Configurable Pulmonary-Tuned Privacy Preservation Algorithm for Mobile Devices

Audio-based automated pulmonary symptom detection has the potential to offer accurate and continuous assessment of patients with lung disease. However, privacy preservation becomes a significant issue when it comes to continuous passive audio recording. Various techniques have been employed to obfuscate the speech within audio in these applications. However, that penalizes the accuracy of detection by affecting of-interest, non-speech audio parts. This is inevitably undesirable as it contradicts the notion of sensing. In this paper, we propose a novel algorithm to achieve the goal by employing a machine-learning-based vowel detection algorithm. The algorithm is implemented in a configurable manner to address many different scenarios of data collection. Logistic regression has been utilized to make the algorithm feasible for on-device implementation. We have shown that our algorithm achieves the goals in that the obfuscated speech is unrecognizable while cough sounds are identifiable by symptom detection models as well as a human ear.