Acoustic models for cough detection perform well on curated benchmarks but can fail silently when deployed in real environments. Speech, environmental noise, and other audio that differs from the training data can trigger confident misclassifications, undermining trust in health-monitoring applications.
To address this, the authors propose SECURE-AI, a framework that applies multi-signal uncertainty quantification to cough detection. Rather than relying on a single confidence score, it integrates multiple signals of uncertainty to better identify when a model's output should not be trusted.
The work, posted on arXiv, is still at the preprint stage and does not report experimental results in the abstract. Still, it highlights a key gap in acoustic health AI: accuracy on benchmarks is not enough if models cannot recognise their own limits in the wild.