Performance Requirements for Cough Classifiers in Real-World Applications

In the context of monitoring respiratory diseases, an unobtrusive cough monitor is an attractive tool. Preferably, such tool requires little or no customization. We address the question of the feasibility of such a device. A large database of sounds including coughs and other events was available. Using deep learning, a general cough classifier was constructed. The plug-and-play feasibility of such cough classifier is addressed by a leave-one-patient-out procedure. For a large part of the cohort (80%), the performance of the classifier is excellent meaning an area under the curve (AUC) of larger than 0.9. On top of that, estimates are derived for its success in practical scenarios by considering the prevalence of cough and the required specificity. It is shown that the acoustic environment can be harsh, requiring very high specificities. From the results, we argue that for real-world applications customization will be required. For part of the population, it suffices to set a patient-specific operation point in generic cough classifier, but for some part a personalized cough classifier will be needed.