# Machine Learning: Predicting With Man-Made Rules

So I’ve started taking a different route to work due to the fact that I am working at a different place. I’ve got to say, different lines have a different feel to them. I guess different districts house different types of people so it’s not much of a revelation. I grew up in a small village in Norfolk where everyone knew everyone else. Naturally, you greet everyone you pass. I’ve toned it down a little in London but it hasn’t fully gone away. If someone gives me eye contact and they look like they will be interesting, I spark up a conversation with them as you learn so much. It became apparent that native Londoners are not so enthusiastic after attending dinner when Uber got it’s license application revoked. People at the table where predicting doomsdayÂ  for the company. I pointed out that it wouldn’t be, as there’s too much business made from it. Not just directly with Uber, but banks also profited by giving loans to Uber drivers to buy their cars. The government was just squeezing more money out of them as a second application is a lot more expensive. The people at the table asked how I knew so much about this. My answer was not impressive. I spoke to an Uber driver recently and asked him about it. He would know more about the mechanics of Uber and he also has an incentive to double-check his facts with different sources as it is his livelihood that we’re talking about. If you can brave the odd rejection by a hardened Londoner who makes out that the only reason to speak to a stranger is to try and manipulate someone into giving their kidney away for free, then I recommend it as it adds another dimension to life.

So imagine my delight when I see a guy on the tube who is reading a machine learning book (yes I’m annoying enough to interrupt readers as well). It turned out that he was new to the field and was reading a chapter on how to calculate a price of a house. I told him to be careful, as it’s a problem using man-made rules. He looked slightly puzzled. This actually highlights an important point that people need to be aware of. Calculating house prices is a regular tutorial and introduction to machine learning. This is because you can get really good results without having to do much work. I’ve seen many people do these tutorials, only rush to apply machine learning to medical outcomes. They then get confused as to why the outcome isn’t as impressive.

This is because the rules calculating house prices is defined by man-made rules. Variables such as square foot space, neighborhood, the age of the house, number of bedrooms, position to public transport and so on all contribute in simple ways as the majority of people need to agree on what increases and decrease the price. Therefore, it’s not surprising that a machine learning algorithm can be accurate at calculating house prices. We defined the concept of price, and we came up with some rules that defined the house price. However, let’s say that we want to predict if someone will get cancer. We’re knowing more about cancer every year, but as a society, we didn’t sit down and define the rules of cancer. Not surprisingly, machine learning in cancer is a lot more complex and is not as accurate. I’m not saying that you shouldn’t attempt hard things with machine learning. What I’m saying is acknowledge this difference. If a machine learning algorithm does well in a field where the rules are man-made, be a bit skeptical if people speculate about this outcome in relation to outcomes where the rules are not man-made. With this it’s a completely different story. This is why I do not get excited when computers beat humans in man-made games such as GO or chess. Yes it’s impressive, but the rules are very well defined.

If you want good returns, trying working on applying AI or just simple programming on navigating the man made clinical rules. We should right now be asking what category does this patient fit into as opposed to, can the computer diagnose cancer. If I was to start a computational project tomorrow with access to the NHS systems (NHS bureaucracy pushed me into financial tech as I just don’t want to wait half my life to get a simple project off the ground) I’d focus heavily on data engineering. Establishing well oiled data pipelines and improving the way in which data can be inputted will lead to safe risk calculations. A system that automatically calculates the wells score and alerts you, or sends a notification to your phone when your patients’ bloods are back have clear advantages. However, the bureaucracy involved in achieving these simple wins is just too much, so I understand why many chase machine learning for tough problems. They have careers to build. Working on a sexy pie in the sky idea that is unlikely to be integrated is a lot better for career development in the NHS than fighting bureaucracy for years.

[mc4wp_form id=”5197″]