A Critical Need for AI in Healthcare

Today more than ever, we need modern, effective diagnostics. The COVID-19 pandemic and its economic impact are having a significant effects on everyday life. The disease could even become endemic, meaning that each year we might see a "cold and flu and COVID-19 season".
So what do we do?
Do we simply continue to shut down the global economy every time the disease flares up? How do we ensure a good quality of care for people in communities that have scarce medical resources?
These questions, especially the second one, are important, and not just for the coronavirus. We are in dire need to modernize our health care systems, and although this is a complex topic that involves scientific, economic, and political factors, there is one issue where there is no controversy at all. We need better testing and diagnostics.
This is where artificial intelligence comes in. AI technologies are ready to deploy in image recognition and can play a critical role in producing an effective response.
Let's first take a look at two examples where AI is already making a difference, and then we can come back to consider the implications for COVID-19.
The FDA and the EU already approved IDx-DR for diabetic retinopathy, which is a complication from diabetes that can lead to blindness. With the IDx-DR system, a technician can take a picture of the person's eyes, and upload them to a server. The server uses to analyze the images and, within one minute, sends back results indicating that the person is either fine or should be referred to a specialist. The benefits are simple and profound. For the patient, they can get tested for the condition at a much lower cost, at an earlier time, and much closer to home. For the healthcare system, the system will pre-screen patients at any point-of-care setting (a doctor's office), so specialists and other hospital resources can focus only on cases that actually need their intervention.
Diabetic retinopathy is a technical term for a condition with a real, human impact.
Video Credit: IDx
In another example, the Center for Systems Biology at Mass. General Hospital in Boston is applying AI and deep learning techniques for oncology. Cancer diagnostics are used not only as a screening tool to diagnose the disease, but also to measure patients' progress throughout the entire treatment cycle. Although we use the term "cancer" as a single disease, it is a catch-all-term for a very wide range of diseases that vary from organ-to-organ and person-to-person. The Weissleder Lab at Mass. General is currently developing a system that, again offers simple but profound capabilities. This system integrates AI processing with cell cytometry, and it can provide results using only a small sample of cells rather than intact tissue. The sample is easier to take, and the results are faster and more accurate. The device for the analysis itself can be made smaller, and it requires little power to operate, so it can be installed in point-of-care settings to avoid having to send samples to clinical labs. This will provide effective cancer screening to a wider range of the population, both in the developed and in the developing world.

Cancer screening, coming to a doctor's office near you.
Image Credit: Weissleder Lab; Nature Publishing Group.
So now back to COVID-19. The first line of defense against the virus is to have a very aggressive testing program. Some countries, like South Korea have been able to scale their testing infrastructure to meet demand while others have struggled to do so. For the test, a medical technician uses a cotton swab to take cells from the person's throat and nose, and the sample is then tested in a lab for the viral RNA. However, we have one major problem in handling positive tests. For many people who are otherwise healthy, a self-quarantine at home seems to be enough. In others, unfortunately, the disease is severe and lethal. For this latter reason, hospitals are inundated beyond capacity by worried patients, and the high density of patients in close quarters stretches the healthcare system beyond capacity.
It would be great, if we could enhance the tests not just to determine who has the virus, but also to know who will have a mild case and who will need medical attention.
Keep in mind, this next part is mostly speculation, but we hope it can add to a constructive dialog, and possibly to solutions that help us all stay happy and healthy.
Since the outbreak started, researchers around the world have made a tremendous effort to help us understand as much as we can about the coronavirus. We now know that SARS and COVID-19 are somewhat similar. In fact, the actual virus names reflect that similarity: SARS-CoV and SARS-CoV-2.
SARS was first identified in 2003, and the following year, the Department of Anatomical and Cellular Pathology at the Chinese University of Hong Kong published a study on the "sputum cytology" of SARS. The sputum is the mixture of saliva and mucus (spit and boogers, basically) coughed up from the lungs when you have a respiratory infection. The researchers found that the white blood cells in the sputum of SARS patients had very unique characteristics compared to healthy patients.
There is much more than just viral RNA in a sample. This information can hopefully be used to determine, or even predict, the severity of COVID-19 in specific individuals.
In this image, the authors found "morphological changes of macrophages, including (A) cytoplasmic foaminess, (B) distinct vacuoles, (C) multinucleation, and (D) ground glass appearance of the nucleus"
Image credit: G M K Tse, et al.; Journal of Clinical Pathology.

These characteristics can be seen in a microscope, so it should be possible to train AI and deep learning models to find these specific features. The hope here is that with further studies, we would be able to use AI-driven imaging analysis to predict the outcome of an infection, so that medical resources can focus only on those in need. These samples would be fairly easy to obtain, and digital microscopy is widely available. Unlike with RNA samples that have to be shipped to a lab, the images could simply be sent to a server on the cloud, greatly increase the access to testing, and improving the actionable information we can get from the results.