The Challenges of Understanding Cough in Continuous Audio recordings: An Interview With Dr. Mindaugas Galvosas
Researchers are exploring ways to use people’s cough noises to aid in the remote diagnosis of specific diseases or disorders. Cough sounds provide crucial information about the respiratory system and its diseases.
However, training a sophisticated artificial intelligence (AI) model to capture and evaluate coughs in everyday environments requires some intense work in defining what a cough is and how it looks in a continuous audio recording.
As an experienced specialist will attest, cough acoustics have plenty of clinical value. The more data available to these algorithms, the more accurate the biomarkers become. However, we must address some challenges associated with digital audio cough recordings to keep gaining ground for this work.
This week, I had a virtual interview with Mindaugas Galvosas, MD, now working in the medical team at Hyfe to explore his cough interpretation expertise.
MS: Thank you for accepting this interview. Please share a little speck about yourself and your interest in cough interpretation with us.
MG: Thank you for the invite. I am a medical doctor from Lithuania with a huge interest in breakthrough health technology and a background in pharmacovigilance and global health organizations. In summer 2021, I joined the medical team at Hyfe and, after a thorough analysis of the available literature on devices for cough, began digging deeper on cough interpretation.
It started with designing a system to annotate continuous audio recordings with cough and other explosive sounds, including respiratory system sounds. Now, we are looking into the signatures of cough sounds specific to particular health disorders.
MS: What is your approach to objective cough interpretation?
MG: Objective cough assessment starts with quantifying the coughs. How much did you cough last day? Or last week? You would only be able to answer these questions today if you are journaling your coughs, audio recordings that you replay and count, or using Hyfe Cough Tracker on your smartphone.
Having objective quantitative data — the number of coughs per hour, per day — allows patients, clinicians, and researchers to understand cough patterns and dynamics, treatment response, and even irritating and alleviating factors.
MS: What metrics should we measure in order to have a better understanding of cough?
MG: To date, cough is evaluated as a symptom quite subjectively “how much do you cough?”, “is it wet or dry?”, “is it with sputum or not?” there are also some validated questionnaires about cough severity. When a coughing patient is present, any doctor will try to better understand the origin, ask about one’s clinical history, do the lung X-ray, and run relevant tests.
Most clinical algorithms work in a way that you first suspect the most common cause for some set of symptoms and typical history. Only later, by differentiating test results and other findings, you come up with the diagnosis and specific targeted treatment.
With cough, sometimes it can be refractory to any treatment possible, sometimes it has a psychological origin, sometimes it is unexplained. Therefore, continuously measuring cough as an objective finding will definitely improve our understanding of cough, its causes, and what works best in treating it.
MS: Cough is a symptom with a plethora of data previously analyzed in a limited way. What are the perks of a precise cough assessment?
MG: If we start quantifying coughs accurately and in a patient-friendly way, we start learning more about the disorder itself and exploring broader treatment strategies. Only after having quantifiable information, which is relevant for the continuity of monitoring, we dig deeper into the understanding of respiratory “signatures,” specific to many respiratory disorders such as asthma, bronchitis, cystic fibrosis, whooping cough, tuberculosis, bronchial carcinoma, and COVID-19.
MS: In your opinion, What is the greatest challenge of automated (e.g., AI) cough interpretation?
MG: There are many challenges here. Firstly, for a machine to objectively understand cough — data is crucial. The amount of data should be huge to represent possible variability among cough sounds, and it should be of great quality, meaning that the cough sounds on which the machine is trained should be defined by a gold-standard method for ground truth — currently, human-annotated cough sounds are used for this purpose. In this process, inter-person and intra-person variability are present if no standard procedure is defined.
Developing and validating such a procedure is a challenge; however, we are confident that we are very close to achieving it. In the near future, there will be cough signature interpretation and challenges related to diagnostic accuracy.
MS: Your work at Hyfe involves using cough sounds to track cough frequencies and identify cough trends; what are the goals and expected outcomes?
MG: Having worked on a few versions of standard operating procedures for cough annotation in continuous audio recordings, our main goals were to define what a cough is: how it sounds and looks on a spectrogram.
In the process, we have discovered that some coughs sound almost identical to throat clears and vice versa — that was a challenging moment, requiring multidisciplinary discussions.
As of today, we are using the SOPs to annotate multiple respiratory sounds: coughs, throat clears, and sneezes. Having such a standard for the annotations is necessary for us to scientifically validate the Hyfe Cough Tracker during the clinical validation studies that are currently ongoing.
We expect to validate the algorithm and scale the use of the application among respiratory health researchers, clinicians, and their patients.
MS: How will machine learning help us understand cough in continuous audio recordings better?
MG: I believe that machine learning (ML) can and will play a very important role in, firstly, finding and classifying the patterns of cough in continuous audio recordings, and secondly, analyzing the cough sound itself — be it a single sound or a set of sounds. These features alone are examples of automation and means of analysis that perform way faster than a human sound annotator would do.
If other data types are added, e.g., health parameters, treatment strategy, and drug adherence, ML can be extremely insightful and helpful in better understanding each coughing person individually. Personalized approaches will be common practice in the future of healthcare.
MS: Which aspects of a cough does the Hyfe Cough tracker currently address?
MG: Today, the Hyfe app, which is free for Android and iOS devices, is useful in cough quantification, noticing the dynamics of continuous cough, and it also lets the user listen to the counted cough sounds. Hyfe research app and its dashboard integration provide more details on cough and are useful for remote patient monitoring and clinical research.
MS: Are there some coughs with unique sounds that you can easily diagnose simply by hearing?
MG: Some scientific publications are showing that asthma, bronchitis, cystic fibrosis, whooping cough could be identified from cough as a symptom and its sound as a data point. Yet, more research and bigger cohorts are needed in this field for these technologies to be accurate, validated, and widely used.
MS: Why is research into cough important for public health, and who else should be involved in these projects?
MG: In my opinion, cough seems to have been neglected as a data-rich symptom. Monitoring and analyzing cough quantitatively will probably be useful in coordinating some actions related to public health measures: determining early outbreaks of respiratory infections and taking necessary actions accordingly.
Currently, the area of cough monitoring could have more support from all stakeholders, to name a few — patient groups in spreading awareness about advancements in “cough journaling” and learning about one’s cough; media in highlight the transition from cough as a “wet/dry” symptom in daily practice to a more elaborate data point; and also collaborative efforts from clinicians and researchers in standardizing the new practices of cough monitoring.
MS: What’s the most interesting thing about a cough that you’ve learned in your research?
MG: Interestingly, it is sometimes very difficult to differentiate between a cough and a throat clear sound when listening to continuous audio recordings, also observing the sound wave and its spectrogram.
After talking to numerous experts in the field, there is no clear distinction between these two. However, currently, in labeling tasks, we treat coughs as sounds with air expulsion present and throat clears as sounds with no audible/visible air expulsion.
MS: What are your opinions on how cough will be understood in the future?
MG: The future of cough research will bring us some insights that we can currently only imagine. Just thinking about implantable sensors and discoveries there is exciting! Today, speaking about monitoring cough through a patient’s smartphone still sounds like a breakthrough solution, and even though many researchers and clinicians are excited and running tests with it, the technology should be validated, and some standards should be agreed upon in the scientific and clinical community. I believe it will not take long.
MS: Is there anything more you’d like to add?
MG: Thank you for the interview. I hope it was interesting.
Conclusion
Coughs can now be observed in real-time, thanks to advances in technology and digital audio. As a result, researchers have been able to investigate cough in greater depth than ever before.
For therapeutic research, digital cough recordings have various advantages. There are no distractions, such as coughing from other patients or outside noises like television or conversations, as there are in a standard cough study.
Furthermore, a recording can be played back multiple times to identify each person’s distinct coughing pattern. Clinicians may be able to treat certain respiratory disorders more successfully and with fewer side effects if they have this information.
Hyfe was developed to use acoustic epidemiology to revolutionize healthcare systems and enhance global health. To that end, Hyfe is developing acoustic diagnostic and monitoring tools that are simple to use, quick to evaluate, accessible to everybody, infinitely scalable, and have a global impact.
These tools are based on the world’s largest and fastest-growing cough dataset, are constantly enhanced through machine learning, and have been thoroughly vetted through collaborations with clinical researchers worldwide.
Many thanks to Dr. Mindaugas Galvosas for his contribution to this article.
