Transforma logo
Subscribe Login

AI gets a mixed scorecard in helping tackle Covid-19

MAR 31, 2020 | Matt Hatton
 
region: ALL sector: ALL technology: ALL

One of the prevailing trends over the last couple of months has been for experts on the technology sector to pitch in with their opinions on the treatment and impact of Covid-19. We don’t do that. We leave the medicine and the epidemiology to the experts. In this current crisis the engineers should focus on building ventilators and perhaps 3D printing masks. However, one of the areas of our coverage that is directly in the firing line is AI and it’s worth digging into the extent to which it has been able to help tackle the spread of the virus. There are some high points, but largely it flags up the inherent deficiencies.

We start naturally with the medical, and probably the #1 ultimate priority is finding a cure. This takes two forms, firstly drug repurposing. Think of the apparent success that the combination of Chloroquine and Hydroxychloroquine seems to have had. Companies such as Benevolent AI are looking at what existing drugs can do and have identified Baricitinib (a drug used to treat arthritis) as a potential treatment. The company analyses scientific literature to identify links between genes, drugs, viruses and transmission vectors and therefore likely effective treatments for Covid-19. Exscientia is undertaking a similar project. Then at the molecular level you have the likes of Insilico Medicine focused on designing molecules to halt replication. Others are using AI to develop new drugs. One process that looked really interesting was the use of reinforcement learning to break down molecules known to act as inhibitors to similar viruses into constituent elements and test against likely characteristics of something that may help against Covid-19. See here for more details.

The main problem is that of finding sufficient training data, which deep learning algorithms need to be effective. The problems here are manifold. There is very little data on the specifics of treatment and diagnosis. Testing is not widespread and the virus is often symptomless. There are also suspicions of improper reporting in some countries. Furthermore, it doesn’t help that the virus is mutating. For deep learning to be effective requires a high degree of continuity, or a hell of a lot more data.

Also immediately pressing (possibly more so) is the need for cheap and effective diagnosis. Typically tests such as Reverse Transcription Polymerase Chain Reaction (PT-PCR) are the standard approach, but these are slow and cumbersome. Much quicker is the use of AI to do image pattern recognition to spot tell-tale signs on scans of the lungs, as done by companies such as Infervision. While it’s not perfect, it is significantly quicker. Furthermore existing data sets from previous viruses seemingly can be used as training data as was the case with training a model at Renmin Hospital at Wuhan University. Ali Baba Group has demonstrated 96% accuracy with its AI application for analysis of Covid-19 CT scans, and was able to deal with one scan every 20 seconds versus 15 minutes for a person. There are promises also of non-invasive diagnosis, but nothing that has yet materialised.

What’s needed to support all of this is a good mechanism for sharing as much data as possible. The Covid-Net Open Access Neural Network, developed by the University of Waterloo and DarwinAI, offers one example, sharing a dataset of thousands of chest scans for anyone who wants to help in its development of an AI tool. One lesson we can take away from this outbreak is to have a better universal source of data under the auspices of the WHO or similar.

AI may have a role to play in predicting patient outcomes once someone is diagnosed with a severe case requiring hospitalisation. There are approaching a million diagnosed cases worldwide. In the EU 30% have been hospitalised. This means we have potentially hundreds of thousands of data points. According to our research at Transforma Insights, less than 30% of AI deployments have access to such a large training set. Of course, the dynamics of every use case are different, so one application might require a million pieces of training data, whereas another might need only a few thousand. However, there is reason to be optimistic. There is a slightly ghoulish question of what such a prediction of outcomes might be used for: is it for resource allocation, targeted pre-emptive action or triage?

The use of AI to support epidemiology has also been attracting a lot of attention, such as through companies like Blue Dot. There are some established areas with existing data sets upon which AI might be trainable, for instance modelling the impact of warm weather on respiratory diseases, or understanding how transmission might occur within hospitals. There is a lot of hypothesis and supposition. What is for certain is that AI is not a magic wand. Especially with epidemiology, but also with the other areas identified, what’s generally needed is firstly some old-fashioned science, and more traditional epidemiological modelling.

Moving away from the medical world. there are a whole load of existing non-medical systems upon which AI will already have been trained that might be tweaked to offer some benefit in dealing with the fallout or mitigating the effects of the virus. Complex systems such as social media, airline booking systems or population movements are well understood and could be used to track trends, for instance predicting how the virus might spread into new territories, or how well government information is being distributed. The challenge is that most of these systems are not currently working within their usual parameters due to the virus’s impact.

Might AI also be used to stabilise financial markets? After all, trading bots are widely used in equity trading. Unfortunately, it’s probably been doing quite the reverse. Throwing a curve-ball at the market the like of which it has never seen before and expecting trading bots to comprehend and act on the information is, shall we say, hopeful. Humans are inherently better placed to cope with the unknown.

Covid-19 exposes a number of weaknesses in AI. For instance, AI does not cope very well with new things, and Covid-19 is very definitely a new thing. By the time we get sufficiently good data for AI to be broadly useful, we’ll probably be over the hump. AI also tends to lack imagination. Human intervention is definitely required to help shape AI responses to novel events.

I suspect that we will learn more about AI from Covid-19 than vice versa.