As someone who has spent his whole career involved in tech companies, I’ve definitely developed a natural curiosity. I love when things that seem like magic actually work as promised, but the feeling of satisfaction is quickly replaced by a thirst to understand the how and why. This has never been truer then since I started working in earnest with Machine Learning and Artificial Intelligence. AI has the potential to completely change our world. While we doing some things at CrushBank with AI that are very cool, I always like to dig into other uses and applications of similar technology. It helps me understand the implications of what we’re doing and sometimes also has the added bonus of better understanding the technology. I thought it might be interesting to look at a real example of something that rates very high in both those categories.
One of the very first lessons I ever learned about AI and something I have been told again and again by experts throughout this journey is that data is king. AI and ML are all about training a computer to identify patterns, connections and relationships; training a computer to predict outcomes based on precedents; and training a computer to simulate human thinking and behavior. Breaking it down and seeing how this works can be really enlightening.
And frankly no industry has seen more investments either, based on the opportunity and potential. Nothing is more top of mind right now than the global pandemic we are still fighting, and most people (whether vaccinated or not) still walk around afraid of what could happen if they became infected. The seemingly random severity of this virus is quite possibly the scariest part of fighting it. Will I die? Be hospitalized? Walk around with a cold for a few days? Or have no symptoms at all? We mostly just don’t know, but not knowing isn’t really an option for most researchers.
This is particularly true in the case of a group of researchers at the University of Copenhagen who are using an AI algorithm to predict with up to 90% accuracy whether someone who is undiagnosed is at risk of dying from COVID-19. The scientists found they could get an incredibly high degree of certainty before a patient is even diagnosed and detailed their process in a scientific study published recently. Looking at how they do this gives us a really interesting view into exactly how this technology works and probably even helps to get us thinking about some other ways – healthcare related or not – that we can leverage the same types of processes to build other models and solve other huge problems.
The researchers in this case studied roughly 4,000 patients who had tested positive. Their first step was to determine the factors that they wanted to account for and analyze in studying the data. They logically focused on a series of external risk factors – things like age, BMI, hypertension, etc. All these data points were used to build the prediction model, which could then predict risk of death at several different stages – at diagnosis, at hospital admission and at ICU admission.
Of the 3,944 patients, 324 died of COVID-19. Breaking that down by gender, they found that all of the men who died were between 73 and 87 years old and had clear underlying risk factors such as high blood pressure and high BMI. Obviously, this is not an absolute, and it’s not overly surprising given what we know 12 months later, but think about the impact if we could have proven this level of accuracy back in April or May of last year. Clearly it could have been very impactful. Not all of their findings were so seemingly obvious. They did determine that organ disfunction, particularly in ICU patients, was a very high marker of potential death. This is the type of insight that is invaluable in developing protocols for triage and treatment of the most severe cases. What’s also interesting and notable is the way the algorithm learns and matures. The researchers have been able to respond quickly to new conditions and data provided by mutating strands, helping doctors react quickly and mitigate potential risks as best as possible as the virus changes and spreads.
Taking a series of known outcomes and training the machine to understand how or why they happened lets us predict what will happen again. It’s not always sexy to break down a mysterious technology into something so seemingly obvious and fundamental, but in this case it’s important because it helps build confidence and reliability into a system. Just saying this machine is going to predict a bunch of outcomes leaves people unsure of what they can and can’t trust. Explaining how and why it works makes the technology real and approachable. Translating that into real-world scenarios really starts the mind thinking: What data do you have in your business and how can it start to resolve large, critical issues?
This post is courtesy of CrushBank CTO David Tan.