Are AI and Machine Learning just buzz words or do they have some functional use?

What are ‘smart’ machines and why do we need them?

So, our previous article on the avenues of MedTech introduced you to the 6 aspects that we felt most important in Medical Technology. Now we want to dig a bit deeper into each of them so that you can gain a better appreciation for the technology and its application.

First up, Artificial Intelligence (AI) and Machine Learning.

Here’s what we know so far:

• Artificial Intelligence produces smart machines that aim to replicate human intelligence with decision making
• Aritifcial Intelligence can be split into Machine Learning sub-sections and Symbolic Subsections
• Deep Learning is a sub-section of Machine Learning
• Machine Learning uses algorithms and large data sets to make decisions (not fully automated as engineer has to validate the decisions) whereas deep learning can self-validate and is fully automated

Now let’s first explore deeper the basics of the technology.

Machine Learning (ML) enables machines to make accurate predictions and subsequent decisions when fed data, but the code underlying this process hasn’t necessarily been written by a human. Even though we’ve stated that it isn’t fully automated, it is important to recognise that there are different forms of ML:

• Supervised Learning Algorithms use what has been learned in the past (labeled data sets) to predict future events. This uses labeled data sets to forecast outputs and after enough training, provide targets for new inputs. The algorithm then compares its output with corrected outputs provided by an engineer to validate the code
• Unsupervised Learning Algorithms use data sets which have no previous labels and studies how other systems develop algorithms which describe the hidden pattern of the data.
• Semi-supervised Learning Algorithms use both types of data sets (labelled and unlabelled) but often small amounts of labelled data.
• Reinforcement Learning Algorithms use trial and error alongside human validation to eventually build an algorithm that can determine correct output measures

An algorithm simply put is a set of rules which need to be followed to solve a problem

Machine learning in healthcare is becoming more and more important especially with the ever-growing access to clinical data. There are hidden patterns that human analysis may not be able to identify. These patterns may translate to associations that can act as predictors for disease or whether certain treatment methods are more succesful than others.

Deep learning is the sub-section of machine learning and artificial intelligence that makes use of artificial neural networks to classify information from text, images or sound.

I know what you’re thinking. What are artificial neural networks? They sound incredibly complicated.

An Artificial Neural Network is key to deep learning and includes many hidden layers that are sometimes not understood by humans

I thought so too. Until someone explained a neural network to me in terms of a computer being able to identify a hand-written number. Whilst a written explanation can be found in more detail here and a video for it can be found here, I want to try and explain this to you as simply as possible.

Recognising the ten hand-written digits, 0–9, has one major problem. Everyone has different styles of handwriting with everyone using different heights, widths and orientations. Some people even write the digits 1 and 7 similarly for example.

If we say that the input to the neural network is a hand-written number, and hold it within a 28x28 grid (each point of the grid is a pixel).

1. Now let’s say that pixel is allocated a decimal (0.1,0.2 etc…) from 0–1 dependent on the shade of grey from the hand-written number, also known as the ‘activation’ of that pixel.
2. Now, all the activation numbers are fed into the neural network as the input layer, which in turn will cause activation of parts of the first round of hidden layer.
3. This will then cause activation of specific parts of the second round of the hidden layer.
4. Finally, this will cause activation of an output layer (the activation in this case corresponds to how accurate the system believes its answer is).
5. This is how a system takes a hand-written number and converts this into a digital number.

Whilst this is an oversimplified version of how a neural network is set-up and functions, it should give you the basic concepts behind how it works. Essentially feeding more and more refined data through a various number of layers and applying algorithms each time to get an output.

Deep learning has a variety of applications and some can truly advance healthcare. We will explore these in more detail in articles to come, but an example of its application is in cancer detection. Advanced versions of these aritificial neural networks are able to take medical images and detect cancer at earlier stages with less misdiagnosis, providing better outcomes for patients.

There is more to Artificial Intelligence which includes Natural Language Processing (NLP) and Computer Vision. We have decided to leave these for another day because we don’t believe it is fundamental to appreciating AI.

It is however, important to appreciate that AI within healthcare warrants being a field by itself and, that truly understanding it doesn’t come from reading a couple of articles. This document by the Academy of Medical Royal Colleges is a good place to start.

Check out our YouTube page to see short explanations on this and many more topics!