Unless you’ve been living under a rock, you will have heard about machine learning. Bizarrely enough, the attitude to machine learning in medicine is just as interesting as the machine learning theory itself. Ignorance, combined with emotion, data fatigue, and desire for people to ride the bandwagon has fostered an everchanging, overgrown diverse jungle of ideologies, opportunism, and politics. If you’re confused, worried about the future, or would just like to make more sense of your social media feed when people post about machine learning, he’s 3 facts you should know to help you navigate through the overgrowth:
Machine learning isn’t the same as AI
Let us start with the most basic fact in order to define what machine learning is. Machine learning is not artificial intelligence. Some people may think that I’m splitting hairs because machine learning is extensively used in AI, but the distinction is real and should be understood.
The first time I practically applied machine learning was in my postgrad when I was working on 3D mapping in surgical robotics. I used machine learning clustering algorithms to filter out the noise that the motion sensor was picking up. That’s right, my first application of machine learning had nothing to do with AI. Machine learning is a model, which means that you can use it to enhance your decision making, without it actually making any decisions. A model is an equation where you feed it some inputs, and it will make a prediction based on it. When I left clinical to study physics, it became clear that physics was mathematical modeling of physical systems. If you want to throw a ball, you’d put in some parameters, and use Newton’s laws with differential equations to predict where the ball would go. Physics gets more complicated with statistical mechanics and quantum theory but essentially it’s the same thing. Machine learning uses a range of models (neural networks, random forests, support vector machines etc), where the weights are refined by regression techniques to try and make the most accurate predictions.
Machine learning gets more complex with reinforcement learning (where it tries to “win” in a simulation), but there’s nothing stopping a clinician to use machine learning to give a probability of an outcome to aid the decision of the clinician.
Machine learning doesn’t strip you of power
Let’s think about the previous point. Machine learning is a model that predicts something and tells you. Do you think this sounds a little freaky? Do you feel like you’re losing control by being influenced by models? Well, you shouldn’t as you already use models in your decision making. The most simple one is speed multiplied by time to get the distance. No doubt you’ve utilized this (even if someone else has done it for you) to make an assessment of how long your journey is going to take. Did you lose power in your decision making? Of course not. In fact, your decision making is empowered with this model. You have to actually think about more things. With the information that this model has given you, you can look at the time you have, and decide if you want to take the longer more scenic route. You can even delve into the question of if the trip really worth your time? Because of this model, you can look at different modes of transport as transport providers would have also used this model to give you rough times of how long the journey would take. Mathematical models have shown again and again throughout history to empower humanity.
Let say you develop a neural network that predicts the probability of a patient having a heart attack (this is already a reality, we can already beat doctors at predicting heart attacks: link). Medical academics can use it to highlight patients who are at high risk and look into correlations that suggest why these patients are high risk. GPs could use it to profile their population, helping them design services that could address their population, and work with high-risk patients to prevent heart attacks. Hospital doctors could also profile their patients. Combining this with audits, and analysis of outcomes and incidents, they could refine protocols, referral pathways, and management plans.
There’s a lot of hype and drama
Because machine learning has lead to a lot of success, there is no shortage of people who talk about it even though they have no experience in it. Practical application is better than just understanding the theory. Some really gritty nuances hit me in the face when I started developing and applying machine learning algorithms for a central London financial tech company. Sadly, I have seen no shortage of people who have never implemented a machine learning algorithm or read the underlying theory behind it giving strong opinions on what machine learning is and what it holds for the future. Because of the lack of training in the traditional academia of the NHS, I found UK healthcare to be the worst offender for this. There are some really good pockets in the NHS where people are really forward thinking. But be warned. There are areas in the NHS where shallow opportunists are so rife, actually knowing about machine learning will go against you. I know clinicians who went back, did math orientated degrees, learned how to code, and then left the NHS because they were constantly pushed out, only to see multiple clinicians who knew nothing, pushed into “innovation” positions because they did some vague “leadership” or “management” course/scheme.
It’s not all doom and gloom. I recommend the department of surgery and cancer at Imperial College London. These guys are very forward thinking. They stay clear of hype, and they are willing to get their hands dirty with coding. They are also willing to teach medics coding [link], learn and develop their own skills. It’s definately a department to watch. It’s a bit of a beacon of hope between the hype and rubbish that gets pushed daily in the NHS. I hope there’s other departments like this out there, and if you know of any please contact me. The problem with departments like the one I’ve linked is that they’re too busy working on projects whilst the opertunists are travelling all over the UK talking about it. Hopefully they can develop a bit more of a unified voice in future. In terms of seeing what a machine learning model looks like, here’s a good link that explores the math behind the most well know neural network [link].
personal mutterings clinical machine learning Imperial college london imperial college London computational medicine machine learning in medicine what clinicians should know about machine learning will machine learning replace doctors
I help clinicians get to grips with coding and tech, I also code for a financial tech firm