Utility of machine learning to identify research for updating a systematic map of evaluations of public health interventions within community pharmacies

ID: 

277

Session: 

Poster session 2

Date: 

Monday 17 September 2018 - 12:30 to 14:00

All authors in correct order:

Stansfield C1, Stokes G1, Thomas J1, Dickson K1, Rees R1
1 EPPI-Centre, Institute of Education, University College London, UK
Presenting author and contact person

Presenting author:

Pynver Fgnafsvryq

Contact person:

Abstract text
Background: An 18-month update search for a systematic map of public health interventions within community pharmacies yielded over 23,000 citations. The map has a broad scope in terms of any public health intervention delivered within a community pharmacy setting, balanced with a focus on outcome evaluations in Organisation for Economic Co-operation and Development (OECD) countries and process evaluations in a UK context. Rather than place further constraints on the search strategy, we applied machine-learning to reduce the workload in screening the resulting records.

Aims: To describe results from applying machine-learning to facilitate prioritised screening against the eligibility criteria for the map and to assess the workload saved by not screening lower-prioritised search results.

Methods: We developed a classifier based on the screening decisions of the original map, and applied it to the citations from the update search in order to generate relevance scores for each citation. The citations were screened in ranked order. We reflect on approach taken, determine precision of relevant citations within the relevance score, and measure the time taken to screen 1000 references of references with relevance scores between 16% to 19%.

Results: We show the included citations in the map relative to the relevancy scores generated by the machine classifier and calculate a total screening reduction. Some challenges with analysing the data relate to differences in processing conference abstracts and duplicate citations.

Discussion: Machine-learning offers the potential for citations from update searches to be ranked in terms of potential relevance, and citations to be screened within available resources. Conceptual issues include uncertainty over the volume of citations to be screened, and a shift from applying limits on the volume of citations from searching to screening. Process issues include developing screening algorithms and citation management, and prioritising studies with a lack of abstracts.

Conclusions: The case study adds to methods in analysing the conceptual shift from searching to screening when undertaking study identification.

Patient or healthcare consumer involvement: The systematic map informs policies on the impact, accessibility and cost-effectiveness of community-based public-health interventions for healthcare consumers.

Relevance to patients and consumers: 

The methods are used to improve the currency of research described in a systematic map. The map describes research evaluations of public health interventions within community pharmacies. The map informs policies on the impact, accessibility and cost-effectiveness of community-based public health interventions for the healthcare consumer.