Using AI to detect COVID-19 tested as a potential method of rapidly and efficiently diagnosing patients.
The rapid global spread of the SARS-CoV-2 virus has presented a huge number of healthcare challenges. One of the greatest difficulties in managing the pandemic is the diagnosis of the disease. Reverse transcriptase polymerase chain reaction tests (RT-PCR) remain the gold-standard in diagnosing coronavirus, however this method has many drawbacks. Results can take up to two days, repeated tests may be needed to rule out false negatives, and sheer number of suspected cases has led to global supply chain issues for RT-PCR kits and reagents.
As a result, attempts are being made to explore alternative approaches to diagnosis. One such alternative is explored in a paper published in Nature Medicine (1). This study investigates the combined use of chest computed tomography (chest CT) with artificial intelligence algorithms (AI) as a potential method of rapidly and efficiently diagnosing COVID-19.
Chest CT scans are currently a key component in assessing COVID-19 disease severity. However, in many early or mild cases, chest CT findings will remain normal despite the presence of active infection. Therefore chest CT alone is not a reliable indicator of infection. It is likely to result in a number of false negatives. Instead, this study examines combining chest CT with other readily available patient data such as exposure history and clinical symptoms.
The research team constructed artificial intelligence algorithms that could pull together these various data sources and give a positive or negative diagnosis. Having constructed the AI algorithms, they tested the diagnostic ability using a variety of approaches. The first was the use of a receiver operating characteristic curve (ROC curve). This plots the sensitivity and specificity of a diagnostic technique, with the overall ability of the technique measured by the area under the curve (AUC). A perfect test has an AUC equal to 1 whereas a non-discriminating test has an AUC equal to 0.5. Generally speaking, an AUC of greater than 0.9 is considered excellent or outstanding. When tested using real patient data, the AI/CT scan approach achieved an AUC of 0.92, indicating outstanding performance.
However, the ROC curve approach weighs sensitivity and specificity equally. The specificity of a diagnostic test refers to its ability to accurately detect the true negatives i.e., to correctly identify those suspected cases that do not actually have the disease. Sensitivity on the other hand refers to the ability to correctly identify the true positives. A more sensitive test will allow fewer positive cases to slip through the cracks. In this case, sensitivity was a more important measure as active cases or true positives may have normal CT scans.
Therefore, the sensitivity of the approach was assessed and compared to the performance of a senior thoracic radiologist. The AI algorithms achieved a sensitivity of over 84%, comparing favourably to the sensitivity of a senior thoracic radiologist (74.6%). Interestingly however, the thoracic radiologist achieved a higher level of specificity (93.8% vs 82.8%). This suggests that the AI model is more likely to generate false positives. But in the case of coronavirus, false negatives are a greater threat to public health than false positives.
Overall, this innovative approach has several key advantages over the RT-PCR method. The AI/CT scan approach has the potential to process large numbers of tests in a short space of time and is not reliant on limited consumable resources such as testing reagents. Furthermore, as we continue to learn more and more about COVID-19, new information can be used to modify the AI algorithms, potentially improving both sensitivity and specificity.
The AI/CT scan approach is unlikely to replace RT-PCR as the gold standard in testing for COVID-19. However, this study provides promising evidence of how it could be used alongside RT-PCR, reducing pressure on resources and lab capacity.
Written by Michael McCarthy
For a range of personal protective equipment, visit www.medofsupply.com
1. Mei X, Lee H-C, Diao K-y, Huang M, Lin B, Liu C, et al. Artificial intelligence–enabled rapid diagnosis of patients with COVID-19. Nat Med. 2020.
Image by PIRO4D from Pixabay
