In 2019, the outbreak of SARS-CoV-2 set off a global pandemic, substantially impacting human lifestyles and posing an unprecedented challenge to health care systems worldwide [1]. Given its high infectivity, the urgent need for rapid and accurate COVID-19 detection has become paramount for effective patient management, timely isolation, and implementation of appropriate public health measures. Consequently, numerous studies have been undertaken to devise efficient methods for detecting infection status quickly and easily. Among these studies, artificial intelligence (AI) models that leverage prominent symptoms, such as coughing, have shown potential [2,3]. The ability to discern COVID-19 solely based on analyzing recorded cough sounds presents a straightforward and swift approach to screening individuals. These AI models offer hope for rapid, large-scale screening and early identification of potential cases. However, it is essential to acknowledge certain limitations in these research trends, particularly regarding variations in symptoms attributed to different SARS-CoV-2 variants [4]. The emergence of new variants has introduced complexities in symptom profiles, making it more challenging to rely solely on cough sounds for accurate COVID-19 detection. As the virus continues to evolve, ongoing research and adaptations in AI models will be crucial to address these challenges and improve the accuracy and effectiveness of diagnostic approaches. Particularly, patients with the Omicron variant have shown a decreased occurrence of characteristic symptoms such as loss of taste and smell, while cold-like symptoms such as sneezing and a blocked nose have increased [5]. In essence, the symptoms of COVID-19 have been increasingly resembling those of a common cold. This implies that the previous AI models designed to detect COVID-19 using cough sounds may not work as effectively or may not be as reliable both currently and in the future.

After the cough sound preprocessing, we developed an AI model to detect COVID-19 from extracted features and time-frequency spectrum using variable frequency complex demodulation (Multimedia Appendix 1 [6-9]). Table S2 in Multimedia Appendix 1 [6-9] summarizes the performance of our proposed AI model according to each data set and data measurement period. The results show that the cross-validation and test results from the Cambridge data set overlap, with an area under the receiver operating curve (AUC) of 0.93. Furthermore, the Virufy data set, which was measured around a similar period as the Cambridge data set, also demonstrated a similar performance: the AUC was 0.92. On the other hand, in the case of the Coswara data set, divided into 3 groups based on the variants, it was observed that the AUC for Alpha was 0.83; for Delta, 0.77; and for Omicron, 0.55; this indicates a decline in performance as the variants progressed. Figures 1 and 2 summarize the overall performance for each data set and measurement period.

The fight against COVID-19 has been marked by the deployment of various traditional diagnostic methods, each serving as a critical pillar in our response to the pandemic. Central among these is the reverse transcription-polymerase chain reaction test, widely regarded as the gold standard for its high specificity and sensitivity [12,13]. Rapid antigen tests and serological assays have also been instrumental, offering quick screening and insight into past infections, respectively. Despite their strengths, these methods have limitations, including resource dependency, time constraints, and varying degrees of accuracy. In this context, AI has emerged as a powerful aid, augmenting traditional diagnostic methods and addressing their shortcomings. By facilitating enhanced data analysis and interpretation, AI algorithms have the potential to streamline reverse transcription-polymerase chain reaction workflows, improve the reliability of antigen tests, and provide a more nuanced understanding of serological data. The integration of AI extends further to the analysis of medical imaging, where it aids in identifying patterns indicative of COVID-19, thus offering a valuable complement to molecular testing [14-16].

Our research has focused on the novel application of AI in analyzing cough sounds, a symptom-based method with promise for noninvasive, rapid screening. We demonstrated that AI models, when trained on cough sounds, could provide a swift approach to preliminary screening, though their performance varied with the emergence of new variants. Despite the performance variability, the inherent adaptability of AI models ensures their lasting relevance. These models can be reconfigured as pretrained frameworks ready to be fine-tuned against emerging viral strains. Such adaptability allows for the rapid deployment of AI tools in response to evolving pathogens, showcasing the potential of AI to serve not only the current pandemic but also as a foundational technology for future outbreaks. The capacity of these AI models to evolve in step with the virus paves the way for an agile diagnostic ecosystem that can quickly adapt to new threats, ultimately contributing to a more resilient global health infrastructure.