The burden of COVID-19 has spread rapidly across the globe. Widespread testing has been recognised as key to understanding and tackling the virus. In order for testing to be instructive, high targets must be reached – the WHO has made multiple calls for global mass testing.
However, quick and accurate testing at such an extensive level remains a challenge. In reaction, artificial intelligence (AI) and machine learning developers and experts have been quick to innovate and develop existing tools in an attempt to advance the speed and efficacy of diagnosis to aid healthcare systems and personnel.
The principal method for diagnosing COVID-19 is an antigen test in which oral and nasal swab samples are taken and analysed for the virus. Results can supposedly be realised within 48 hours. However, conducting tests and obtaining quick results relies on having adequate infrastructure (testing kits, labs, equipment) and human resources.
This poses a challenge in less developed countries, where health systems are weaker and under-funded. Even in Europe and the US, many countries have faced a shortage of testing kits. Conducting tests on the presence of symptoms or potential contact with the virus can also miss asymptomatic and pre-symptomatic cases. This enhances the risk of further transmission, particularly now as social distancing measures ease and schools and offices begin reopening. The accuracy of such tests has also come into question, and there is a high chance of false negatives occurring.
Where hospital resources are more scarce, proposed AI systems can deliver rapid results. Patients with a positive result can be isolated immediately whilst the burden on resources is limited. Crucially, a new testing kit is not needed for each patient. In addition to autonomous diagnosis, the AI model can provide a ‘second opinion’ to physicians in cases where a computed tomography (CT) scan is either negative (in the early course of infection) or shows nonspecific findings.
How does it work?
COVID-19 diagnosis models are utilising AI in the form of image recognition algorithms. CT scans that consist of multiple chest x-rays are analysed for abnormalities representing COVID-19. Based on images in the training data, the algorithm analyses the new scan for similarities and categorises the image to produce a positive or negative result. This can take up to as little as 30 seconds. This is not the first time AI has been applied in this way – image recognition has been harnessed for medical diagnosis in a number of solutions, including for detecting malnutrition. Similarly, x-rays have also been used to effectively detect pneumonia.
It is worth mentioning, there are other functions of AI being utilised for diagnostic purposes. For example, methods for predictive analysis are using blood samples and other vital signs (as in a recent Oxford study), and similarly, socio-demographic data is being used to generate baseline predictions to inform test pooling.
Technical and ethical challenges
As with all applications of AI, there are number of challenges and ethical considerations.
A machine learning algorithm is only as good as the data it is trained with, in regard to quantity and quality. A lack of training data compromises accuracy and reliability and for best results, researchers and developers need to collaborate and share data. It has also been suggested that there are variants in how COVID-19 presents across demographic groups and the likely severity of infection. However, disparities across gender, age, ethnicity and other demographics are yet to be fully understood.
It is, therefore, crucial to include a wide sample of cases in the training data to mitigate categorisation bias and inconclusive results for those likely to be under-represented or missed entirely. With the novelty of COVID-19, such data may be difficult to obtain and developers may not recognise the need to actively source data from minority groups. Whilst sharing data is encouraged, data shared across contexts may still not provide a full or accurate representation of the target population. In the accelerated development of solutions, compromises on the use of sub-optimal data sets may occur in a trade-off for speed over accuracy.
CT scans combine multiple x-rays which may not conform to a consistent structure across samples. This messy data can trip up the algorithm when analysing to produce less accurate results. Additionally, scans that produce a positive result for COVID-19 may not distinguish between abnormalities that present similarly in overlapping respiratory illnesses such as influenza, H1N1 and other SARS viruses.
Acceptance of outcomes regarding an individual often relies on transparency – the ability to explain how a result was achieved. This may be disregarded in the quick development of solutions, as retaining full explainability of the AI’s process would cost time. Furthermore, should discrepancies occur, human accountability is imperative. Quick uptake and development of this technology for combating COVID-19 additionally means that relevant policy, regulation and best practices lag behind and data privacy may be overlooked.
Examples
| Mount Sinai | Mount Sinai researchers at Icahn School of Medicine have developed AI algorithms that integrate chest CT findings with clinical information such as age, temperature, symptoms and exposure history. They have processed over 900 scans (as of March 2020), delivering accuracy on par with that of an experienced radiologist. However, limitations have been recognised, including a small sample size and bias in the training data. This could limit the model’s ability to distinguish COVID-19 from other respiratory issues. |
| Behold.ai | Allegedly, Behold.ai runs the world’s largest deployment of a radiology-based AI diagnostic solution for COVID-19. The Behold.ai algorithm has been developed using over 30,000 image samples, all of which have been reviewed by consultant radiology clinicians. The platform is capable of sorting images into normal and abnormal categories in under 30 seconds. The algorithm does not detect COVID-19 in isolation but instead reduces the burden on healthcare systems by working as a fast and accurate triage tool. Behold.ai will supply its algorithm to Apollo Hospitals’ network of 71 hospitals and some local government-run hospitals in India. |
| Damo Academy – Alibaba | The research arm of Chinese company, Alibaba, has trained an AI system to recognise COVID-19. It claims to be 96% accurate in its ability to process over 300 – 400 scans in 20 – 30 seconds. It is said to have aided 26 Chinese hospitals review over 30,000 cases. |
The development of AI as a diagnostic tool remains in the nascent stages and many remain sceptical in its ability to deliver accurate results when used for testing for COVID-19. However, even in solutions where overlapping illnesses cause false positives, quick triage may be of significant aid to healthcare systems.
AI and machine learning can be utilised for COVID-19 diagnosis through various applications. Although AI, and specifically AI imaging, might not provide a stand-alone solution in the fight to beat coronavirus, it can serve as a valuable ‘second opinion’.
The advent of COVID-19 provides a new field for research and opportunities to develop existing technology. Whilst many applications of AI in the response to COVID-19 have greater prominence (e.g. for surveillance and tracking the spread of the virus), utilising the situation to advance AI as a diagnostic tool improves its significant promise to transform healthcare systems and positively impact healthcare outcomes for patients.