Nutritional informatics: mining supermarket sales data as a nutritional assessment method

Many nutritional assessment techniques, including food frequency questionnaires (FFQs) and 24-hour dietary recalls have innate limitations such as expensive protocols, high respondent burden, and self-reporting biases. Supermarket sales data have shown promise as a new, indirect, inexpensive nutritional assessment method in recent studies. The goals of the research in this dissertation were to link nutritional content to supermarket sales data and to determine the relationship between supermarket purchases and traditional nutritional measures through correlation and regression analyses. Nutritional data was mapped to sales data at the nutrient and food group levels. One year retrospective supermarket sales data, household food inventory data, and FFQ results were then obtained for 50 households recruited for the study. A correlation analysis was completed to compare percentage of food groups purchased over 52 weeks against food groups in the household inventory and in the FFQ results. Additionally, stepwise regression models were created to predict BMI, energy intake, fat intake, and saturated fat intake based on supermarket sales data. Nutritional content was mapped to 100% of the supermarket sales data at the food group level and at 69% for the nutrient level. The correlation coefficients between the household inventory and sales data over the course of 52 weeks ranged from -0.13 to 0.83 with an average value of 0.23 at week 32, while correlation for the comparison between the FFQ and sales data ranged from -0.17 to 0.47 with an average of 0.23 at 32 weeks. 5 The regression models to predict BMI, energy intake, fat intake, and saturated fat intake each yielded significant results for several food group purchases from the sales data. Mapping nutritional content to sales data was successful, given that there are potential strategies to increase the linkage for nutrient data. The correlation results are in line with other studies comparing nutritional assessment methods against each other and the regression models produced many significant food groups that are substantiated by multiple studies. Overall, the work presented gives an excellent starting point for further informatics research into the untapped potential of supermarket sales data as a nutritional assessment method and public health tool.