How machine-learning can speed up diagnosis and treatment

With the COVID-19 pandemic, where do you see the main opportunities for using artificial intelligence?

First, there is a huge opportunity of AI in diagnosis. For instance, using x-ray or CT images to detect the virus could speed up diagnosis and avoid the need to wait for days for results. Another example is “artificial noses”, which are electronic sensors to detect an illness by smell. This technique is used for diseases such as tuberculosis and provides almost instantaneous diagnosis. There are other ways in which machine learning can help: AI can detect patterns in the voice, to detect breathing difficulties or evidence of recent coughing, or in facial expressions, to detect mental health conditions such as depression.

The current crisis is showing effects in the mental health arena, and AI has a huge role to play in helping us there. Chatbots, for example, could help in combating loneliness among people who may have no contact with friends or family in social distancing situations, and chatbots may also help detect pointers of depression in the speech. But AI and virtual reality are also being largely used to assess, diagnose and even treat mental health conditions such as phobias or anxiety.

Which leads to the role of this technique in medical treatment. AI can help decide when ventilation is necessary and under what conditions. It can help in drug and vaccine design. More broadly, with AI, we could predict treatment outcomes – that would allow us to choose the right treatment for each patient and, ultimately, increase their chances of survival. Last but not least, when we have the pandemic under control, AI could be used, for example, to monitor the long-term implications of this virus, such as cardio-vascular and neurological health conditions.

How close are we to putting some of these solutions into practice?

It takes time. We need to be sure that these technologies really work before we start using them. There are a lot of resources going into this area, and for the moment, we are still at the experimental stage. In the first instance, these technologies may be deployed as a support for medical experts, and as accuracy improves, certain suitable tasks may be partially or even fully automatized. With cancer, there are prototypes for diagnostic tools that are already more accurate than human medical experts. So, it can definitely be done. The big difference is the need of massive amounts of data – that is the drawback. By comparison, humans need much less data to make diagnoses.

How quickly can we mobilize that data?

The question is whether there is sufficient data available for research. Hospitals have a lot of health data, but that is protected by data privacy in most places. We could also collect more data on how this virus is transmitted – but the tracing apps, for example, will not allow us to do that, because they were not designed for that purpose. We are currently researching the possibility of using smaller data sets in machine learning and this is also an active area of research for my own group. One other possibility is generating artificial data that mimics natural data. AI could help detect new outbreaks, using people’s online posts from their social media accounts or Google searches, but again that raises privacy issues. There are opportunities now to put data to good use – opportunities that did not exist twenty years ago. On the data privacy question, we already give a lot of data away on Facebook – there is a good argument for making more personal data available, if people are explicitly willing to do so, to support research.

Thomas Lukasiewicz

AXA Chair in Explainable Artificial Intelligence (AI) for Healthcare, University of Oxford (UK)

Professor Thomas Lukasiewicz is Professor of Computer Science and Head of the Intelligent Systems Laboratory at the University of Oxford in the UK. He is also a Turing Fellow at the Alan Turing Institute in London and a Fellow of the European Association for Artificial Intelligence. His research focuses on AI and machine learning, particularly for healthcare applications. He is also Associate Editor for the Artificial Intelligence (AIJ) journal and the Journal of Artificial Intelligence Research (JAIR).