Artificial intelligence enabled COVID-19 detection: techniques, challenges and use cases

The COVID-19 pandemic underscored the critical need for rapid and efficient diagnostic tools. The limitations of existing methods, such as RT-PCR, which suffers from long turnaround times, limited availability of kits, and the requirement for specialized labs and trained personnel, prompted the exploration of AI-powered solutions. Machine learning (ML) and deep learning (DL) offer promising alternatives due to their flexibility, adaptability, and potential for automation. This paper investigates the application of ML and DL models in addressing the challenges of COVID-19 detection across various domains, including early warning systems, tracking and prediction, data dashboards, diagnosis and prognosis, treatment and cure development, and social control measures. The research was motivated by the lack of a comprehensive survey that encompasses all these areas and addresses the cost-effectiveness and high effectiveness of these techniques. The rapid advancements in deep learning models and their potential to significantly transform the healthcare industry further fueled this review.

The paper reviews several existing studies on COVID-19 detection. Giri et al. (2020) compared diagnostic methods based on sensitivity, specificity, and throughput. Other studies, such as Shah et al. (2021) and Udugama et al. (2020), provided in-depth reviews of AI techniques applied to various aspects of the pandemic response, including early warning systems and social control. However, these studies lacked a comprehensive comparison of existing state-of-the-art detection methods across different modalities and a discussion of the unresolved challenges and limitations.

The authors conducted a comprehensive literature review using Google Scholar, IEEE Xplore, ScienceDirect, and Springer, focusing on peer-reviewed articles. Their search criteria included keywords such as "Detection techniques for COVID-19," "Applications of Machine Learning in COVID," and "Deep Learning technologies for COVID-19." The search process involved refining the search strings to ensure relevance and manually screening review papers to identify relevant studies. The review covers laboratory tests (RT-PCR), blood tests using machine learning, and AI techniques, including neural networks, CNNs, and related technologies like RNNs, KNN, decision trees, linear and logistic regression, ResNet50, VGG19, federated learning, and transfer learning. A comparative analysis of datasets used in various studies was performed, categorizing them into RT-PCR, X-ray, and blood test datasets. The analysis assessed the ease of use, affordability, and accuracy of different detection techniques. The paper also includes a research gap analysis and discusses the challenges and limitations of current AI-based COVID-19 detection methods.

The review identified several key findings: 1. **RT-PCR:** While the gold standard, RT-PCR suffers from limitations including long turnaround times, kit shortages, and the need for specialized infrastructure. ML models offer a faster, cheaper alternative. 2. **Blood tests:** ML models trained on hematological parameters (e.g., WBC, LDH, CRP, platelets, D-dimer) can aid in COVID-19 detection and severity assessment, but accuracy is limited by dataset size and quality. 3. **Chest X-rays and CT scans:** DL models, particularly CNNs, are effective in analyzing lung X-rays and CT scans to detect COVID-19. However, accuracy depends on the quality and size of training datasets, and the models may not generalize well to different populations or equipment. 4. **Dataset limitations:** A major challenge is the limited availability of large, high-quality datasets for training and testing AI models. Data augmentation and synthetic data generation techniques were used, but these have limitations, impacting model generalization. 5. **Model accuracy:** The accuracy of AI-based detection models varies significantly, often falling short of RT-PCR's accuracy, especially when trained on small datasets. High accuracy rates reported in some papers may be attributed to unbalanced datasets. 6. **Ease of use and affordability:** AI-based models, once trained, offer the potential for faster, cheaper, and more accessible COVID-19 detection compared to RT-PCR, making them particularly suitable for resource-constrained settings. However, the computational resources required for training and deploying complex models may pose a barrier. 7. **Research gaps:** The review highlights gaps in the uniformity of AI-based COVID-19 detection algorithms, the availability of training data, the evaluation of real-world performance, and the transparency of AI models.

The findings of this review highlight the potential of AI-based methods for COVID-19 detection but emphasize the critical need for addressing the limitations associated with data availability, model accuracy, and generalization. While ML and DL models show promise in offering rapid and cost-effective diagnostic tools, particularly in resource-limited settings, the accuracy of these models remains below that of RT-PCR in many studies. The reliance on data augmentation and synthetic data generation techniques raises concerns about the generalizability of these models to different populations and clinical settings. Future research should focus on developing robust methods for data collection and curation to create large, representative datasets for training AI models. Improving model architectures and employing techniques like transfer learning could also enhance accuracy and generalizability. Rigorous validation through clinical trials and real-world implementation are crucial to assess the effectiveness and reliability of AI-based COVID-19 detection systems.

AI-based methods, particularly ML and DL, have shown significant promise in COVID-19 detection, offering faster and more accessible diagnostics. However, challenges remain, primarily concerning data availability and model accuracy. Future research must prioritize creating larger, high-quality datasets, improving model architectures, and conducting thorough clinical validation. The integration of various data sources, including clinical information and image data from different modalities, holds considerable potential for enhanced diagnostic accuracy and generalizability. The development of user-friendly, easily deployable AI tools would significantly improve the accessibility and efficacy of COVID-19 detection in diverse settings.

The review is based on the available literature, and the quality and generalizability of the findings are subject to the limitations of the included studies. The accuracy and performance of the AI models discussed are highly dependent on the quality and size of the training datasets used, and dataset bias may affect the results. Furthermore, the review may not capture all relevant studies on the topic due to the rapid evolution of research in this area.