All articles Business

AI for Machine-Based Health Diagnosis Part 2

Part 2: Learning Techniques - No singular silver bullet

This is the second in a multi-part series on applied Artificial Intelligence for Health Care Diagnosis. This first part covered the basic terms and concepts that are important building blocks for understanding the historic and more contemporary advances.


It is important to understand that today there is no one “magic” AI learning technique that can be applied to all domains and problems. While Hollywood may have you believe otherwise, human judgement on machine learning remains in vogue and necessary. Therefore, a variety of learning techniques are often used by scientists and engineers today within a product offering. This second part of the series drills down into some specific learning techniques and their applications and limitations.

Machine learning is one type of artificial intelligence. Often writing a program to solve a task that a human cannot describe explicitly can be extremely difficult. Ai enables systems to learn and improve based on experience, without the necessity for defining specific rules or pathways. There are 4 types of learning 1) Supervised Learning, 2) Unsupervised Learning 3) Semi-supervised learning and 4) Reinforcement learning. One or more of these types of learning can be used depending on many factors such as the availability of good quality data or the problem that is trying to be solved. Healthcare data is often messy, filled with unnecessary information or is totally unstructured. Selecting the right form of learning for a system particularly in healthcare is always important to consider. Let’s talk more about the 4 types of learning one could adopt.

Supervised Learning

Supervised learning learns by example. As children we undergo supervised learning when we are at kindergarten. For example, we are shown an image of a triangle and told it is called a “Triangle”, or that adding 2 and 2 equals to 4. We are presented with the problem and the solution at the start. The next time we are shown a triangle we already know what it is based on what we had learnt previously. Supervised learning is essentially the same thing.

A training set of data acts as the “gold standard” outcomes that are that we hope to get. The algorithm learns by comparing its own output with the gold standard examples we have previously given to find errors. It will modify the model to adjust for errors it made and so “learns” from experience. This form of learning is normally used when historical data is present, and this data has been labelled with the “correct” answers. This type of data in healthcare is extremely hard to come by. This is because healthcare data often contains multiple variables, changes over time or the “correct” answer is not suitably definable. Supervised learning has been used before to detect breast cancers in x-rays. A labelled set of breast cancers (which have been confirmed as cancers histologically) is used to train a supervised machine learning model. A separate set of x-rays (with the labels not shown) is then presented to the model, and the algorithms results are compared to the correct answers.

Supervised learning can be further divided by how often the system learns (one shot vs many shot) or the form of data used (Regression vs Classification). Supervised machine learning makes judgements that humans can relate to (since we train them). However, these types of models will have trouble dealing with new information. If a system designed to categorize photos of cats and dogs is presented with an image of a horse, it will incorrectly lump the photo into only one of the two categories it knows.

 Unsupervised Learning

Unsupervised learning tasks find structure in data that a human may not be able to. This is because the “correct” answers are not possible to obtain, or they don’t occur naturally (e.g. price of house vs square feet). The data is normally more complex and harder to understand at a glance.

The concept of unsupervised learning is sometimes a difficult concept to get one’s head around. For example, if we present an AI system with a bowl of fruit and ask it to categorize each item in the bowl, this could be done in one of several ways. The system could choose to categorize the fruit based on shape of the fruit, the size or main color, or a combination of a number of those. These categorizations would not necessarily be incorrect.

The algorithm is not provided with a historical sample of labelled data which it is able to reference against. Instead, a large volume of data is presented at once and the algorithm is asked to find structure in the data. Because of the need of vast amounts of data, and programming expertise, unsupervised learning is a more complex process and is used less than the supervised alternative. Unsupervised learning is the type of AI people envisage when they talk about AI “learning”, or true general artificial intelligence.

To develop an unsupervised machine learning system a training set is provided, but not labelled, and a list of outcomes do not need to be defined. The system is totally blind to what the outcomes should be, or how they should be arranged. Instead, the algorithm infers structure by finding trends and associations between the data provided.

Imagine our supervised learning example, except the teacher is not present. You are presented with a bunch of shapes (squares, triangles, circles etc.) but have never been taught about their names, or how they are associated. Using your cognitive ability, one would be able to group these shapes based on the number of edges, or their size or even their color, without any formal prior training. This is a similar to how an unsupervised learning system would derive outcomes.

The more data that is provided, the more accurate the categorizations become (in theory).

Semi-Supervised Learning (SSL)

Think of semi-supervised learning as a combination of both methods above. The reason semi-supervised learning exists is because in many instances the types of problems an AI system is designed to solve require a balance of both approaches. In healthcare reference data to solve a problem is sometimes available, but often it is incomplete or might be inaccurate. Semi-supervised learning is used in this case to help, as it can learn from both the “correct” data using supervised learning and fill in the outstanding gaps using unsupervised.

The volume of unlabeled data outweighs the labelled because it is harder to obtain. There may not be enough data enough data for a supervised learning approach only. A drawback to supervised learning alone is that is often subject to human bias (we normally get humans to label the data). Using SSL is useful for reducing human bias in the process.

 Reinforcement Learning

Reinforcement learning is how we train our pets; when they do something right, we offer treats and when they do something wrong, we do not. Using treats, they learn the desired behavior. In a similar way reinforcement learning in AI rewards correct answers and penalizes for the incorrect. Therefore, it learns without directly being taught by a human. The aim of a reinforcement learning system is to seek the greatest reward and minimize penalty. This is a fairly complex type of learning requiring 3 pieces. 1) The agent (or the decision maker) i.e. the pet, 2) the environment (everything the agent interacts with) i.e. our pets environment and 3) actions (what the agent can do) i.e. sit or stand.

The agent will figure out the best way to solve the problem based on its aim of maximizing reward. The learning occurs via this reward feedback referred to as the reward signal. This form of AI is often used for gaming where the algorithm discovers which steps are required to lead to a maximum reward and win the game. However, this method requires a lot of trial and error and as a result, this process requires multiple repetition and the ability to “fail”. This process is referred to as a Markov Decision Process.

We hope this installation provides a better understanding of what “learning” means in the artificial intelligence setting.  We have talked about the 4 common types of learning methods and how they are adopted.

 And Next…. Part 3 ahead…

In part 3 of our series we will take a closer look into deep learning and what this means in healthcare, and how neural networks fit into this every complex jigsaw.

About the Authors

Dr. Sunir Gohil is both an MD and a PhD in Data Science and is a member of the Clinical AI group at Babylon Health in London.

Brian E. Skiba is a member of the Babylon Health US team and is based in Austin, Texas.