Data-Driven Insight and Prediction

We will develop risk prediction models to predict negative care outcomes such as mortality, increased care needs or hospitalisation. These tools have great potential to ensure the delivery of the right care to the right person in the most cost-effective way. This work package will focus around building on the rapid increase in availability of routine data and improvements in available computing power. Analysis of the new routine data will bring us to new insights into health, vulnerability and care in later life which will enable better predictions of care needs and more effectively targeted interventions.

What are our intentions?

The overall aims of this work package are:

1. To use research survey data to explore and understand how later life trajectories of frailty, wellbeing and social participation relate to each other and are influenced by factors such as housing, wealth, income, care and neighbourhood.
2. To use routine data to develop and validate a set of risk prediction tools, drawing on a range of quantifiable factors of negative outcomes informed by the survey analysis, for use in health and social care delivery.

Why is this important?

Looking back after someone has died, it is relatively easy to see that there are several reasonably distinct routes or trajectories that individuals take on the road to death. It is much harder to predict in advance what an individual’s trajectory or outcomes will be. Risk stratification is an essential element of any complex intervention to maximise function, quality of life and independence in later life. Existing prediction tools, however, are not good enough, lacking either precision in identifying those at risk of adverse outcomes or sufficient validation for confident use in applied settings and few have been deployed as large scale interventions to improve care.

How will we achieve this?

We plan a programme of data-focused work drawing on theoretical perspectives and modelling methods from social science, epidemiology, and machine learning. There will be two strands of related work, drawing on the strengths of survey and routine data as well as statistics and machine learning. Survey data, such as the English Longitudinal Study for Ageing, provides rich detail in terms of health measures and socio-economic circumstances, while routine data has sample size, timeliness and coverage that should be used in risk prediction models to identify those individuals who require some form of intervention. Similarly, statistical techniques offer theoretically informed models, while machine learning offers flexible data-driven insight. We will adopt machine learning techniques to uncover new associations and relationships whilst validating findings using theoretically-informed statistical models.

Our expectation is that:

1. The complementary strengths and weaknesses of data, method and theory will bring new insights to improve effectiveness of existing risk insight models that have so far been found lacking in performance and validation.
2. We make a significant methodological contribution by combining different data and different approaches to quantitative modelling to yield stronger risk prediction models.