Developing a Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence (MELD)

Background Multimorbidity occurs earlier in the lifecourse among people from more disadvantaged backgrounds.

Wider determinants such as housing, education and work influence development of long-term physical and mental health conditions across the lifecourse. Growing evidence suggests that the sequence of accrual of conditions varies considerably and influences outcomes. Sentinel conditions (the first long-term condition occurring in the lifecourse) may play a particularly important role.

Understanding how wider determinants, sentinel conditions and accrual sequences affect risk of early-onset (by age 50) of complex (four or more conditions) and burdensome (combining mental and physical conditions) multiple long-term condition multimorbidity (MLTC-M) is vital to inform when and how to deploy Public Health interventions.

Data at sufficient scale that links whole lifecourse events, including wider determinants, with clinical information on long-term conditions is lacking.

Development Award Aim To establish the necessary environment, principles, systems, methods and team in which to use artificial intelligence (AI) techniques to 'connect' longitudinal cohort data with routine NHS data and identify lifecourse causes of early-onset complex MLTC-M, to then identify optimal timepoints for public health interventions within the future Research Collaboration.

Development Award Objectives (4 work-packages) 1.Acquire and curate data from the Care and Health Information Exchange Analytics database (CHIA – routine GP data) and the 1970 British Cohort Study (BCS70) in analysis-ready format (WP1&2) 2.Test optimal clustering algorithms in CHIA that identify early-onset complex MLTC-M (WP1) 3.Identify three exemplar burdensome early-onset MLTC-M clusters in CHIA (WP3&4) 4.Identify people in BCS70 with the three exemplar clusters and test their relationship with selected early life determinants using a Directed Acyclic Graph-based approach (WP3) 5.Explore if sentinel conditions and long-term condition accrual sequence can be identified and characterised in CHIA (WP3) 6.Explore whether the nature and determinants of sentinel conditions can be characterised in BCS70 (WP3) 7.Develop AI transfer learning methods that allow extrapolation of inferences from BCS70 to CHIA and vice versa (WP1&3) 8.Build capacity for engagement and co-production of the Research Collaboration structure, objectives and outputs (WP4) Research Collaboration questions: Within the context of early-onset, complex MLTC-M clusters: 1.How is the nature of clusters influenced by the timing and nature of sentinel condition occurrence and timing and accrual sequence of further conditions? 2.How are demographic, socioeconomic, behavioural and environmental determinants acting across the lifecourse associated with age of onset and nature of sentinel conditions, timing and sequence of subsequent conditions, and nature of the resulting cluster? 3.How is risk of developing additional conditions associated with different steps and pathways in the development of MLTC-M? 4.Can transfer learning AI methods within an AI-enabled linked dataset be used to reliably apply causal inferences about early lifecourse determinants of MLTC-M within large routine healthcare datasets to identify critical windows of opportunities for prevention?

Timeline 8 months from January 2021 Impact and Dissemination We will disseminate via academic routes and scale these proof of principle methods within the Research Collaboration by AI-linking other birth cohorts and larger routine datasets. Identifying key time points for public health interventions would significantly influence public health policy.