I have emerged, blinking, from the darkness of grant/paper writing purgatory (a.k.a December to March in Australia). It is time to get the blog going again, and to make up for the long gap in posts I’m going to start with the big one. The question I get every time I tell a colleague what I am working on, every time I give a lecture, every time I chat with someone new on social media.
“Will computers make doctors obsolete, and if so, how soon?”
Over the course of the coming few blogposts, I intend to give my best answer to that question. I hope I can do it justice, because I don’t really think it has been adequately explored elsewhere.
This is going to take me a number of posts, which as I have said previously I will endeavor to keep short and sweet (below 2000 words, hopefully). Expect most posts to cover one or two concepts, rather than lay out the entire argument in one giant wall-o-text.
I will try to be as serious and rigorous as I can be, taking the current lay of the land and predicting the future. But I should point out right away that there is a lot of uncertainty in the near and medium term. In the long-term the answer is pretty clear, but in the near or short term … anyone who says they know for sure is trying to sell you something.
The uncertainty exists because we actually have very little evidence to base our predictions on. Exploring the evidence will come a bit later in this series, but for now it is fair to simply say that even those rare few who have expertise in all of the areas that are relevant to the question are struggling to make good predictions.
So what I will try to do is assess the arguments that have been made and present you with the evidence that we have, so you can build your own (hopefully informed) opinion. Like I have said previously, medicine is huge and complex, so the discussion is going to touch on computer science, cognition, culture, regulation, clinical practice and more. Hopefully there will be a little something for everyone*.
Today I just want to introduce the topic. It seems to be a topic about which very few experts are willing to make serious, falsifiable predictions.
Very few, but not none.
Meet Professor Geoffrey Hinton.
Lunatic fringe? He stuck with neural networks through the “AI winter”, but that is a bit much 🙂
As the picture states, Prof Hinton is one of the “godfathers” of deep learning (there are three or four of them, depending on who you ask). He is one of the most respected, knowledgeable and experienced researchers in the field. He is one of the most cited modern computer scientists (Scholar says almost 170,000 citations!). He manages Google Brain Toronto and is the Chief Scientific Officer at the new $150M Vector Institute. His list of doctoral students contains some of the biggest names in the field, including the research director of OpenAI and the director of AI at Apple. Famously, an important technique called dropout, which is widely used in deep learning today, was first described as a footnote in one of his powerpoint presentations.
So when he says something about deep learning, it is a pretty good default position to take his opinion seriously.
And he says this:
“People should stop training radiologists now. It is just completely obvious that within five years, deep learning is going to do better than radiologists.”
Geoffrey Hinton, Machine Learning and Market for Intelligence Conference 2016
Now that is a serious prediction (although Prof Hinton did also say “it might be ten years”). This is the medpocalypse scenario, where entire fields of medicine just disappear in the next five or ten years, presumably with most of the other disciplines close behind. We will refer back to this scenario throughout the series.
As a quick aside for the less medically informed: A radiologist is a specialist doctor, they read the x-rays and MRIs to diagnose you. A radiographer is not a doctor, they operate the scanners. These similar sounding words are a common source of confusion.
Anyway, how do we even judge a claim like this? I mean, this describes an unprecedented disruption of a professional occupation. Never before have we seen whole disciplines cease to exist in less time than it takes to train a new worker (a radiologist has at least ten years of medical training, and usually five in radiology specifically).
To tackle a statement like this, we will need to get some basics out of the way. Over the next post or two I will talk terminology, explain some new concepts, and try to build an understanding of what exactly Prof Hinton is describing.
If you want to brush up on a bit of background information, and see how I generally think about these sort of things, I strongly recommend reading a few of my previous posts. In particular:
Do machines actually beat doctors?
The three phases of medical AI research
These posts will provide some grounding for this discussion, and the fact that my 2017 predictions are looking pretty on the money so far should hopefully afford me enough credibility to keep you reading 🙂
Speaking of credibility, if you want to know more about me, check out About Me. TL:DR, I am a radiologist (a medical doctor!) who does medical AI research (trying to become the other sort of doctor too). Being in both camps, I feel fairly well placed to address this topic.
I am really excited to have finally started this series, and I hope you will all enjoy it as we go forward. I will aim to release a post every week, so this should keep me busy for a while!
List of other posts in the series:
Part 2: Understanding Medicine
Part 3: Understanding Automation
Part 4: Radiology Escape Velocity
Part 6: The Bleeding Edge of Medical AI Research, Google and Diabetic Retinopathy
Part 7: The Bleeding Edge of Medical AI Research, Stanford and Skin Cancer
11 thoughts on “The End of Human Doctors – Introduction”
Very interesting post. Your credibility in medicine and also AI puts you in a unique position to discuss of impact of AI. I was wondering if you would give me your thoughts on this work, where the goal was to predict the ‘diagnoses’ (after the fact) by looking the notes collected by doctors and nurses (it uses MIMIC III dataset).
Condensed Memory Networks for Clinical Diagnostic Inferencing
I think the article is a very interesting one, and definitely a component in medical automation. Messy, natural language data is a major barrier in medical AI, and any work that tackles it is good (and this paper had particularly good results).
We still aren’t there yet with this sort of data though, natural language understanding is a long way behind image understanding in maturity. Lots of work to do!
Awesome. Looking forward to more posts, we’ve been waiting for a while!
Sorry to keep you waiting 🙂
Looking forward to the next post!