Cough monitoring, namely for the measurement of cough frequency, has seen increased interest and use in clinical trials to provide a greater understanding of treatment effects on subjects’ cough. As sponsors and researchers evaluate cough monitoring technologies for their studies, an important aspect is the cough counting method used to determine cough frequency.
There are currently three main approaches used for objective cough counting: manual (human) cough counting, semi-automated cough counting, and fully automated cough counting. Each method has benefits and limitations, differing in accuracy, efficiency, and scalability, with varying applications for each. Below, we explore these approaches and their best use cases in life sciences and research.
Manual Cough Counting
Manual cough counting involves human annotators—typically trained respiratory therapists, pulmonologists, or clinical researchers—who listen to complete audio recordings to identify cough events. Because experienced professionals listen to complete 24-hour recordings and identify coughs, this method is considered the gold standard by the FDA in terms of accuracy.
Manual counting, however, is more time-consuming and labor-intensive, making it less practical for large datasets. Most periods of audio recordings will not contain coughs, even among patients with extremely high cough rates – yet the stochastic nature of coughing necessitates that all recordings be examined. Although manual counting presents risks for patient privacy, there are methods available to preserve patient privacy, such as using voice activity detection (VAD) tools to identify and render patient speech undiscernible while simultaneously preserving cough sounds.
| Benefits | Limitations | Best Applications |
| – Highly accurate when performed by trained annotators – Useful for validating complementary cough counting methods with regulatory guidance | – Time-intensive, requiring hours of annotation per dataset – Expensive, as it relies on trained human experts | – Validation of semi-automated cough counting methods with regulatory guidance – Earlier phase cough studies with low subject enrollment |
Semi-Automated Cough Counting
Semi-automated cough counting offers the benefits of manual cough counting but with greater efficiency.
Image 2: Strados Labs semi-automated cough counting process
In the case of Strados Labs’ semi-automated process, an algorithm pre-screens complete audio recordings and identifies portions unlikely to contain cough (such as those quiet times captured while patients are sleeping) before human annotation occurs. The non-cough containing portions of the recordings are subsequently withheld from annotation, reducing the total annotation time and cost while ostensibly retaining all the cough events. Note that all original data is still retained, allowing for future quality control review and audit. In some semi-automated systems, including Strados Labs’ process, a separate voice activity detection (VAD) algorithm is applied to the audio dataset to render speech unintelligible and preserve patient privacy.
Due to its balance of cough count accuracy, annotation efficiency, and patient privacy, the semi-automated cough counting approach has been widely used in later phase pharmaceutical clinical trials where the number of subjects is higher; note, this method does require additional validation to the FDA for each study so it may be impractical for earlier phase studies.
| Benefits | Limitations | Best Applications |
| – More efficient and cost-effective than fully manual counting – Retains human oversight, ensuring strong accuracy – Patient privacy is preserved through speech obfuscation filters – Clear regulatory pathway for pharmaceutical clinical trials – Retains all data from cough recording, allowing for future review, audit, or algorithm performance tracking | – Human annotation of processed audio datasets is still time-intensive – Validation of the processes for 1) withholding non-cough sections of the recordings, and 2) obfuscating speech, may be required by the FDA for each study | – Cough trials where cough frequency is a primary endpoint (often later phase studies) – Respiratory trials (e.g., IPF) where cough frequency is an exploratory endpoint or measured in a sub-study |
Fully Automated Cough Counting
Fully automated cough counting relies on machine learning to detect and quantify coughs, without human annotation. Machine learning models are typically trained on large datasets containing hundreds of thousands of cough sounds, allowing them to distinguish coughs from speech, throat clearing, environmental noise, and other respiratory sounds.
While fully automated systems offer the least accuracy of the three approaches discussed, they offer the greatest efficiency and scalability. Fully automated cough counting allows for processing of thousands of hours of cough recordings in a fraction of the time compared to manual or semi-automated cough counting, making it practical for longitudinal cough monitoring studies, high-volume subject screening, and near real-time remote monitoring. The regulatory validation requirements and pathway for using this approach in pharmaceutical clinical trials are less clear than those of the semi-automated approach.
| Benefits | Limitations | Best Applications |
| – Scalable, capable of processing large amounts of data – Cost-effective, eliminating the need for human annotators – Facilitates longitudinal cough monitoring over weeks or months | – Potential for false positives or missed coughs, especially in noisy environments – Less clear endpoint regulatory acceptance and pathway, especially as a primary or secondary endpoint – Depending on the underlying cough monitoring technology, audio data may not be retained for future study and/or verification of semi-automated results – Automated cough detection algorithms, if running directly on the audio recording device, may reduce overall battery life & monitoring time | – Longitudinal research studies requiring continuous cough monitoring over extended periods, e.g. on the order of weeks or more – Fast, inexpensive, and high-volume screening for patients in clinical trials or research studies – Exploratory studies where regulatory validation requirements may be less stringent – Near-real time monitoring of symptoms e.g. coughs and cough bouts |
Which Approach is Best?
Identifying the right cough counting method for your study depends on the specific needs of the study or clinical application:
- Manual cough counting is practical for early phase cough trials with fewer subjects, and/or validating semi-automated cough counting methods for later phase trials based on regulatory guidance. (See Strados Validation Guidance for more information).
- Semi-automated cough counting offers a balance between cough count accuracy and efficiency. The semi-automated method is typically used in later-phase cough studies where 24-hour cough frequency is the primary endpoint and where subject enrollment is higher; it is also applicable for measuring cough as a secondary endpoint, exploratory endpoint, or in a sub-study for a respiratory indication where cough is prevalent and burdensome, such as idiopathic pulmonary fibrosis or bronchiectasis.
- Fully automated cough counting offers maximal efficiency with generally lower accuracy relative to manual and semi-automated cough counting. Fully automated cough counting is useful for cost-effective screening of subjects, and/or for research studies where cough might be monitored over weeks or months.
At Strados Labs, we offer all approaches – manual, semi-automated cough counting and automated cough counting from lung sound recordings captured by our FDA 510(k) cleared RESP® Biosensor and other devices such as watches. Interested in learning more about our cough monitoring solutions? Contact us
Author
Tom deLaubenfels, PhD
Director of Data Science
Strados Labs