Multi-Modal Approaches for Improving the Robustness of Audio-Based Covid-19 Detection Systems

Research efforts have demonstrated promising results employing artificial intelligence to detect COVID-19 with respiratory sounds, but these experimental milestones have not translated to real world impact due to a host of acoustical and data-driven challenges. Therefore, it is imperative that audio-based COVID-19 screening systems incorporate additional information to supplement decision-making, particularly when deployed under adverse conditions. Symptom checkers are a possible alternative for predicting COVID-19 but they are ineffective at screening asymptomatic subjects. In this paper we investigate the performance of COVID detection systems using audio, symptom, and multi-modal data. The results suggests that simple feature-based fusion approaches (AUC=0.786) slightly outperform symptom-based classifiers (AUC=0.779) in overall classification performance, but are significantly more valuable than symptom-based classifiers when screening asymptomatic subjects with a 42.4% accuracy advantage compared to 9.4%.