Get expert insights into cough science, monitoring and research
When it comes to accuracy of cough monitoring, sponsors wonder whether retaining audio recordings of cough is the best way to validate whether every cough is detected accurately. We will discuss audio retention in the following article, covering privacy, consent, data governance, and the possibility that being recorded changes how participants behave, and therefore what gets measured.
The FDA-adopted ICH E6(R2) guideline defines source data as “all information in original records and certified copies of original records of clinical findings, observations, or other activities in a clinical trial necessary for the reconstruction and evaluation of the trial”. For cough monitoring, the relevant clinical observation is that a cough occurred at a particular time.
In Hyfe’s approach, the timestamp produced by a locked, validated algorithm records that observation. The audio is the sensor signal used to detect it, for which we treat the timestamp as source data. FDA guidance also recognizes that digital health technologies can produce clinically relevant data as algorithm-derived events (FDA, 2023).
Let’s take cough annotation as an example of why replaying audio does not necessarily lead to better outcomes. In a Hyfe multicenter validation study (Chaccour et al., 2025), two independent trained annotators reviewed continuous recordings. Their agreement was high (inter-labeler Pearson correlation 0.96, intra-labeler 0.98, with fewer than two disagreements per hour; Sánchez-Olivieri et al., 2023), but under a strict protocol: the annotators worked blind to each other, and disagreements had to be formally arbitrated by a third expert. This example shows that retaining audio recordings does not lead to a self-explanatory ground truth. Human re-analysis still needs predefined definitions, trained human analysts, and procedures for handling disagreements. Confidence in cough measurements comes from validating the measurement process; simply being able to replay individual recordings does not establish that the system measures cough reliably.
Once cough events are processed by an algorithm, the question changes from Can someone replay every event? to Has the measurement process been shown to work? Analytical validation examines algorithms that convert measurements captured from sensors into physiological measurements (Goldsack et al., 2020). In the context of cough monitoring, analytical validation entails testing of the cough-detection algorithm.
Retaining audio is, nevertheless, helpful during the process of developing and validating a cough-detection system. For example, trained reviewers may listen to recordings, label cough events according to a predefined protocol, and then compare those labels with the algorithm’s detections.
This does not mean that every audio recording must be retained in case it is needed for future reviews. According to FDA’s guideline, a DHT should be validated to provide objective evidence that it appropriately assesses the intended clinical event in the target population; depending on the technology and its context of use, this validation may include comparison with appropriate reference measurements (FDA, 2023). If someone later listens to a particular recording and decides whether an individual timestamp was really a cough, they are performing a new human assessment of that event. That assessment has its own rules, uncertainty, reviewer variability, and possibly adjudication.
Another reason that retaining cough sounds is not appropriate in clinical trials has to do with the basic statistics of technology validation. Retaining sounds that were deemed to be coughs only provides supporting evidence for the question “were the coughs identified actually coughs” (aka specificity) and does not address the bigger question of “were all coughs identified” (aka sensitivity). Thus, to determine the performance of a cough monitor in a clinical trial would require human verification of the entire monitoring period - which is impractical.
First, there are privacy and consent considerations. Retained audio must be stored securely, access-controlled, and governed. Participants need to understand what is recorded, how it will be used, and who can access it (FDA, 2023). In the EU, the GDPR requires personal data to be limited to what is necessary for the purpose (Art. 5(1)(c)).
Second, there are the limits of continuous monitoring. One day of monitoring can miss important intra-subject day-to-day cough variability. Continuous monitoring over days or weeks can improve cough measurements, but keeping audio for that long can complicate consent requirements and increase storage, annotation, and analysis burdens. Such constraints can limit monitoring duration which is counter productive to enabling the optimal endpoints
Third and finally, our data show that monitoring can change participants’ behavior. In a Phase 2b chronic cough trial, participants in the placebo arm coughed more on days when they also wore a 24-hour audio recorder than on surrounding days when they wore only the watch (geometric mean ratio 1.48, 95% CI 1.27–1.73, at screening; 1.42, 1.24–1.61, at the run-in visit). The ratio was lower later in the study (1.16, 1.04–1.28, at day 15; 1.26, 1.14–1.38, at day 28). We attribute this to participants’ awareness of being observed, a pattern consistent with the Hawthorne effect (Wickström & Bendix, 2000). Continuous recording of conversations and other sounds may make monitoring more intrusive and amplify this effect. Whether audio retention itself adds reactivity remains to be tested.
Retaining more granular data does not automatically make a cough measurement more rigorous or otherwise useful. At the same time, retaining continuous audio introduces additional risks and obligations.
The cough timestamp produced by a locked, validated algorithm is the source data, not a proxy for it. Retaining raw audio doesn't add verification value; it adds privacy exposure, consent complexity, and a chain-of-custody problem of its own, while solving nothing that algorithm validation hasn't already solved once, upfront.
Chaccour, C., Sánchez-Olivieri, I., Siegel, S., et al. (2025). Validation and accuracy of the Hyfe cough monitoring system: a multicenter clinical study. Scientific Reports, 15, 880. Link.
European Parliament and Council of the European Union. (2016). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88. Link.
Frankel, M., Brew, J., Ellis, J., Shaw, J., Small, P., & Batycky, R. (2027). The Hawthorne effect of conspicuous short-term cough monitoring [Abstract submitted]. American Thoracic Society International Conference, 2027.
Goldsack, J. C., Coravos, A., Bakker, J. P., et al. (2020). Verification, analytical validation, and clinical validation (V3): the foundation of determining fit-for-purpose for Biometric Monitoring Technologies. npj Digital Medicine, 3, 55. Link.
International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use. (2016). Integrated addendum to ICH E6(R1): Guideline for good clinical practice E6(R2). Link
U.S. Food and Drug Administration. (2013). Electronic Source Data in Clinical Investigations: Guidance for Industry. Link
U.S. Food and Drug Administration. (2023). Digital Health Technologies for Remote Data Acquisition in Clinical Investigations. Link
