One year ago, I left the start-up where I had been working on an AI-driven companion to accompany patients through their cancer treatments.
When I left, I was deeply frustrated with the start-up environment surrounding AI in Healthcare. I was still convinced that AI could help in this space, but all I was seeing was teams going down what I considered to be the wrong paths.
Marco Schmidt, the CSO of Biotx.ai, and I have published an opinion article in the Pharma Boardroom online trade journal. We have been discussing the topic ever since my first keynote talk at the launch of the Charité Digital Health Forum in 2018 (a video of the same talk is available from PyData 2018).
Something I’ve struggled with on and off over the 20 years that I have been making mathematical models is explaining those models to others. I have tried to bring people along and develop their understanding. But mainly what I observed was that, some people just got it and others did not.
I have certainly improved my own skill at explaining. This comes down to having streamlined stories and simpler take-home messages. Telling a clearer story certainly improves my audiences’ self-satisfaction, but ultimately some of them get the whole message and others do not.
I have mentioned, in the past, that I am a huge fan of Nate Silver. Something which he used to repeat quite frequently, on their podcast, is a sort of predictive modelling tautology:
The best prediction of the future is no change.
Nate Silver [Paraphrasing]
This concept has even got a probabilistic and philosophical theory behind it. All other things being equal, over the long history of time, the next moment from now is not likely to be any different from right now. If we repeat this process often enough then we will be right more often than we are wrong. In essence, we are accepting that there is continuity (and perhaps causality) in our experience of the natural world. Political scientist David Runciman even explored the concept in his recent work of political theory.
I originally took this statement in the manner in which, I hope, it was intended. But behind every great phrase there is often an enticing problem. Thinking over this phrase has led me to realise that there are three basic types of predictive models and each one of them has a fundamentally different purpose and indeed parameterisation.
This is my third attempt, over the course of 9 months, to write this article. The first attempt foundered on my desire to go into detail on whether explanation or explanability is a good characteristic of a model or not. I confess, this was overly motivated by my personal frustration at having worked with somebody who, “never let the facts get in the way of a good story.” The second attempt got lost in a forest of anecdotes from previous projects. I was trying so hard to knit them together that I failed to make a point. Today, I want to focus on the single most important thing that I have learned about developing decision making models.
I am asked quite often how I see Data Science in the biomedical industry. I have, of course, many answers each of which is context dependent. However one theme which I find frequently recurring is a sort of straw-man debate which seems to inherently attract technical practitioners.
The debate is usually structured as follows: How do you see the validation of medical AI products working in practice? Answer: clinical trials, test-validation sets, blah, blah But doesn’t this lead to enormous overheads? Answer: yes, but there are shortcuts But if you take these shortcuts then don’t you run the risk of running into costly failures when you finally run the clinical trials? It goes on….
Working in industrial research is usually very motivating but occasionally it is also frustrating. You’ve just done something really cool but you’re not allowed to tell anybody outside the company about it. Indeed, in a small company there might not be anybody inside of the company who can even appreciate it!
I have worked on roughly 4 really cool projects since leaving academia at the end of 2017. And apart from some basic mentions in my blog (e.g. here and here) most of what I have done has been known only to a few key stakeholders.
I had the opportunity to talk recently with a relatively advanced researcher in machine learning methods. The conversation turned briefly to the study of embeddings when he mentioned that most of his work involves things that can be embedded in Euclidean space. Since I’ve been spending a bit of time thinking about embeddings recently, I asked him some questions to get the official ML take on the subject. I was resonably gratified to learn that – although most ML engineers don’t think much about embeddings – the research on this topic considers the embedding to be tightly bound to the network architecture. It is not possible to study abstract embeddings, divorced from applications. I fully agree with this point-of-view.
Randomised controlled trials (RCTs) have been the gold standard for statistical evidence, of treatment effect, for over 100 years. Their strength is in their attempt to avoid major sources of bias in a comparison of the evidence. However, they are costly to run, particularly in the domain of personalised medicine, to which medical AI products typically belong.