Using loyalty card records and machine learning to understand how self-medication behaviours vary seasonally in England, 2012-2014

Authors: Alec Davies*, Geographic Data Science Lab, Department of Geography and Planning, University of Liverpool , Mark A Green, Geographic Data Science Lab, Department of Geography and Planning, University of Liverpool, Alex D Singleton, Geographic Data Science Lab, Department of Geography and Planning, University of Liverpool
Topics: Medical and Health Geography, Geographic Information Science and Systems, Geography and Urban Health
Keywords: Self-medication, minor ailments, regression trees, machine learning, XGBoost, Health Geography, Geographic Data Science
Session Type: Paper
Day: 4/4/2019
Start / End Time: 9:55 AM / 11:35 AM
Room: Congressional A, Omni, West
Presentation File: No File Uploaded

Data driven population health surveillance is a core part of monitoring and identifying trends and behaviour in disease and illnesses. Administrative and survey data are traditionally used however they have limitations of cost, resolution and temporal coverage. The increasing self-medication movement brings opportunities to understand the prevalence and treatment of minor ailments. Loyalty card records offer one solution for revealing novel insights about self-medication behaviour whilst addressing these data limitations. We used loyalty card records from a national high street retailer to examine how purchasing of over-the-counter medicines for coughs and colds and hay fever varied in England (2012-2014). Analyses were undertaken at the Lower Super Output Area level allowing exploration of ~300 retail, social, demographic and environmental predictors of purchasing. Gradient boosted trees (XGBoost) were used to predict future monthly purchasing. Clear purchasing seasonality was observed for both outcomes reflecting the climatic drivers of their associated minor ailments. Coughs and colds exhibited wider exposure through higher purchasing proportions. Dynamic models performed best, however where previous year behaviour differs greatly (training data) predictions witnessed higher error. The most important features were consistent across models (e.g. previous sales, temperature, seasonality). Feature importance ranking had the greatest difference where seasons changed. Loyalty card records are shown as a valuable data source to supplement our understanding of self-medication behaviors, through efficient, cheap and objective purchasing data. They offer promise for the monitor of minor ailments prevalence, as well as revealing insights about the drivers of purchasing behaviors.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login