The Supervised Learning Requirement

• Everything we’ve done requires a target, a known outcome to train against.
  • Survived - Yes/No
  • Readmitted - Yes/No
  • Length of stay - 4.2 days
• But what happens when you don’t have labels?
• And what if your data isn’t a table of rows and columns at all?

There’s Much More to Machine Learning

• Supervised learning is one tool in a larger toolkit.
• The right tool depends on:
  • Question
  • Data Type
  • Outcome Type

Types of Machine Learning

• Supervised Learning - predict a labelled outcome
• Unsupervised Learning - find structure in unlabelled data
• Recommender Systems - predict what a user will want next
• Time Series Forecasting - predict future values from historical patterns
• Deep Learning - learn from complex, unstructured data (text, images, audio)
  • Natural Language Processing (NLP)
  • Computer Vision
  • Large Language Models (LLMs)
• Reinforcement Learning - learn from actions and rewards
• There is overlap between these different tpes of machine learning (and there are lots of other types too).

When Labels Don’t Exist

• Real scenarios where there is no target column:
  • You have patient records with no outcome flagged, but want to know if there are distinct groups.
  • You have GP referral data and no one has defined what “high-need” looks like yet.
  • You have staff survey responses and want to find response patterns.
• You are no longer trying to predict an outcome.

What is Unsupervised Learning?

• Unsupervised Learning does not need a target variable.
• The algorithm finds latent structure (groupings, patterns, relationships) in the features themselves.
• You are not predicting an outcome; you are describing the data.

Types of Unsupervised Learning

• Clustering - group observations into clusters that are similar to each other and different from other clusters.
  • K-means, hierarchical clustering, DBSCAN
• Dimensionality reduction - find a lower-dimensional representation of the data that preserves its structure.
  • PCA, t-SNE, UMAP

Clustering in Healthcare

• There are lots of examples where clustering in particular can be useful in healthcare.
  • Segment patients by care needs - low-need, moderate, high-need, complex multimorbidity.
  • Group GP surgeries by demand profile - high-attender, stable, growing.
  • Identify similar hospitals for benchmarking.
• Clusters emerge from the data - you don’t define them in advance.

What Are Recommender Systems?

• Predict what a user is likely to want, based on their behaviour and the behaviour of similar users.
• This is a very common type of machine learning that everyone will have interacted with.
• The technology behind Netflix recommendations, Amazon product suggestions, Spotify playlists.
• Two main approaches:
  • Collaborative filtering - users who liked X also liked Y
  • Content-based filtering - recommend items similar to what you’ve already engaged with

Recommenders in Healthcare

• Surface relevant NICE guidelines based on a patient’s profile.
• Match patients to clinical trials based on eligibility and prior enrolment patterns.
• Staff training recommendations based on role, grade, and learning history.
• Recommend clinical decision support resources to clinicians based on case type.

What is Time Series Forecasting?

• Time series forecasting predicts future outcome values using (primarily) past outcome values.
• Data indexed by time - A&E attendances by week, bed occupancy by day, prescribing volumes by month.
• A primary challenge is separating trend, seasonality, and noise from each other.

Common Approaches

• Traditional, Statistical Methods - ARIMA, exponential smoothing
  • Well-established, interpretable, good for single series
• Machine Learning Approaches - gradient boosting or neural networks applied to time-indexed features
  • Often best on large collections of related series
• Prophet - designed by Meta for demand forecasting, blending the traditional with the modern
  • Handles seasonality and holiday effects simply
  • Robust to missing data, trend shifts, and outliers
• NHS examples - elective recovery planning, workforce demand modelling, medicine supply forecasting.

What is Deep Learning?

• A class of machine learning models built from neural networks with many layers.
• Each layer learns progressively more abstract representations of the input.
• Excels where traditional methods struggle - raw text, images, audio, video.
• Requires much more data and compute than classical methods, but has transformed what’s possible.

Natural Language Processing

• Natural Language Processing applies deep learning to text.
• Core tasks:
  • Classification - is this feedback negative?
  • Named entity recognition - find names, entities, organistions etc. in free text
  • Summarisation - condense a long, complex text to key points
• NHS examples - automated clinical coding, patient experience analysis, discharge summary review.

Computer Vision

• Computer vision applies deep learning to image data.
• Core tasks:
  • Classification - is this scan normal or abnormal?
  • Detection - locate and flag an abnormality
  • Segmentation - outline a tumour or organ boundary at the pixel level
• NHS examples - diabetic retinopathy screening, chest X-ray triage, skin lesion detection.

Large Language Models

• LLMs (ChatGPT, Claude, Gemini) are deep learning models trained on vast amounts of text.
• They learn to predict the next token and in doing so, develop broad language understanding and reasoning capability.
• Built on the transformer architecture, which processes entire sequences simultaneously using attention mechanisms.
• LLMs are not a separate category - they are the current frontier of NLP within deep learning.

LLMs in Healthcare

• There are lots of areas where LLMs could be usedd in healthcare.
  • Drafting and summarising clinical correspondence.
  • Answering questions about guidelines and pathways.
  • Coding clinical text to ICD-10 or SNOMED CT.
  • Patient-facing chatbots and triage support.
• However, it’s important to recognise the risks involved with these tools too .
  • Hallucination risk, data governance, clinical liability - none of which are solved problems.

What is Reinforcement Learning?

• An agent learns by taking actions in an environment and receiving rewards or penalties.
• Not trained on a fixed dataset - learns through trial and error over many interactions.
• The goal of reinforcement learning is to learn a policy (a decision rule) that maximises cumulative reward.
• The technology behind AlphaGo, game-playing AI, and robotics (such as self-driving cars).

Reinforcement Learning in Healthcare

• Treatment sequencing - which therapy to give next, given a patient’s response history.
• Sepsis treatment protocols - learning dosing strategies from outcome data.
• Appointment and theatre scheduling optimisation.
• Less common than supervised or unsupervised methods, but it is a growing research area.

Match the Method to the Question

Summary

• There is more to machine learning than supervised learning.
• Machine learning can be applied to a wide range of questions, data types, and outcome types.
• Knowing what is possible is half the battle!
• Identify the question first, and then the data/outcome type. You can then pick the method you need.

Thank You!

Contact:

Code & Slides:

/NHS-South-Central-and-West/code-club

… And don’t forget to give us your feedback.