Domiciliary Cough Monitoring for the Prediction of COPD Exacerbations

Chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide [1]. AE-COPD negatively impact patients’ quality of life and are a leading cause of health service utilisation, adding to the global challenge [2]. Attempts have been made to identify impending AE-COPD with the objective of early intervention to improve outcomes. Remote monitoring of patients’ symptoms and physiological parameters (telemonitoring) may be used for this purpose with ‘alerts’ conveyed to a clinician when worsening symptoms/signs are detected.

However, there is no compelling evidence that existing telemonitoring strategies can detect meaningful change in clinical condition. Randomised controlled trials of telemonitoring in COPD have failed to demonstrate a reduction in hospitalisations [3‐6]. Indeed, an increase in the number of community treated exacerbations and increased health service utilisation have been reported in patients receiving telehealth [3, 6]. One problem with existing telemonitoring approaches is the frequency of ‘false alerts’ i.e. those not representing an impending exacerbation that will require treatment. In a well-conducted COPD telemonitoring trial, community teams dealt with 2441 alerts from 97 participants over a 1-year period (25 alerts per patient per year) [3] thus leading to high resource utilisation.

Cough and sputum production are common in COPD [7, 8] and are associated with exacerbations, lung function decline and risk of death [9]. 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 [10]. Calverley et al. demonstrated a mean increase in symptom score of around 1 unit (scale: 0–4) at the time of exacerbation across a range of symptoms including cough, breathlessness, chest tightness and night-time wakening. We have previously demonstrated a similar change in cough score measured using a 5-point Likert scale during AE-COPD convalescence [11]. Whilst change of this magnitude can be detected at a population level; in an individual patient it is unlikely to prove clinically useful. Cough is the symptom with the highest concordance with AE-COPD, independent of COPD severity [10]. It is therefore a most promising candidate for predicting AE-COPD.

We have previously described a system that analyses acoustic signal from microphones placed in the domestic environment to identify and record cough events [12]. This system can detect individual long-term cough count trends and demonstrate a reduction in cough frequency during AE-COPD convalescence [11]. In the current study, we monitored cough frequency in exacerbation prone COPD patients over 90 days and assessed the predictive performance of an automated AE-COPD alert system based on cough count and compare it with an established telehealth questionnaire.

We report the first longitudinal study of continual daily cough monitoring in COPD patients specifically designed to assess cough frequency in the AE-COPD prodrome. Observational studies have shown that patients report worsening cough for several days before starting treatment [10]. For the first time, we confirm an increase in objectively measured cough frequency occurs using a cough monitoring system that analyses ambient sound with a static microphone located in the persons sleeping quarters.

The cough monitoring alert system detected increased cough frequency a mean of 3.4 days before approximately half of AE-COPD. This was achieved with fewer false alerts than a paradigm based on questionnaires. The majority of questionnaire-based alerts in our predictive window of 8-days prior to an AE-COPD can be explained as alerts occurring by chance. Indeed the questionnaire-based system generated one false alert every 10 days, so the chance of an alert occurring in the window prior to an AE-COPD can often be explained by random association. Conversely, there is a high probability that an alert generated by cough monitoring signifies a genuine change in clinical state associated with impending AE-COPD. We detected 14 alerts pre-exacerbation with only 13 false alerts during the study. The false alert rate was one per 100 days. The average lead time of 3.4 days suggests that cough monitoring-based alerts could prompt a meaningful early clinical response.

Cough monitoring in this study was limited to the room where participants slept. It was considered that there may be inter and intra-participant variation in the amount of time spent in this room. However, analysing data based on patient proximity to the microphone or during set night time hours did not alter the overall classifier performance [13].

Cough monitoring-based alerts are individualised by establishing each patient’s baseline cough count and identifying deviation from it. The daily symptom questionnaire attempts to do the same by asking patients to report a change from their normal state. However, it was clear that some patients continually answered ‘yes’ to questions such as ‘Are you more breathless than usual today?’. In order to identify if establishing an individualised baseline was the reason for favourable cough monitoring-based alert performance, we subsequently analysed the daily symptom questionnaire data using the methodology used for cough monitor alert generation i.e. looking for an increase in symptom score over an individualised baseline. The questionnaire success rate dropped to 0.41 and the false alarm rate to 0.015. This performance remained numerically inferior to the cough monitoring system, but most notable was the reduction in lead time to an average of 1 day giving little predictive value.

Participants in this study have severe COPD with a high symptom burden and experienced frequent exacerbations. This provided an opportunity to observe cough frequency during AE-COPD but made establishing a baseline during clinical stability challenging. Some participants were excluded from analysis because they did not have adequate time between AE-COPD to allow assessment of a baseline. The studied population is a challenging cohort due to their clinical instability. It is possible that performance will improve further in individuals with lower exacerbation frequency.

In this study, we define AE-COPD according to treatment and healthcare utilisation. This means that mild AE-COPD, defined as symptom worsening for more than two consecutive days but not leading to treatment with systemic steroids or antibiotics [15], are not included in our analysis. Although these events are often reported in studies where symptom diaries are completed, their importance in the context of clinical interventions generated through telemonitoring is not clear. One could argue that questionnaire-based alerts that did not precede a moderate or severe AE-COPD were not false-alerts, but alerted clinicians to a mild AE-COPD. However, these events resolved without the need for additional interventions or health care utilisation. For the purpose of telemonitoring, a system that differentiates mild AE-COPD that will resolve spontaneously from impending moderate-severe AE-COPD that may benefit from early treatment provides optimal healthcare resource use.

In addition to the favourable performance of cough monitoring-based exacerbation prediction; auditory cough monitoring is a passive process that requires no direct user input. A drawback of conventional telemonitoring is the reliance on patients to correctly interpret and respond to questionnaires and input correctly recorded physiological measurements. We have previously reported poor adherence with home spirometry monitoring [12] and in this study, we observed a number of participants that repetitively reported their symptoms as having ‘increased’ in severity, limiting the utility of questionnaire-based reporting.

In previous studies, we have observed reluctance from patients and their carers to engage with a system which continuously monitors all sounds within the home environment [12]. There are a number of ethical and data protection issues around this method of observation. Early concerns about the acceptability of having audio recording equipment within the home appear to have been adequately addressed by limiting recorded sound to 1 s excerpts; ensuring no possibility of conversational speech being recorded.

This study has a number of limitations. Firstly, it was a derivation cohort used to develop the alert system based on the results of objective cough monitoring. System performance will need validation in a different cohort of patients; ideally, a heterogeneous group of COPD patients to determine the generalisability of the method. The static domiciliary system used may well be applicable to a group of severe, mainly house-bound individuals and further work will be required to determine whether performance is maintained in less sedentary patients. The impact of cough among other family members may be a greater factor when a less severe cohort of patients is studied. Although we very successfully reduced the number of false alerts (increasing specificity), the sensitivity of the system may be less than optimal. Finally, the hardware used in this study was ‘off-the-shelf’ and as such was prone to remedial technical failure in two out of the 30 systems deployed. Reliability may be improved by bespoke technology.

An increased cough count can be detected in the prodrome of around half AE-COPD using a cough monitoring system that analyses ambient sound in patients’ sleeping quarters. This can be achieved with a low false alert rate, meaning that each alert can be considered meaningful by clinicians. This contrasts with questionnaire-based alerts, which are difficult to interpret due to the high false positive rate.

The cough monitoring-based alert system described in this study warrants further assessment in a validation cohort. If alert performance is confirmed, cough monitoring will offer the first objective, passive monitoring strategy for COPD patients that is capable of generating meaningful clinical alerts of impending AE-COPD.