Work

Contrasting deep learning audio models for direct respiratory insufficiency detection versus blood oxygen saturation estimation

Year: 2025

Type: article

Abstract: Background:Recently, several deep learning audio models have been proposed to detect respiratory insufficiency in COVID-19 and general patients through voice and speech analysis, achieving accuracy of 97.4% in the COVID case and 95% in the general case.Objectives:By employing state-of-the-art deep learning architectures for voice and speech audio analysis, this study contrasts the tasks of respiratory insufficiency (RI) detection and blood oxygen saturation (SpO2) estimation.Methods:We use pretrained audio neural networks (CNN6, CNN10, and CNN14) and the Masked Autoencoder (Audio-MAE) for RI detection. These models were pretrained on AudioSet, with transfer learning playing a key role in preventing overfitting. We perform binary classification for RI detection, consistent with prior works. As deep learning models perform better in classification than regression, we perform a simple $92\%$ SpO2-threshold binary classification. This aims to evaluate the feasibility of SpO2 estimation from voice and speech.Results:For RI detection, the models considered achieve near-perfect accuracy (99.9% on COVID RI and $98.6\%$ on general RI), surpassing previous results. Yet, for the task of estimating SpO2 (binary threshold classification), the accuracy stays below 70% and an MCC-correlation below 0.3. This sharp contrast suggests the presence of multiple features in the audios which are useful for RI detection, but cannot be used for SpO2 estimation. Alternatively, one might consider the hypothesis that patient treatment masked SpO2 levels, while voice and speech still retained RI cues.Conclusion:The near-perfect RI detection performance suggests low-cost and automated methods can be developed for assisting patient triage. Moreover, while automated audio analysis offers a valuable assessment of a patient's RI status, it does not provide as accurate information about SpO2 levels, highlighting the limits of voice and speech biomarkers across different diagnostic tasks under current technologies.

Source: Intelligence-Based Medicine

DOI: 10.1016/j.ibmed.2025.100331