Speech as a Biomarker for COVID-19 Detection Using Machine Learning.

Area of Science:

• Biomedical Signal Processing
• Machine Learning Applications
• Respiratory Illness Diagnostics

Background:

• Physiological changes associated with COVID-19 can alter speech characteristics.
• Speech spectral features, analyzed via Short-Time Fourier Transform (STFT), offer potential biomarkers.
• Distinguishing between healthy and COVID-19 positive individuals using speech is an emerging diagnostic approach.

Purpose of the Study:

• To investigate the efficacy of speech spectral features for COVID-19 diagnosis.
• To apply and evaluate machine learning classification algorithms for identifying COVID-19 positive individuals.
• To optimize models for minimizing misclassification of positive cases.

Main Methods:

• Statistical analysis of speech spectral features derived from STFT.
• Utilizing speech samples from both healthy and asymptomatic COVID-19 positive individuals.
• Training and evaluating five state-of-the-art machine learning classifiers, including Decision Forest.

Main Results:

• Higher Root Mean Square (RMS) error in statistical distribution fitting for COVID-19 positive speech samples compared to healthy samples.
• Performance evaluation of multiple machine learning algorithms, with parameter tuning to prioritize correct identification of positive cases.
• Decision Forest algorithm achieved the highest recall of 0.7892 in classifying COVID-19 positive individuals.

Conclusions:

• Speech spectral analysis, combined with machine learning, demonstrates potential for non-invasive COVID-19 detection.
• The study highlights the statistical differences in speech patterns between infected and healthy individuals.
• Optimized machine learning models, particularly Decision Forest, can effectively aid in the early identification of COVID-19.