Abstract
Introduction Cough is a hallmark symptom of multiple respiratory conditions. Technological advances have made it possible to monitor cough using smartphone applications continuously and objectively. However, these technologies have not yet been evaluated for their feasibility in real-world research conditions.
Methods We conducted 14-day continuous cough monitoring of presumptive TB patients enrolled as outpatients in a multi-center prospective tuberculosis diagnostics trial in Uganda, South Africa, Vietnam, the Philippines and India. Participants were given a smartphone with the Hyfe Research application installed, a carrier sling, and training on how to use both the phone and the cough application. We evaluated how often and how long participants recorded, and when coughs were detected. We also compared multiple cough frequency metrics for their performance in detecting changes in cough frequency.
Results Of 693 participants who consented to continuous cough monitoring for 14 days, 568 (82.0%) completed at least one recording session successfully. Among these 568 participants, 421 (74.1%) recorded on at least 10 of the 14 days in the recording period for a median of 23.9 hours per day. We detected coughs in a diurnal pattern with a median of 1 cough per hour detected between 6am and 2pm local time and a median of 0 coughs for all other hours. We found median coughs per hour to be the cough frequency metric most robust to outlier values while still detecting changes in cough frequency during the recording period. Implementation challenges include the visibility of cough monitoring devices, nighttime recording and concurrent charging of devices.
Conclusions Continuous cough monitoring using a smartphone application is feasible in low- and middle-income country settings.
Introduction
Cough is a hallmark symptom of multiple respiratory condition with a substantial global burden.1 Further, for conditions like tuberculosis (TB), cough evolves throughout the course of the disease and may contain useful clinical information.2 Thus, accurately measuring the burden and characteristics of cough is of significant clinical and research interest. Traditionally, cough is measured subjectively through symptom screening or cough severity assessment tools. By design, these assessments can only be performed sporadically and rely on patient self-report.
The development of smartphone-based artificial intelligence applications (apps) which can detect and analyze cough newly allow for continuous and objective cough monitoring. 3 Through such applications, it is conceivably possible to develop diagnostic and treatment monitoring tools based on cough sounds and patterns. There are currently several validated continuous cough monitoring technologies available on the market which use smartphone apps4–6 as well as continuous cough monitoring products which use a free-standing recording device.7
Continuous cough monitoring tools allow researchers to objectively measure the timing and frequency of coughs. This is distinct from efforts to develop cough classification algorithms for diagnostic processes.8,9 Monitoring cough count and timing patterns allows researchers to evaluate how these features may change over time and in response to treatment. Many cough monitoring tools also record participant coughs as sound files which allows for research into how specific acoustic features of cough associate with patient clinical characteristics and outcomes.
While these new tools offer promising objective insight into patient cough patterns, based on experience with several ongoing studies, implementation challenges remain. It has not previously been established whether participants are willing to carry phones or recording devices during their daily life. Further, the devices have not been extensively tested for their ability to identify and record cough in resource-limited settings. There are additional challenges to conducting cough monitoring in low-resource settings where participants may not have stable access to internet and electricity.
The diagnostic performance of continuous cough monitoring with the Hyfe Research App will be reported at the conclusion of the trial. In this report, we describe the operational and implementation experiences with 14-day continuous cough monitoring of presumptive TB patients as part of the ongoing international multi-center TB diagnostics trial in five high TB-burden low- and middle-income countries..
Methods
Setting
The Rapid Research in Diagnostics Development for TB Network (R2D2 TB Network) study10 evaluates novel TB diagnostics in various stages of development among patients with presumptive TB in five low- and middle-income countries: Uganda, South Africa, Vietnam, the Philippines and India.
Study Population
Eligible participants are 18 years or older and have new or worsening cough for at least two weeks. Participants are excluded if they have completed latent or active TB treatment in last 12 months, taken medication with anti-mycobacterial activity in last 2 weeks (isoniazid, rifampicin, rifapentine, fluoroquinolones), reside >20 km from a study site, are unwilling to return for follow-up, or do not provide informed consent.
Procedures
Demographic and clinical data
At enrollment in the R2D2 TB Network study, participants undergo a baseline clinical exam which includes sputum collection for TB testing (Xpert MTB/RIF Ultra and liquid culture), blood collection, and a digital chest radiograph. Study staff also record participants’ age, gender, height, weight, smoking status and duration of cough. HIV diagnosis is made either based on participant self-report of a positive HIV diagnosis or a positive test result. Diabetes diagnosis is made either based on participant self-report of a diabetes diagnosis or an HbA1C above 6.5%.
Cough data monitoring
Continuous cough monitoring is offered to all participants enrolled in the R2D2 trial using the Hyfe Research app.11 When in recording mode, the Hyfe app continuously listens for explosive sounds (i.e., sudden increases in sound above background levels). A validated machine learning algorithm12 classifies explosive sounds as coughs and recordings of these cough sounds are saved. Recordings do not exceed 0.5 seconds in length to protect participant privacy. Cough data, including sound files, are uploaded to a secure server accessible to study staff. The app can identify and recording cough sounds without requiring an active internet connection. The Hyfe app stores data until the phone is connected to the internet at which point the Hyfe app will upload data to the server. The app maintains the data even if the off or the battery dies.
Participants who provide consent are asked to use the Hyfe app to record cough sounds for 14 days. Participants are loaned a smartphone and a charger if they do not have one or prefer not to use a personal device. The phone model(s) used by each site are given in Table 1.
List of phone models used for continuous cough collection by site
Study smartphones are restricted such that they cannot download additional applications, use the camera or browse the internet. Study staff train participants in the use and maintenance of the loaned smartphone. Participants are provided with a carrier bag designed to be worn around the neck (Figure 1) during the day to facilitate recording of cough sounds.
Picture of carrier bag used to keep phone near participants’ faces during recording.
Participants are instructed to insert the phone into the carrier such that the phone’s microphone is directed towards their face. Participants are also instructed to position the phone close to their head when sleeping and study staff recommend charging the phone during sleep. Participants are asked to record for as many hours as possible but are shown how to pause and restart Hyfe app recordings. Full standard operating procedures are provided in supplemental materials. If a participant completed at least one recording session with a duration longer than 5 minutes and returned the phone to have the Hyfe data uploaded, they were considered to have completed longitudinal recording.
Data analysis
Data on recording periods and detected coughs were downloaded from the study server. These data were linked to participant enrollment data collected and stored on a secure REDCap13 server. This work reflects continuous cough data collected and uploaded between May 4th, 2021 and April 25th, 2022. All participant recording sessions were censored to 14 days.
In order to characterize participant recording behavior and timing of detected coughs, summary statistics were calculated. To account for the non-normal distribution of these parameters, medians and interquartile ranges (IQR) were used.
Several metrics were considered for measuring cough frequency. The desired metric would robustly summarize cough frequency for a given day in the recording period and show any trends over the 14 days in the recording period. We compared the total number of coughs over the total recording time (coughs per hour, CPH), the maximum number of coughs in any recorded hour (maximum cough per hour, maxCPH), and the median number of coughs in any recorded hour (median cough per hour, medCPH). Each metric was calculated for each day that a participant had recorded.
Summary statistics of recording time and cough detection, visualizations and cough frequency metrics were analyzed in R version 4.2.0.14
Ethics statement
Ethical approval for this study was obtained from institutional review boards (IRB) in the US and in each study site. In the US, approval was obtained from the University of California San Francisco IRB (# 20-32670). In Vietnam, approval was obtained from the Ministry of Health Ethical Committee for National Biological Medical Research (94/CN-HÐÐÐ), the National Lung Hospital Ethical Committee for Biological Medical Research (566/2020/NCKH) and the Hanoi Department of Health, Hanoi Lung Hospital Science and Technology Initiative Committee (22/BVPHN). In India, approval was obtained from Christian Medical College IRB (13256). In South Africa, approval was obtained from Stellenbosch University Health Research Ethics Committee (17047). In Uganda, approval was obtained from Makerere University, College of Health Sciences, School of Medicine, Research Ethics Committee (2020-182). In the Philippines, approval was obtained from De La Salle Health Sciences Institute Independent Ethics Committee (2020-33-02-A).
Funding statement
Research reported in this publication was supported by the National Institute of Allergy and Infectious Diseases (NIAID) of the National Institutes of Health (NIH) under Award Number 5U01AI152087. SH is supported by a Canadian Institute of Health Research Postdoctoral Fellowship, a UCSF Division of Pulmonary and Critical Medicine T32 Fellowship, and a UCSF Center for Tuberculosis and NIH/NIAID R25AI147375 award.
Results
Of 1436 participants enrolled in the R2D2 TB Network study, 693 (48.2%) consented to continuous cough monitoring for 14 days. Of these, 568 (82.0%) completed at least one recording session, returned their phone, and had their data uploaded from the smartphone to the study server. The median age of these participants was 40 years (interquartile range [IQR] 28.5 – 54.0) and 179 (31.5%) were female (Table 2). The median BMI was 22.2 (IQR 18.9 – 24.9) and 123 (12.7%) participants had smoked within the past week. 76 (13.4%) participants were living with diabetes, and 68 (12.0%) were living with HIV. The median duration of cough before care-seeking was 28.0 days (IQR 17.0 – 60.0).
Demographic summary of participants
Operational considerations for continuous cough collection
Participants who declined to participate in continuous cough recording were significantly older (median 44.0 years vs. 40.0 years), more likely to be male (61.6% vs. 47.4%) and had significantly lower BMI (median 20.8 vs. 22.2) (Table 2). Participants who declined to participate were also more likely to have smoked in the past 7 days (88.5% vs. 78.2%) and have a longer duration of cough (median 30.0 days vs. 28.0 days). There was no significant difference in prevalence of diabetes or HIV across the two groups.
We did not collect data on reasons for declining to participate in continuous cough monitoring. However, study staff reported that some older participants cited discomfort with smartphones and using mobile applications. Other participants, especially younger females, did not wish to explain to their family or spouse why they temporarily had a second phone. Finally, some participants said that they did not feel they could wear the phone in its carrier sling at their place of work or school. In Uganda specifically, participants cited fear of losing the phone and inconsistent access to electricity for charging as reasons to decline consent.
In order to monitor data quality, summary reports were generated for each site biweekly describing participants’ recording behavior and cough detection (code provided in supplemental materials). Through these reports some sites discovered that the number of recording hours for participants would decrease in the middle of the recording period. Support calls to participants to remind them to continue recording were helpful in increasing the number of recording hours for participants.
In South Africa, one participant had their phone stolen. To protect participants, longitudinal recording was stopped at this site. Discussions are ongoing about how phones could be carried less visibly while still keeping the phone microphone close to participants’ mouths.
Participant recording adherence
Overall, 74.1% of participants had recording sessions on at least 10 days, though only 45.4% of participants had recording session on all 14 days (Figure 2). This varied by site, with South Africa, Uganda and India having generally fewer recording days per participant. The median number of recording days per participant across all sites was 13.0 (IQR 9.0 – 14.0).
Proportion of participants recording at least X days
In aggregate, participants recorded consistently throughout the day (Figure 3). Recording behavior was similar across each country. The median number of recording hours per day across all sites was 23.9 (IQR 13.6 – 24.0, Table 3).
Histogram of recording time by site. Time has been converted to local time zone for each site. Count refers to the number of participants recording in a given hour across the recording period.
Median number of recording hours per day
Cough detection
Despite recording evenly across the day and night, the majority of coughs were detected during the day. The median number of coughs detected (among days where participants recorded) was 1 cough per hour between 6 am and 2 pm (Figure 4). During all other hours the median number of coughs detected was 0. Each country also demonstrated a unique pattern of cough timing (Figure 5).
Plot of median coughs detected per hour. Black line indicates median value. Grey region reflects interquartile range. Time has been converted to local time zone for each site.
Density plot of timing of cough detection by study site. Time has been converted to local time zone for each site.
Comparison of cough frequency metrics
Coughs per hour (CPH) and maximum coughs per hour (maxCPH) were both susceptible to extreme values likely driven by participant coughing fits (Figure 6, A and B). CPH is also sensitive to short daily recording durations which, as the denominator in this metric, can cause artificially high CPH values. The final metric, median coughs per hour (medCPH), was more robust to outliers with a maximum observed value of 123 (Figure 6, C). MedCPH also demonstrated the expected trend in decreasing cough frequency across the recording period (Figure 6, D). For these reasons, medCPH is our preferred cough frequency metric.
Comparison of cough frequency metrics. Figures show a boxplot of each metric for each day in the recording period. A shows the total coughs per hour metric, B the maximum coughs per hour metric, and C the median coughs per hour metric. A, B and C are presented with the same y-axis to facilitate comparison of outlier values. D shows a close view of the median coughs per hour metric to demonstrate the trend in the metric across the recording period.
Discussion
In this study, we have shown that it is feasible to conduct high-quality continuous cough monitoring for a 14-day period using the Hyfe Research app at multiple international sites among presumptive TB patients. Participants consistently recorded throughout the day with a median recording duration of 23.9 hours. Nearly 75% of patients recorded on at least 10 of the 14 days in the recording period. Most coughs were detected during the daytime and each country demonstrated a unique distribution of cough throughout the day. Among the three evaluated metrics, we found median coughs per hour to have the best performance for measuring daily cough frequency.
We were able to successfully coach participants in how to use the Hyfe app. Further, participants recorded consistently, and we detected coughs throughout our observation period.
We were also able to provide all patients with a carrying case that held the phone in the ideal location to detect and record coughs. We have verified that the sound files are of sufficient quality to extract acoustic features for future analyses.
However, there are several challenges which need to be addressed to improve this innovative means of data collection. We did not detect coughs during the night. This may be that patients are genuinely coughing less frequently during deep sleep, or it may mean that the phones were not located close enough to where participants were sleeping. Participants may have been limited in where they could position the phone during the night because they needed to be able to charge the phones. We also encountered issues with the visibility and convenience of our phone carriers. Having the phone visible in South Africa unfortunately caused one participant to be robbed. Some participants cited the visibility of the phone as a contributing reason for declining to participate in continuous cough monitoring. A more subtle recording device such as a lapel microphone or wrist band may address these concerns.
Further, there were some significant demographic differences between participants who did and did not consent to continuous cough monitoring. Older participants, males, those who smoked and those with longer cough durations were more likely to decline to participate. This could negatively impact the generalizability of analyses of cough patterns. In future studies, additional recruitment procedures like more extensive training with smartphones and the Hyfe app should be considered to address technology literacy concerns.
Based on our experiences, we believe there are several key aspects to ensuring feasible, high-quality continuous cough monitoring. Participants need to be trained how to use smartphones and the recording app; extra care should be taken with patients who feel uncomfortable with smartphone technology. Generally, participants were willing to use carrier slings which maintained optimal position for cough detection and recording. Participant recording behavior improves with continued support from study staff in the form of check-ins. Consistently checking data quality metrics also allows study teams to quickly respond to issues in data collection. In future studies, data collection could be improved by devices that can be carried less conspicuously to reduce the risk of theft and burden on participants while maintaining the quality of cough monitoring. Additionally, studies would benefit from devices which do not require charging for the duration of the recording period. This would reduce any time that the participant must be away from the device and ensure that the device could be kept close to the participant during sleep.
In summary, continuous cough monitoring is a promising avenue of research for multiple respiratory conditions, and we have demonstrated that this form of data collection is feasible at five international sites.
Data Availability
All data produced in the present study are available upon reasonable request to the authors.
