Google Trends can improve surveillance of Type 2 diabetes

Abstract

Recent studies demonstrate that people are increasingly looking online to assess their health, with reasons varying from personal preferences and beliefs to inability to book a timely appointment with their local medical practice. Records of these activities represent a new source of data about the health of populations, but which is currently unaccounted for by disease surveillance models. This could potentially be useful as evidence of individuals' perception of bodily changes and self-diagnosis of early symptoms of an emerging disease. We make use of the Experian geodemographic Mosaic dataset in order to extract Type 2 diabetes candidate risk variables and compare their temporal relationships with the search keywords, used to describe early symptoms of the disease on Google. Our results demonstrate that Google Trends can detect early signs of diabetes by monitoring combinations of keywords, associated with searches for hypertension treatment and poor living conditions; Combined search semantics, related to obesity, how to quit smoking and improve living conditions (deprivation) can be also employed, however, may lead to less accurate results.

Figure 1

Structure of the Experian Mosaic Public Sector database.

Figure 2

Medically- (top) and self- (bottom) diagnosed populations in Central London. Maps produced with the Experian Mosaic Public Sector data, using ArcGIS Desktop v.10: ESRI 2011. Redlands, CA: Environmental Systems Research Institute http://www.esri.com/software/arcgis/arcgis-for-desktop.

Figure 3

Medically- and self-diagnosed diabetes variables and EMPS risk variables, used by two main UK models Cambridge Risk Score and QDScore. Darker and lighter bars correspond to medically-diagnosed and self-diagnosed diabetes variables respectively. (*p < 0.1; **p < 0.05). This illustration was produced in R ggplot v2.1.0.

Figure 4

Candidate risk variables, constituting modeling scenarios without (left) and with (right) self-diagnosis variable. All regression coefficients presented have CI 95% and higher.

Figure 5

Correlation propensity between keywords, associated with the risk factors, contributing to model scenarios without and with self-diagnosis variable (ISB), and generic search term ‘diabetes’. Fisher-transformed rs indicate stronger correlation trends for the case of the modeling scenario, comprising self-diagnosis variable (‘With ISB’), leading to the conclusion that self-diagnosis can be seen as a mediator for highlighting more dynamic behavioural and lifestyle risk factors. Data source: Google Trends.

Figure 6

Correlation propensity between keywords, associated with the main risk topics and generic search term ‘diabetes’. Fisher-transformed rs indicate stronger correlation trends for the case of the modeling scenario, comprising self-diagnosis variable (‘With ISB’). Topics consist of the following risk factor components: Poverty (Without ISB): ‘Barriers to housing’, ‘Poor indoors living conditions’; Medical condition (Without ISB): ‘Obesity’, ‘Rheumatism’, ‘Acne’, ‘Eczema’, ‘Anxiety’; Lifestyle (Without ISB): ‘Smoking’; Ethnicity (Without ISB): ‘Pakistani’, ‘Irish’, ‘Celtic’, ‘Black Caribbean’; Poverty (With ISB): ‘Barriers to housing’, ‘Poor indoors living conditions’; Medical condition (With ISB): ‘Insomnia’, ‘Dermatitis’, ‘Acne’, ‘High cholesterol’, ‘Anxiety’; Lifestyle (With ISB): ‘Sometimes/Rarely diet’, ‘Use slimming products’, ‘Trying to lose weight’; Ethnicity (With ISB): ‘Irish’, ‘Somali’, ‘Sikh’, ‘Eastern European’, ‘Black Caribbean’. Data source: Google Trends.