Introduction to Predictive Modelling in Health and Social Care
By Connecting for Health | 2012
Predictive modelling can be used to estimate the probability of a range of future events happening for individuals. These events can relate to an individual’s health or social need. For example:
- Jack has a 1.2% chance of having an emergency admission in the next 12 months
- Jill has a 0.5% chance of developing diabetes
Predictive modelling can also be used to estimate the probability of a range of future events happening within populations. Individuals with similar probabilities within a population can then be grouped together into strata (as is done in the well known Kaiser Pyramid in health care). For example:
- 3.2% of people in Leeds have a 60% chance of having an emergency admission in the next 12 months
Many of these events have a direct impact on the need, form and timing of health and social care delivery. Therefore by being able to predict future events interventions can be planned and executed to:
- Optimise individual health and social outcomes – for an individual ensure they have the best treatment and care in a proactive rather than reactive manner
- Optimise population health and social outcomes – for a care commissioner help balance their population needs with resources so they can drive down health inequalities and target interventions to individuals who are most at risk
- Increase efficiency and effectiveness of health and social service delivery – care providers and commissioners can reduce ineffective delivery by matching delivery to predicted need at an individual level
- Reduce cost of health and social service delivery – care providers and commissioners can reduce cost by reducing inappropriate and ineffective service delivery to individuals who will not benefit while minimising future costs by early identification and intervention for individuals who will benefit
Predictive modelling is therefore applicable in both commissioner and provider contexts, potentially across all health and social care settings and all health and social care professionals.
Predictive modelling technically generates a “probability” of an event happening in the future. A probability ranges from 0.0 (will never happen) to 1.0 (is absolutely guaranteed to happen).
Any probability in-between is a guess or an estimate, it is not a guarantee. The subject of probability is a complex one and surprisingly there is no agreement among statisticians and mathematicians as to what probability actually means or how it should be interpreted – for example the Frequentist and Bayesian schools of probability. A probability value might therefore best be viewed as the likelihood of something happening in the future – likelihood is used here in the common vernacular sense, statistics has its own specific definition of likelihood. A value close to 1.0 is very likely to occur, but there is no guarantee it will.
A value close to 0.0 is very unlikely to occur, but again there is no guarantee it won’t.
For a value close to 0.5 it is very difficult to decide if it is likely or unlikely to occur.
For example the probability of matching all six numbers in the lottery is 0.000000000000005 (but only if you actually buy a lottery ticket), which looks very unlikely to occur. However the probability of flipping heads in a coin toss is 0.5, which makes it difficult to decide if it is likely or unlikely.
Probabilities can also be expressed as a percentage, for example the probability of flipping heads in a coin toss is 50%. Sometimes the word “chance” is used instead of probability.
The term “Risk” is often used when referring to the probability of an adverse or unwanted event happening in the future. In health care, diseases or symptoms of disease would be regarded by most people as an unwanted event in their lives and the probability or chance of them happening would be seen as a risk. A consequence of morbidity is usually treatment and intervention from health organisations. Such treatments or interventions can be viewed as proxies or indicators for the causal morbidity and thus are often viewed as unwanted events in their own right with associated risk. For example most people and health professionals would regard emergency admission to a hospital, whatever the underlying cause, to be an unwanted event.
Probability and risk can be presented as an absolute value or a relative value. For example:
- Jack has a 1.2% absolute risk of having an emergency admission in the next 12 months
- Jack has a 0.001% relative risk of having an emergency admission in the next 12 months compared to the average risk of emergency admission in the next 12 months for a specific CCG/PCT population
When making any statement of probability or risk it is important to be clear about the following:
- The subject the probability applies to – in the examples above it is Jack.
- The event the probability is predicting – in the first example above it is emergency admission in the next 12 months. Most events have some time constraint placed on them. The example is constrained to the next 12 months. Some events have an implicit time constraint, for example developing diabetes is implicitly constrained to the mortality of the individual. For explicit time constraints it is often important to know when the time constraint starts, this is often when the prediction has been made.
- The type of value – a pure probability or a percentage for example.
- The value of the probability/risk.
- If the probability/risk is a relative measure then it is important to identify what it is relative to as illustrated in the second example above.
There are many ways to make a prediction in health or social care. Many health professionals make intuitive predictions based on their own skills and experience. There are many well known issues with reliability, accuracy, repeatability and scalability associated with such intuitive predictions which makes predictions based on some formal model or algorithm preferable.
Such formal predictive models are created by analysing a large set of historical data that contains many different data items that could act as input variables (sometimes called predictors) and the actual associated values of the event being predicted (output variable) for a large number of individuals. The analysis selects the subset of input variables that best seem to predict the output variable values and creates an algorithm or model that describes how to calculate a predicted outcome value based on values for the input variables.
Once a predictive model has been created it can be implemented (coded) into a prediction tool. Input variable values can then be feed in for an individual and the tool will calculate and output the risk of the event happening for that individual.
Was this article useful?13