Nighttime Cough Characteristics in COPD Patients

1. Introduction

Cough is an important symptom occurring in many respiratory diseases and associated with exacerbations, lung function decline and risk of death [1,2,3]. Increased cough is common during AE-COPD and a prodrome of increasing symptoms including cough can be seen for up to 2 weeks before AE-COPD. Thus questionnaires for monitoring the status of asthma (ACT: asthma control test) and COPD (CAT: COPD assessment test) include an assessment on cough. Calverley et al. [4] considered questionnaire reported symptoms like cough, breathlessness, chest tightness and night-time wakening and found a mean increase in symptom score of around 1 unit (scale: 0–4) at the time of exacerbation. This change of magnitude was found at a population level; in an individual patient it is unlikely to prove clinically useful.

Reasons why questionnaire symptoms do not perform as hoped may be various. There is the coarse quantisation of the response; increasing the level of detail (like a visual analogue scale: VAS) is uncommon as the required introspection will likely become an issue: posing questions that are hard to answer will typically come at the cost of decreased adherence. The interpretation of the labels associated with the questions are usually left to the user and this results in subjectivity. Furthermore, the moment that a patient tends to report a significant increase in the burden of a symptom (1 step on the questionnaire scale), is likely also the moment to seek advice meaning that no additional lead time would exist even if the questionnaire scale would be apt.

Out of the questions from the CAT (cough, sputum, chest tightness, night-time awakening), cough is the symptom with the highest concordance with AE-COPD, independent of COPD severity [4]. Cough is also a symptom that can be detected in an automated way as shown in many technical studies on prototype cough detection systems, for reviews see [5,6]. An automated system would mitigate issues associated with subjectivity, introspection and reporting. Sensors that are used are sound pressure sensors and/or accelerometers attached to the chest. Various systems are currently around such as wearable recorders [7], and wearable sensors, e.g., [8,9,10].

For patient comfort and adherence, an unobtrusive and hassle-free system is preferred. Since day- and night-time coughing is correlated, it would suffice to measure either. Measuring night-time has several advantages. It can be done truly hassle-free by a stationary system in the home of a patient. Also, the nighttime is a period with identical behaviour (sleeping) while the day-time is agenda driven and may involve all kind of activities inducing unexplained day-to-day variability or requiring extra channels to collect relevant contextual information. Finally, using a microphone as sensor in the sleeping quarters of a patient yields an off-body system and the night-time is usually also the most quiet period of a day making sensing and detection more robust.

For these reasons, the development of an unobtrusive stationary off-body microphone-based monitoring system was pioneered and has been reported in [11,12,13]. It holds the promise of constituting an element in a exacerbation prediction tool 1. A further validation of this system was undertaken in a double-blind clinical study. This paper addresses part of the outcomes of this study. It focuses on aspects of the cough monitor and cough behaviour (patient specific patterns) such as the variability of the cough density over the course of a monitoring session and the validation of an earlier reported empirical relation between mean and standard deviation in case of a stable chronic coughing patient.

The outline of the paper is standard. Firstly, the clinical study set-up is discussed, its devices are detailed and the data processing is described. Section 3 describes the findings and is followed by a discussion (Section 4).

2. Methods and Materials

2.1. Data Collection

We conducted a prospective longitudinal double-blind study of continual cough monitoring in COPD patients experienced in telemonitoring. To provide a reasonable chance of detecting AE-COPD, participants were studied for 12 weeks using domiciliary cough monitoring and asked to complete daily questionnaires each morning. If no exacerbation occurred in the first 12 weeks, the participants were asked to continue for at most another 3 months.

The study is double blind in the sense that i) cough data was analysed without any knowledge of the patient status or condition and ii) moderate and severe COPD exacerbations were identified retrospectively without access to the cough monitor data. The primary aim of the trial was to validate a (causal) alert mechanism for exacerbations based on cough trend data [12].

The current paper addresses secondary aims of the study: better understanding of cough behaviour and associated potential improvements of the alert mechanism. The main contributions of this paper are twofold. In the first place, we demonstrate patient variability in nighttime coughs patterns in the form of a case study and show its dependence on health status. Secondly, we validate the empirical relation between mean cough count and cough count variation [12]. Both have been done using only the data of the cough monitor, i.e., without access to the medical data, questionnaire data and identified COPD exacerbations.

The study was reviewed and approved by the North East-York Research Ethics Committee (REC Ref.: 21/YH/0203), the United Kingdom Health Research Authority and the Internal Committee Biomedical Experiments of Philips Research. Informed consent was obtained from all participants involved in the study. The patient target was set at $n=40$ and 2 patients failed to finish the trial. One withdrew almost immediately and the other one changed home leaving equipment behind. The data collection process ran from August 2022 to June 2024. Several issues delayed a speedy progress, including COVID, political changes affecting equipment, and organizational changes at Philips.

2.2. Cough Monitor

The used cough monitor is a stationary system placed in the sleeping quarters of the participant. It is a successor to systems used in earlier studies [11,13,15] targeting unobtrusive and hassle-free monitoring of cough. The prototype consisted of a single board computer (ASUS tinkerboard 2G) with a USB measurement microphone (Dayton IMM6) and a cellular dongle (Huawei E5330), see Figure 1. Feature extraction and type of cough classifier were as in earlier trials [11,13,16].

To prevent issues with power outage, the system was created such that it would start up automatically. At the installation, it is provided with a start and stop time that were set to 9PM and 9AM, respectively. It was assumed this covers the time in bed of all patients and such a monitoring period is called a session. To generate time stamps and to start and stop monitoring at the correct moments, timing information is required. This information is drawn from the cloud. In view of the long period over which the monitor is intended to operate in a stand-alone mode, the system is shut down and booted every day which is an effective means to prevent memory leakage. A consequence is that if the connection to the cloud is not established, the timing information is not available and the monitor will not start a session.

For the first 8 participants, the dongle contained an IoT SIM card (TruPhone) operating over the 2G network. Later participants used a 5G SIM card (Vodafone, UK). Data were transferred to the cloud (AWS) from where data were downloaded to a proprietary system for analysis. To ensure privacy, only time stamps and audio features are transmitted at moments where the sound scape changes. We refer to this as acoustic events or transients. For a limited number of these events, also a one second audio snippet is recorded and transmitted. The limited use of short snippets prevents listening in to any conversation while enabling checking for audio issues with the devices or its set-up (e.g., a ticking clock next to the monitor can be detrimental) and creating a personalised classifier.

At the server side, the data is collected and a classifier is trained based on the first monitoring days as described in [12,16]. In these previous trials, solid personalized classifiers were attained when trained with around 200 coughs. With such target in mind, the number of days that requires annotation (based on the snippet-feature combination) is variable because the number of acoustic events is highly dependent on the patient [17].

A high-level schematic overview of the processing is given in Figure 2. On the left-hand side there is the processing in the patient’s home. It consists of the audio processing based on an audio processing library A. The audio processing captures the audio and extracts features and snippets. This part is a mature software component and has been deployed successfully in several trials. The audio processor is embedded in a scheduler with interface and the communication part including encryption. This has been developed for experimental purposes only. On the right-hand side, there is a data receiving and development unit D where data is received, decrypted and stored. Furthermore, there is an annotation and classifier training system to develop the personalized classifier based on the snippets and the associated subset of audio features. The central cough monitor processing is in unit C containing the trained (personalized) classifier that classifies each received audio feature and creates cough counts based on classifier outputs and associated timestamps. The counts form a time series and are input to the alert mechanism.

2.3. Cough Classifier

The cough monitor is closely tied to the operation of the human hearing system as this defines the groundtruth of what a cough is [18]. This starts with the features that are extraction from the audio in the home of the patients. These incorporate a variant of Mel frequency cepstral coefficients (MFCC). Next to these, energy levels before and at the acoustic event and the (local) density of events are extracted. The deployed classifier is a personalised classifier. A generic cough classifier would essentially require a one-class classifier, i.e., only using information on the sound of a cough without considering the environment in which the cough is being uttered. The neglect of using environmental information presumably leads to worse detection performance and is certainly not in line with our knowledge on human perceptual processing which actively uses contextual information [19].

For annotation, an audio-visual interface was used. The audio is essential to annotate if the features belong to a cough, and not to an environmental sound or a vocal sound of the patient which is not a cough (like: throat clearance, sigh, moan, sneeze, burp, speech, laughter). The signal waveform is provided where a typical cough consists of three phases: an explosive part, an intermediate stage and a voiced phase. Not only is the visualization helpful because of this specific pattern, it is also instrumental to obtain a single identification of the cough over time and not multiple. The explosive phase of the cough is defined as the target as this is the acoustic response to the opening of the vocal chords after pressure build-up: a requirement for a cough by definition. It is also the most easily identified part of the cough: the intermediate and voiced phase are not always clearly present in sound and visuals. To obtain a unique signature for a cough, the position of the feature extraction is shown in the graph and the annotator checks if this corresponds to the explosive phase in case the sound resembles a cough.

For each patient, a classifier was trained using an extreme gradient boosted decision tree classifier (XGBoost) where default settings were used in the training. After training the classifier, it is executed on the entire feature set using a relatively high threshold $T=0.9$. The outcome is a list of time stamps of detected coughs. To create a profile of the cough density over the monitoring periods, the following procedure was used. One hour periods were selected and shifted over the monitoring period with a 15 minute update. The detected coughs in that period were counted. Data from the same time interval in all sessions were collected and treated as a random variable: mean, standard deviation, median and quartiles were calculated. Also the $\alpha$-trimmed mean was calculated with $\alpha =0.25$.

Note that we prefer to speak about cough densities rather than cough frequencies. Frequency is term connected to regularly occurring events, periodicity and quasi-stationarity. Since the number of coughs within a given timespan is far from periodic or equally spaced over time, the term cough density is preferred.

2.4. Alert Mechanism

The alert mechanism defined in [12] uses a threshold that is based on a period of low cough counts not covering any exacerbation days. This method is not useable in the present situation due to the double-blind character of the study. To solve this, an adaptive threshold was developed which is also more appealing for long-term use in real-life to allow for, e.g., aging and seasonal effects. The method works as follows. The time series is mapped to the B-scale to be able to work with an identical outlier setting for all patients [12]. Data of the latest L sessions is input to determine a baseline. The oldest $L-D$ sessions are used to define the current threshold, D is the delay or dead zone. The median is taken over the $L-D$ observations and this median is input to a first-order IIR filter with pole $p_t$. An offset O is added to define the dynamic threshold level T.

An alert is raised based on the smoothed cough counts on the B-scale using a first-order IIR filter with pole p. If this smoothed cough count exceeds the threshold at least twice in the latest three sessions, an alert is raised. The settings for the mechanism are given in Table 1. We note that the parameter $\alpha$, $\beta$, $L-D$, p correspond exactly to those in [12]. The parameters associated with the adaptive mechanism (D, $p_t$ and O) were set based on the cough data only, i.e., without knowledge of diagnosed exacerbations. To illustrate the behaviour, Figure 3 shows three examples of cough count (B-scale), smoothed cough count and constructed dynamic threshold T. All three examples show alert days as indicated by the red circles. The red circles are on the black smoothed line, as this is the input to the alert mechanism. Clinical aspects of the study and validation of this particular alert mechanism will be covered elsewhere. In the remainder, we will concentrate on the more general aspects of cough counting: cough density profiles and validation of the earlier proposed cough scale (B-scale used in Figure 3).

3. Results

3.1. Patient Characteristics

From the 40 enrolled patients, one withdrew and another one didn’t finish the trial. The characteristics of the remaining 38 patients are given in Table 2. Compared to our earlier studies [11,13], less severe patients were involved in this trial.

Not all patients provided useable cough data. This is a consequence of the set up of the study (retrospective without interventional mechanism), behaviour of patients (absence (e.g., holidays), moving) and system issues (disconnected power, interrupted data connection). From the 38 patients, 4 were immediately excluded from further analysis. There were 3 patients (P035, P038, P039) with severe connectivity issues and patient P012 had connectivity and RF interference issues. It made the data too scarce to run the alert mechanism and, in most cases, there was even too few data to train a classifier.

In Figure 4 the distribution of duration of the monitoring (number of days from first and last received data), the number of days within that period that data was received, and the number of missing days (i.e., the difference between duration and monitoring sessions) is shown in a boxplot. The median of missing days is about 10, on a median of 84 days. Though the patient can turn off the power, it is presumed that the main cause for missing days is that the system was unable to schedule its daily monitoring session due to a not guaranteed mobile connection.

3.2. Annotation and Classifier Performance

For each patient, a trained classifier was created by annotating the snippets that were transmitted. During annotation it was found that the monitor had issues with RF interference which appeared in the audio as spike trains. Next to patient P012 (already ruled out because of connectivity issues), this interference made the audio snippets (and audio features) for P020 questionable. With other patients, RF interference occurred much more rarely and is expected to have no or little impact on the results.

It was also observed that for one patient (P018) the dominant respiratory acoustic event was not a cough but what sounded as a combination of a cough with a throat clearance. The sound and waveform patterns observed in annotation thus did not match to this case as the annotator was instructed to identify the explosive part in a normal three-stage acoustic event [20] exhibiting intermediate and voiced phases after the explosive phase. No attempt was made to define an annotation process suited to this singular case. The data of this patient were excluded from further analysis. For more details on the annotation, see Appendix A.

Settting P018 and P020 aside, the number of patients with cough data becomes 32. From these patients, 12 had a partner and 20 not. From the 12 partners, four were reported as coughers, four were not, and for the other four cases this is unknown.

The experience from earlier experiments indicated that solid classifiers were obtained when around 200 coughs where used in the training. The number of snippets that were annotated as coughs ranged from 178 to 278 over the patients. The number of annotated non-coughs ranged from 256 to 8779, clearly marking the huge range of cough prevalence in the annotated snippet set.

Another observation that was made during data screening and annotation was that it is questionable whether patient P027 was actually occupying the sleeping quiarters after an initial period of about 8 weeks: both acoustic events and coughs dropped dramatically for almost all days. With other patients, there are also questions around presence having days with abnormal low number of acoustic events and coughs. No mechanism was constructed to rule out these days.

Figure 5 gives the boxplots representing the sensitivity, specificity, accuracy and positive predictive value as determined in the training with a threshold setting of 0.9 for cough classification. The fraction of acoustic events classified as coughs (positive rate: PR) on the entire data set is also included. These statistics highlight our basic notions in the design of an alert mechanism: a high PPV is essential if the cough counts are to act as its substrate. In view of the low cough prevalence [17] and robustness for exposure to untrained acoustic events [20], a high threshold (0.9) was set for the cough classifier. This ensures a high PPV by favouring specificity over sensitivity. The sensitivity ranges from 0.26 to 0.85, the sensitivity is high (all except 2 above 0.975), the full PPV range stretches from 0.8 to 0.98 and the accuracy is in the range 0.86 to 0.98 (except for 2 patients).

In Figure 6, the prevalence of coughs as determined during annotation is compared to the positive rate. We note that the positive rate (detected coughs) is low: for 75% of the patients below 10% of the acoustic events. There is one clear outlier where a quiet acoustic environment is combined with a large amount of coughs resulting in a positive rate of nearly 0.4. The positive rate is smaller than the annotation prevalence: the median is 6% while the median cough annotation prevalence is 13%. This is mostly due to a low sensitivity induced by the preferred high threshold setting (median: 0.54). Other factors that contribute to the difference are that prevalence determined from annotation covers only a small part of full data while the positive rate is determined over all acoustic events and that the snippet selection is biased slightly towards the louder events in order to facilitate annotation.

3.3. Cough Counts

Each detected cough carries a time stamp. This enables statistics of these cough counts where we consider mean and variance of the counts over a monitor session, and cough density profiles over the monitoring period (9PM to 9AM). The amount of coughing depends heavily on the participant.

For each patient, it was checked if an alert was raised by our prototype alert mechanism: in that case the cough count is not stable over the days. As the start of the alert track is delayed because the system needs time to collect baseline characteristics, the trend plots may reveal considerable variation in the beginning without an alert being generated. Therefore, a visual inspection of the data was undertaken (by the first author) to verify if the first part of the cough data did not have an upward or downward trend. In case of absence of alert and absence of trend, the patient cough data is labelled as stable. An exception to this rule was P021 showing an enormous spread in cough counts in the middle of the observation period and this was labelled as unstable as well. Only 6 out of the 32 cough trends were labelled as stable, about half (17 out of 32) had at least one alert. An overview of this data is in Appendix B.

In case of stable cough counts over the days, the averaging over the day creates a cough density profile presumed to be characteristic for that particular patient during normal respiratory condition. Figure 7 shows the profiles of participants with a stable cough count. In the cough density plots, the median density profiles are provided (expressed in coughs per hour) together with the quartile ranges and the $\alpha$-trimmed mean with $\alpha =0.25$. We observe that these curves are not uniform but exhibit peaks at different time instances. Some graphs have peaks in the morning hours, some in the evening, the latter presumably corresponding to the time the patient goes to sleep. Some patients hardly cough at all in the sleeping hours, yet some have a steady floor in the cough density. The diversity of patterns agrees well with the observations in [12]. For completeness, the set of all cough density patterns is provided in Appendix B.

In [12] a linear relation was found between mean and standard deviation of cough counts of patients that had no medical issues during the monitoring period. To validate this empirical relation for the present cough data, the current data and the earlier found model are jointly plotted in Figure 8. The data of the assumed stable patients follow the earlier reported linear trend while the data of the other patients are located dominantly above this line.

The linear relationship suggested a scale transformation to create a cough count scale (called the B-scale) where a unit step has the same meaning anywhere on the scale [12]. To show the effect of the scale, we provide the cough count quartile ranges as a function of the median in Figure 9. The dashed line representing the median plus a fixed offset runs nicely congruent with the third quartile, especially for the patients with a stable cough count. For the non-stable patients, the bars sometimes extend above the line, and sometimes agree with it. This is assumed to be caused by the use of robust statistics implying that if the data is mostly stable, the used metrics will follow that of the stable patients. Note that there are two patients (P017, P027) with low median cough count and bars below the blue line. This may be caused by prolonged absence of the patient.

Like said, creating profiles by averaging over days for patients having respiratory issues during the observation period may not provide a meaningful density profile. To test this hypothesis, the data of each patient was split into two equal parts: sessions with highest and lowest counts. A profile was created for these two data sets. To compare these profiles visually, they were normed such that the mean density equals 1 cough/h. In Figure 10, the profiles are shown for patients having stable cough counts. Our interpretation of the density profiles are qualitatively rather similar. In contrast, four selected profiles are shown from the patients with an unstable cough trend in Figure 11. Here we observe that clear shifts in the densities occur. The top-right graph reveals that for days with a high cough count an additional peak occurs around midnight. The bottom-right shows a heightened peak in the morning hours. The bottom-left completely changes its character from coughing dominantly in the evening (at around 10PM) to coughing all night long. The top-left graph has a change in timing: the evening peak occurs earlier and an even clearer shift in the two peaks in the morning. This would be compatible with a patient going to bed earlier and sleeping less long in case of increased cough count and thus presumed respiratory issue.

4. Discussion

For COPD patients with an appreciable nighttime cough count a cough density profile can be constructed. These density functions tend to be rather uniform (flat), have one peak (the morning or evening cough) or double-peaked (morning and evening cough). The results confirm patterns shown in an earlier study [12] and demonstrate a large variation over the population.

The deployed method of creating density profiles uses averaging over all days. To prevent bias, we considered only patients with data where the alert mechanism wasn’t triggered and visual inspection didn’t contradict the assumption of similar behaviour at different days. Results from all other patients are presented in Appendix B. Even though robust statistical measures were used, they may be biased and therefore should preferably not be used as a description of the patient’s normal cough density. Nevertheless, also these curves show pronounced patient-specific peaks.

Note that for some patients the density profile is relatively large or even maximum at/near start or stop of the monitoring period. This is an undesired monitoring setting as it means that the count can be largely influenced by slight changes in habits of the patient. It may be associated with acoustically open spaces or deviant sleeping hours.

For stable patients, an empirical relation has been proposed between mean and standard deviation of the nighttime cough count [12] (for unstable patients, this is not meaningful: mean and standard deviation estimates are rather sensitive to outliers). This relation was verified by considering data from cough counts where the cough count trends were identified as stable.

The earlier introduced B-scale for cough count was applied to the present data. It confirms that this scale enables defining a patient-independent offset which, when added to a robust statistic of the mean (baseline), defines a cut-off level identifying outliers.

Lastly we considered the dependence of the cough count per night on the density. This confimed that the shapes of the density profiles were only slightly dependent on the strength for stable patients. For unstable patients, the shape of the density profiles may vary considerably with the total amount of coughs per night, and thus presumably with the patient’s respiratory health.

5. Conclusions

We have analyzed data from COPD patients ($N=32$) obtained by a stationary nighttime cough monitor. Cough density profiles were created demonstrating a large degree of variation of cough behaviour across patients. The empirical relation between mean and standard deviation of the nightime cough count was confirmed. The effect of mapping the cough scale for uniform outlier detection was illustrated. These findings deepen our insights and may serve as guidelines in the design for clinical applications like alert mechanisms for chronic coughers. The data also suggest that not only the intensity (amount of coughs in a day or night) but also the shape of cough density profiles may vary substantially with the patient’s respiratory state. Validation of the alert mechanism incorporating these insights will be the next step.