A novel approach using machine learning to produce Living Evidence Maps for what works to reduce health inequalities in primary care

Problem

There has been a 300% increase in primary care inequalities studies in the last decade, with almost 3000 articles published last year. Systematic reviews, particularly on broad topics such as this, have a limited ability to synthesise this increasing body of literature in a timely fashion without going out of date. Innovative methods are needed to address this problem. Machine learning (ML) software can be used to regularly and accurately identify relevant literature to create living evidence maps. We aimed to create living evidence maps of interventions to address health inequalities in primary care to support the implementation of equity-focused policy and practice.

Approach

Six systematic reviews were used to train a ML algorithm, using EPPI-Reviewer software, to identify studies investigating what works to address inequalities in primary care. Additional studies were identified using network graph searches which use forward and backwards citation tracking and related article searches. The ML software prioritised potential studies in OpenAlex according to their similarity to the training material. The prioritised studies were screened on title-abstract and full text, and manually coded for their intervention type, disadvantaged group and health and care outcomes. Additionally, as part of auditing the tool, manual searches for relevant studies were included and coded. Using EPPI-Visualiser software, the codes for interventions, disadvantaged groups and health and care outcomes were mapped to allow researchers to identify both evidence to inform practice, and gaps in the research. Evidence briefs, aimed at policymakers and practitioners, are then produced based on the living evidence maps. Our living evidence maps are published open access.

Findings

We have included 329 systematic reviews (SR) and 3 umbrella reviews (UR). The disadvantaged groups with most publications were ethnic minority groups (n=210) and socioeconomically disadvantaged groups (n=118). The interventions with most publications included advice and counselling (n=162) and education (n=156). Outcomes most studied included behaviour change (n=98) and diabetes (n=86). Topic maps with the highest number of publications include advice and counselling interventions for ethnic minorities (SR = 107, UR = 2), education interventions for ethnic minorities (SR = 106), community and link worker interventions for ethnic minorities (SR = 86), and advice and counselling interventions for socioeconomically deprived groups (SR = 62, UR = 2). There was no evidence found for gypsy, roma or traveller communities; autism; or ADHD.

Consequences

Our novel ML-driven living evidence maps generate a comprehensive evidence base, more efficient and up to date than traditional systematic reviews. These decrease time spent synthesising evidence, allowing efforts to be focused on action to reduce health inequalities. Our living evidence maps facilitate the production of evidence briefs for policymakers and practitioners, and identify evidence gaps for researchers and funders, highlighting opportunities to address primary care inequalities in research and practice.