly, as the monitoring duration increased, respiration also exhibited enhanced performance, with respiration at 30 days achieving the best results (accuracy 0.913, Youden index 0.792). However, there was minimal variation in the classification performance of HR across different timescales, maintaining an accuracy rate of approximately 0.67 and a Youden index of around 0.15. On the other hand, AUC was utilized to be evaluated. The findings demonstrated that all AUC values exceeded 0.5 and exhibited statistical significance. Notably, the classification performance of HRV improved with longer monitoring duration, specifically presented as HRV₃₀ > HRV₁₄ > HRV₇ > HRV₁. Similarly, the performance of respiration gradually tended to increase with extended monitoring duration, indicated by respiration over 30 days/respiration over 14 days > respiration over 14 days/respiration over 7 days > respiration over 1 day. However, there were no statistically significant differences between respiration over 14 days and respiration over 7 days or between respiration over 14 days and respiration over 30 days. Furthermore, the performance of HR also progressively enhanced with prolonged monitoring time (HR₃₀/HR₁₄/HR₇ > HR₁). Nevertheless, there were no statistically significant differences among HR₇, HR₁₄, and HR₃₀, as shown in Table 3.

Table 3 and Figures 3B-3E demonstrate that the classification performance, as evaluated by the accuracy, Youden index, and AUC, consistently showed respiration > HRV > HR over a single timescale. This consistent pattern was observed over all 4 timescales. In addition, the sensitivity of HR remained consistently low over the 4 timescales, ranging from approximately 0.2 to 0.3, indicating a high potential for missed diagnoses rate (1–sensitivity) of up to 0.7-0.8. Specifically, the sensitivity of HR₁ was 0.177, with a corresponding missed diagnosis rate of 0.823. These findings highlight that only using HR features for the diagnosis of patients with COPD may lead to a significant increase in missed diagnosis rates exceeding 80%.

The results shown in Figure 4 indicate that the classification performance using the 7-day monitored features (HRV, HR, respiration; AUC=0.976) is already comparable to that of the 30-day monitored features (HRV, HR, respiration; AUC=0.989). Therefore, considering that the performance of monitoring longer than 7 days tends to saturate, while increasing the monitoring time leads to an increase in the dropout rate, we recommend continuous monitoring for 7 days during daily home life, which can already provide valuable assistance in the auxiliary diagnosis of COPD.

We classified the case and control groups by a single measured vital sign feature during the 7-day long-term monitoring period. As shown in Table 4, the results revealed that the HF₇ exhibited the most effective classification performance (accuracy 0.830, Youden index 0.602), followed by TP₇ (accuracy 0.792, Youden index 0.474) among the 6 features (HRV) evaluated by the accuracy and Youden index. Among the features set (HRV, HR, respiration)₇, RRF₇ demonstrated the best performance (accuracy 0.898, Youden index 0.733), followed by HF₇, while HR₇ showed the poorest performance (accuracy 0.674, Youden index 0.147). Furthermore, both HR₇ and ULF₇ exhibited very low sensitivity with the missed diagnosis rate exceeding 70%, indicating that using HR₇ or ULF₇ for classification and prediction could lead to higher chances of missed diagnosis.

Additionally, AUC was used to evaluate the effect of the features on the classification performance. As depicted in Figure 5, the AUCs of all 7-day time-scale monitored features were greater than 0.5 and demonstrated statistical significance (P<.001). The highest AUC value was observed for RRF₇ at 0.955. The order of the classification effects was as follows: RRF₇ > HF₇ > RR₇ > LF₇/TP₇ > SDNN₇ > VLF₇ > ULF₇/HR₇. Notably, there were no statistically significant differences between LF₇ and TP₇, ULF₇, and HR₇.

Considering that the classification performance of HF₇, RR₇, and RRF₇ was superior to that of other features, which reflected the potential values of such features for the diagnosis of COPD, we further analyzed the cutoff values based on the maximum Youden index for HF₇, RR₇, and RRF₇. This serves as a reference for auxiliary diagnosis in patients with COPD. The determined cutoff value for HF₇ was 1316.30 nU, which corresponded to sensitivity and specificity values of 0.836 and 0.810, respectively. The cutoff value for RR₇ was 16.3 times/min, accompanied by sensitivity and specificity values of 0.783 and 0.803, respectively. As for RRF₇, the designated cutoff value was set at 0.1 minute with corresponding sensitivity and specificity values of 0.938 and 0.925, respectively. Samples exceeding these respective cutoff values would be assigned to the case group, while the others would be assigned to the control group (as illustrated in Table 5).

The primary purpose of this study was to explore the feasibility of bed sensor–aided continuous night-sleep monitoring of vital signs (HRV, HR, and respiration) to assist in the remote diagnosis of COPD. Our findings show that continuous night-sleep monitoring has significant potential in the diagnosis of COPD, and the statistical means of RRF, HF, and RR are the crucial features. By extracting statistical features from the continuously recorded HRV, HR, and respiration over 7, 14, and 30 days, the classification performance of the vital signs for COPD diagnosis was evaluated and compared in different timescales, and the cutoff values of the most important features were obtained as a reference point for in-home diagnosis of COPD. Our prospective cohort experimental analysis verified that the classification performance of the diagnosis of COPD increases directionally proportional to the monitoring duration of vital signs recording. This outcome is expected due to the influence of various factors such as environment, daytime emotions, and activities on the recorded vital signs during a single night sleep [26,27]. Comparatively, by increasing the statistical timescale with regard to feature extraction, the derived statistical features can reflect the status of individual health more accurately [28,29]. Under the tradeoff between the performance of diagnosis and the feasibility of continuous monitoring, it is recommended that a 7-day daily home-based continuous monitoring is sufficient to assist in the diagnosis of COPD.

According to previous research, the resting HR and respiration of patients with COPD are higher than those of healthy individuals. However, there is still a lack of literature support regarding the comparison of the correlation between these indicators and patients with COPD [30]. In this study, we discovered that long-term monitored features such as RR, HF, and RRF were more significant for diagnosing COPD than HR alone. Using either RR or HF as features, the classification performance accuracy already exceeded 80%. Furthermore, when utilizing RRF for classification, the accuracy reached 90%. In comparison, if HR was used for classification purposes only, the accuracy dropped to 67%, with a missed diagnosis rate exceeding 70%.

The resting RR for healthy individuals ranges from 12 to 20 beats per minute [31]. However, our findings showed that RR exceeded 16 beats per minute, which can be viewed as an important feature for the diagnosis of COPD. This finding is expected since patients with COPD are more prone to experiencing rapid breathing [32]. The cutoff value for RRF in this study was determined as 0.1 minute, indicating that any presence of rapid breathing during night sleep significantly increases the probability of a COPD diagnosis. The corresponding sensitivity and specificity values reach up to 94% and 93%, respectively.

Patients with COPD exhibit abnormal autonomic nervous function and reduced HRV [33,34]. Studies have demonstrated a decrease in sympathetic regulation among stable patients with COPD, suggesting a potential shift toward parasympathetic regulation [35]. By contrast, our findings show that patients with COPD exhibit an elevated HRV compared to healthy controls. The possible reasons are 2-fold. On the one hand, the HRV indicators of patients with COPD were increased due to cardiac compensation since all patients with severe cardiovascular diseases were excluded in this study. This phenomenon was also validated by a pioneer study [36]. On the other hand, as pointed out by [37], parasympathetic regulation is more active in patients with acute exacerbation of COPD compared to those with stable COPD. In this study, although the events of acute exacerbation in the case group were excluded before analysis, potential labelling errors of the acute exacerbation cases existed, since the recording of periodic follow-up might have been incomplete due to out-of-hospital visits or recall bias, and small amount of monitoring data during acute exacerbations were included, which could account for the observed increase in HRV among patients with COPD.

LF reflects the regulation of the sympathetic nerve, HF reflects the activity of the parasympathetic nerve, and LF/HF reflects the relative activity of the sympathetic nerve and the parasympathetic nerve, that is, the balance between the sympathetic nerve and the parasympathetic nerve [36]. The results of our study showed that there was no statistically significant differences in the LF/HF characteristics between the case group and the control group, which indicated that in patients with COPD, the LF and HF both increased, thus resulting in the relatively constant LF/HF indices. This is also consistent with previous studies, which demonstrated no significant difference between the LF/HF of patients with COPD and the control groups [37].

This study is the first prospective observational study based on continuous vital signs monitoring during night sleep to assist in-home diagnosis of COPD. With the help of noncontact sensors placed under the mattress, we performed continuous night sleep monitoring for individuals with COPD and healthy individuals over weeks and employed the statistical means of the monitored HR, respiration, and HRV at different timescales as features instead of the short-term measurement in a single day. Comparatively, the proposed long-term monitoring could reduce the uncertainty and inaccuracy caused by various interference factors present in short-term monitoring sessions and thereby improve the performance of the auxiliary diagnosis of COPD.

This study has some limitations. The majority of the individuals in the case group were patients with COPD recruited from the hospital. Due to the characteristics of COPD, the proportion of old and lean men in the case group was higher, while individuals in the control group were normal primarily. Statistical comparisons revealed uneven distributions regarding age, sex, and BMI at baseline between the 2 groups, potentially introducing a selection bias. To address this issue, we accounted for these factors during data analysis by adjusting them prior to analysis. Consequently, we observed an improvement in the classification performance while observing minor impacts on the comparison of the classification effects for different features among our patients.

In this paper, we report the utilization of bed sensors for continuous recording of vital signs, including HR, respiration, and HRV during nighttime, thereby aiming to provide auxiliary diagnosis for COPD. Compared to single-time and single-day measurement, the long-term statistical features derived from continuous vital signs recordings demonstrate superior performance in the diagnosis of COPD. Moreover, the longer the timescale used for constructing the statistical features, the better is the classification performance for COPD diagnosis. Feature importance analysis showed that the statistical features of respiration and HRV contributed significantly more than the HR features in the COPD diagnosis. Specifically, RR, RRF, and HF exhibited the highest contribution to COPD diagnosis and hold important reference for home-assisted diagnosis of COPD. Additionally, this study determines the cutoff values for important features, which can provide a reference range for clinical COPD diagnosis. In principle, this design may also be applicable to other cardiovascular diseases such as heart failure, but the important features for disease diagnosis may be different.