Development of a Digital Biomarker for Symptomatic COVID- 19 via Frequency-Aware Spectral Features and Machine Learning Tuning

Abstract

Cough-based analysis has emerged as a non-invasive, low-cost solution for identifying symptomatic respiratory illnesses, including COVID-19.However, current models often suffer from low interpretability, frequency insensitivity, and poor performance under class imbalance. To addressthese gaps, this study aims to develop a frequency-aware classification framework tailored for symptomatic COVID-19 detection using cough sounds as a potential digital biomarker. The method involves decomposing audio into four spectral bands—[300–400 Hz], [300–600 Hz], [450–1000 Hz], and [1100–1650 Hz]—followed by spectral analysis, feature extraction, and classification via XGBoost. Advanced training strategies incorporating Bayesian optimization, SMOTE up-sampling, and threshold moving were applied. Results show the [450–1000 Hz] band yielded the highest performance, with precision of 0.96 and recall of 0.97, outperforming state-of-the-art baselines. This study concludes that integrating frequency segmentation with interpretable training enhancements can significantly improve symptomatic COVID-19 detection, establishing cough acoustics as a viable digital biomarker for scalable deployment in real-world, resource-constrained healthcare environments.