What is machine learning?
Machine learning is a branch of artificial intelligence that gives computers the ability to learn from data without being explicitly programmed at every step. Machine learning algorithms can identify patterns in data, generating insights to ultimately aid better decision making.
What is statistical modelling?
Statistical modelling aims to identify and quantify relationships between variables. Models show the strength of the relation between a predictor (such as age) and what we are trying to predict (such as death). Statistical models can make predictions about future values, but they rely on a set of assumptions about the data, for example that the data follows a particular distribution of values.
Machine learning, on the other hand, makes fewer assumptions about the data. This enables it to work with large numbers of variables, but this can come at a cost, as model outputs become harder to interpret. However, much research has been done on this, and techniques to overcome these issues have been developed.
Why is machine learning being used in the health sector, and why are NHS organisations interested in applying it?
The health sector has a wealth of data, but organisations often lack the capability to analyse and interpret it. This is where machine learning algorithms can help – they can be used to analyse vast amounts of data quickly, and often in an automated way, which saves time and money which is important in today’s resource limited NHS. Machine learning is proving valuable in imaging, for example, where it is helping to identify the presence of malignant tumours in scans of breasts, lungs, skin and brain many times faster than clinicians are capable of.
What can machine learning tell us that other approaches can’t?
Machine learning can identify patterns in data that it may not be possible to easily identify through traditional statistical techniques. In some situations, it can also generate more accurate, detailed predictions compared with traditional techniques, particularly when there is a large number of variables.
How does Dr Foster use machine learning?
We are using machine learning for this exact purpose – to identify patterns in the data that we may otherwise miss and to improve the predictions that we make. These range from predicting outcomes for patients, such as the risk of an emergency admission or adverse event, to predicting healthcare cost at a population level. Previously, decisions were often made solely based on human assumptions, but using data provides additional insights and can aid more effective, evidence-based decision making. At Dr Foster, we combine clinical expertise with machine learning techniques to get accurate and clinically relevant models.
How is machine learning providing insights, and what new applications of machine learning is Dr Foster looking into?
We have developed a model using both GP and hospital primary and secondary care data to predict the likelihood of a patient being admitted to hospital in the next 12 months. We are developing a similar predictive model for Lambeth GP Federation using general practice data only. Models like this are usually built using hospital data, as previous emergency admissions are a strong indicator of future emergency admissions, but initial results using only GP data are promising.
We are also developing clustering techniques to identify segments of the population with similar needs and healthcare demands in order to direct resources and services to these segments. This will help improve outcomes and care experience and also reduce per capita costs. Imperial College London has used these clustering techniques to segment heart failure patients into groups of patients with similar utilisation of different health and social care services. This is helping them to determine whether heart failure patients receiving integrated services have better outcomes than patients whose care is delivered by separate departments.
In 2020, we will develop predictive models using machine learning techniques to predict outcomes for hospital patients in close to real time. Outcomes are likely to include delayed discharge, death and readmission. These models will help clinicians in hospitals to make better decisions about their patients in order to prevent these negative outcomes from occurring.
How should we ensure there is a common understanding about the use of machine learning in the NHS?
Forums and conferences about machine learning in the NHS present opportunities to discuss what machine learning is and isn’t, so that there is a common understanding. We hope to continue publishing content around this subject that will help to establish some common ground. Industry groups and networks (such as TechUK) may also be useful to enable suppliers to agree on definitions.
What’s next for machine learning in the NHS?
NHSX aims to accelerate the adoption of artificial intelligence technologies within a standardised and ethical framework, so we should be seeing more widespread use of machine learning techniques across the NHS. This will facilitate more personalised healthcare by generating accurate predictions at patient level, which will enable better decision making and, ultimately, better outcomes.
For more information on the work Dr Foster is doing on machine learning and predictive analytics, email firstname.lastname@example.org