Development of a dementia risk detection tool using primary care electronic health records

Problem

Increasing prevalence of dementia, related to an ageing population, is a major healthcare challenge. Primary care is the main route through which individuals are identified or subsequently diagnosed with dementia by a GP or specialist referral services. Evidence suggests numerous risk factors are associated with development of Dementia and many multi-factorial dementia risk models have been proposed. However, predictive ability has generally been quite limited, few have exploited routinely collected general practice data or focused on early detection, none incorporate longitudinal trends in health, and only one has utilised the potential of Machine Learning (ML) methods. We aim to develop an improved healthcare record-based tool for estimating patient risk of developing dementia with a focus on earlier identification of those at risk.

Approach

The Clinical Practice Research Datalink (CPRD) is an anonymised primary care electronic patient record database capturing events from healthcare interactions. We identified patients aged 60-95 years contributing to CPRD between 01/01/2005 and 31/12/2017 along with data on a diagnosis of dementia over the period. A set of more than 150 potential predictors were identified from published systematic reviews, relevant individual research studies, and advice from dementia experts, Model building is using both traditional logistic regression analysis and machine learning (ML) techniques. ML is being carried out by co-investigators at the University of London in parallel to the traditional modelling approach based at the University of Manchester.

Findings

Between 01/01/2005 and 31/12/2007, 2,005,756 adults aged ≥60 years contributed to CPRD and fulfilled inclusion criteria. Of this cohort, 7,621 (3.4%) were identified as having a dementia diagnosis. After assessment for feasibility and reliability, we developed and implemented Clinical Readcode lists for 60 risk factors, broadly classified into demographic and social factors, physical and mental health status, consulting patterns, and treatments received. Additional variables were constructed for many factors to capture longitudinal trends alongside current status. Under univariate analysis, the great majority of factors have demonstrated an association with progression to a dementia diagnosis within a 5 or 10 year period, including most classes of prescribed drugs. Multivariate analysis and full predictive modelling using regression and ML methods is underway and will be presented.

Consequences

Tools to aid in the early identification of people at high risk of dementia are urgently needed. Screening in primary care on the basis of information in the electronic health record has great potential, provided a sufficiently accurate tool can be developed. Such a tool is also greatly needed to identify high-risk individuals for invitation into clinical trials of promising treatments. Success in developing a markedly improved tool for the prediction of dementia may also lead to utilising the same techniques to develop improved risk tools for many other health conditions.