(0806) COVOICE-19 Project: COVID Screening by Voice and Artificial Intelligence

Introduction: The COVOICE-19 project aims to improve the rapid screening of COVID-19 through AI-based voice classification techniques. This research investigates the link between specific vocal features and COVID-19, enabling swift non-invasive testing methods.

Methods: The study was conducted in 2021, encompassing both hospitalized and outpatient individuals, with 204 recordings of sustained phonation (/ah/). The study was conducted in 2 Portuguese hospitals, with nasal swabs (RT-PCR) for COVID-19 resulting in 100 positive and 104 negative samples. The voice recording/analysis was based on vocal characteristics. The independent variables included an array of vocal features, such as Mel-Frequency Cepstral Coefficients (MFCCs) and spectral properties. Preliminary analyses feature engineering and acoustic analysis, creating a dataset with 152 features. This dataset was used to train machine-learning classification models for predictive analysis, with the mean of relevant metrics calculated through K-Fold Cross Validation with K = 10 and 10 repetitions across 100 model fits with a 95% confidence interval. The SHapley Additive exPlanations (SHAP) framework was used to improve the interpretability of the models.

Results: The results showed a validation accuracy of 90% with 85% sensitivity and 97% specificity for the top-performing model (Multilayer Perceptron). The Random Forest model achieved a validation accuracy of 86%. The models were capable of discerning COVID-19 positive cases from negative ones. The acoustic/subjective analysis of voice revealed significant alterations in vocal characteristics, such as roughness and tension, in COVID-19 positive compared to the control.

Conclusions: .This project underscores the potential of AI assisting in COVID-19 diagnosis through voice analysis. The study's findings reveal a strong correlation between specific vocal features and COVID-19. Also, AI models demonstrating high accuracy, sensitivity, and specificity. These results show that non-invasive and quick testing methods are feasible, opening the door for AI-driven research in diagnostic methods.