Validity of AI-based long-term cough monitoring for clinical use

Chronic cough is a debilitating condition that affects quality of life and often signals underlying respiratory disease. Clinical studies currently rely on audio recordings reviewed by human analysists to objectively quantify cough frequency, but this process is time-consuming and limits long-term monitoring. Emerging automated cough measurement tools enable multi-day measurements and therefore offer a more accurate assessment of a patients actual cough burden. At the same time, they are expected to show varying performance across patients. The tradeoff between algorithmic accuracy and the number of measurement days required to obtain a valid estimate of an individuals objective cough burden has not previously been explored. In this study, we aimed to validate the performance of our proprietary cough detection algorithm, part of an automated cough monitoring system, in patients with chronic cough. We also investigated how algorithm performance and daily variability in cough frequency affect the optimal number of monitoring days. The algorithm was evaluated under real-life conditions in 51 patients with chronic cough, achieving a median sensitivity of 0.93 and a median precision of 0.94. By using a confidence score to identify unreliable data and applying bootstrapping simulations to model variability, we found that the largest gains in measuring true cough burden were achieved with three days of monitoring, with diminishing returns beyond seven days. In conclusion, because cough frequency fluctuates daily, automated cough counting tools can enhance human-annotated data by enabling multi-day monitoring to more accurately capture true cough burden.