Large-scale and high-resolution analysis of food purchases and health outcomes

4.1 Datasets

4.2 Estimating eating habits of an area

To estimate the eating habits of people living in an area, we pool together all the food items purchased by its residents and look at the nutritional properties of the average item. We do so by defining six metrics below.

For the sake of reproducibility of our analysis, we publicly share our data aggregated at MSOA level.

5.1 Relative abundance of nutrients

The nutrition guidelines by the World Health Organization (WHO) recommend to limit the relative energy supply derived from each nutrient within specific ranges [54]; for example, fats should contribute no more than 30% to the total intake. By plotting the distributions of the \(f_{\mathrm{nutrient}\text{-}\mathrm{calories}}@a\) values across neighborhoods (Fig. 1), one sees that Londoners, on average, buy a healthy share of protein, yet they buy unhealthy nutrients (e.g., sugar, fat, saturated fat) more than the recommended limits, and carbohydrates and fibres less than the recommended amounts. The extent to which residents collectively depart from recommended limits changes across the city and is defined as \(\mathrm {departure}\text{-}\mathrm{nutrient}@a\):

$$ \begin{aligned} \textstyle\begin{cases} | f_{\mathrm{nutrient}\text{-}\mathrm{calories}}@a - \max_{\mathrm {nutrient}}| &\text{if } f_{\mathrm{nutrient}\text{-}\mathrm{calories}}@a > \max_{\mathrm{nutrient}}, \\ 0 &\text{if } \min_{\mathrm{nutrient}} \leq f_{\mathrm{nutrient}\text{-}\mathrm{calories}}@a \leq\max_{\mathrm{nutrient}},\\ | \min_{\mathrm{nutrient}} - f_{\mathrm{nutrient}\text{-}\mathrm {calories}}@a | &\text{if } f_{\mathrm{nutrient}\text{-}\mathrm {calories}}@a < \min_{\mathrm{nutrient}}, \end{cases}\displaystyle \end{aligned} $$

(8)

That is, an area’s departure from the nutrient’s recommended level is zero, if the fraction of area a’s total calories coming from nutrient is within the recommended min-max band for that nutrient (i.e., within \([\min_{\mathrm{nutrient}}, \max_{\mathrm {nutrient}}]\)). The departure from the recommended levels of, for example, fats and sugars are mapped in Fig. 2.

To dwell on the health impact of such departure, we now match food purchases with health outcomes and start to answer our seven research questions.

We compute the Spearman rank correlation between disease prevalence (as per Formula (1)) and all the food-related metrics (as per Formulae (2)–(7)). As shown in Fig. 3, calorie consumption is strongly correlated with cholesterol and hypertension, while calorie concentration is strongly correlated with diabetes (RQs1+2). To check whether the relationships between calories and chronic diseases are linear or not, we produce a set of x-y plots arranged in three columns and nine rows in Fig. 4. Each column corresponds to one of the three chronic diseases, and each row corresponds to one of the features derived from our food purchases. For example, in the first row, we have plots that relate hypertension, cholesterol, and diabetes to calorie concentration; in the second row, instead, we have plots that relate these three diseases to calorie consumption. In both cases, we see that, as calories increase, the prevalence of any of the three diseases increases, as one would expect. To ease interpretation of these x-y plots, we rescaled the x-axis. For example, we normalize the average item’s weight in area a as follows:

$$ \mathrm{relative}\text{-}\mathrm{item}\text{-}\mathrm{weight}@a = \frac {\mathrm{item}\text{-}\mathrm{weight}@a- \mu(\mathrm{item}\text{-}\mathrm{weight})}{\mu(\mathrm{item}\text{-}\mathrm{weight})}, $$

(9)

where \(\mu(\mathrm{item}\text{-}\mathrm{weight})\) is the average weight across all areas. If the rescaled value is zero, then the area’s weight is equal to the average value in London. If the value is 0.1, then the area’s weight is 10% higher than the average value. By observing, for example, the resulting plot related to calorie concentration (first row in Fig. 4), we see that, as the consumption exceeds the average value (i.e., \(x>0\)), the per-capita prevalence of any of the three diseases considerably increases.

We also see that the four main nutrients are associated with the three diseases in expected ways: carbohydrates, fat and sugar are positively associated with the three diseases, while fibres are negatively associated (RQs3+5) (Fig. 3). Indeed, the prevalence of each of the diseases increases as carbohydrates (fourth row in Fig. 4) increases. The relationships between the diseases and fat (third row in Fig. 4) is not as clear cut as one would have expected though: the relationships with fat and saturated fat come with high variability (RQ4). By contrast, more proteins (sixth row) and fibres (seventh row) are associated with lower disease prevalence.

To go beyond individual nutrients, we also find that both nutrient diversity (mapped in Fig. 5) and average item weight show high correlations, which of course have opposite signs: item weight (which correlates with calorie consumption) is positively correlated with disease prevalence, while nutrient diversity is negatively correlated (RQs6+7). Indeed, in Fig. 4, one sees that the prevalence of any of the three diseases rapidly decreases with nutrient diversity, while it considerably increases with calorie concentration and average item weight. This is further confirmed by the quadrant in Fig. 6, which places areas according to the prevalence of pairs of nutrients (computed with Formulae (3), (4), and (5)), and colors them according to the prevalence of diabetes (computed with Formula (1)): residents of the City (central London) and Chelsea (West London), who tend to be highly educated and well-off, consume fibres and diversify nutrients; those of Newham, which is a deprived yet rapidly developing neighborhood in East London, consume considerable quantities of calories, do not diversify their nutrients, and end up with a high prevalence of diabetes; interestingly, the residents of Hackney, which is a deprived yet highly-educated neighborhood in East London, enjoy healthier eating habits (i.e., low consumption of carbohydrates and high nutrient diversity) and do not suffer from diabetes as much as Newham’s residents do.

So far we have considered our food-related metrics individually. However, these metrics are not orthogonal, and the presence of one is generally associated with the presence of another. In fact, by correlating the presence of a nutrient with the presence of another (cross-correlation matrix in Fig. 7), we see that an item’s average weight (on the first row of the correlation matrix) is generally not related to any nutrient, as one would expect; carbohydrates (second row) are associated with calorie concentration and sugar (sugar is indeed one type of carbohydrate); high calorie concentration (third row), in turn, comes with food high in carbohydrates, fat, and sugar; and nutrient diversity (last row) is generally found in food high in proteins and fibres.

5.2 Predicting medical prescriptions from nutrients

As a next step, we go beyond studying correlations and aim at predicting the number of prescriptions from the food data. To do that, we first need to account for the dependencies between nutrients. Also, we should account for any factor other than nutrients that impacts health outcomes. The literature typically controls for socio-economic conditions, which have been shown to be a proxy for access to knowledge and capabilities, including access to nutritional information and physical exercising [55]. To account for all these aspects, we use linear regression analysis. The prevalence of each of the three chronic diseases is the outcome variable of an ordinary least squares regression, while our food-related metrics and a set of control variables are the predictor variables. Where necessary, predictor variables undergo a logarithmic transformation, and in addition, we apply a min-max rescaling of each variable in the range \([0,1]\), which allows us to assess the relative influence of each factor: the larger the absolute value of the coefficient associated to a feature, the higher the relative importance of feature in predicting the outcome.

To better make sense of the predictive power of our features, we run a sensitivity analysis where we measure the \(R^{2}\) values for regressions run with different feature sets (Fig. 8). First, we note that demographic features alone are considerably less predictive than nutrients alone, although they improve the overall accuracy when combined with nutrients. Second, and more interestingly, the prediction performance of only the combination of nutrient diversity and calorie intake is comparable to the more complex combination of all the individual nutrients.

6.1 Main results

All the previous results suggest that, in London, socio-economic conditions matter far less than what people eat. As opposed to having high levels of education or of median income, eating less calories and opting for a diet with diverse nutrients are both strongly associated with healthy areas. Indeed, one of the surprising results from the regression analysis is that income is not a significant predictor. Previous studies have shown that income is correlated with general health conditions including mental disorders, self-reported bad health, and lower chances of long-term survival [57]. Yet not all diseases equally interact with socio-economic variables. In contrast to the prevalence of cancer, the prevalence of illnesses connected to the metabolic syndrome such as circulatory diseases or high cholesterol does not significantly change across socio-economic groups [58]. Indeed, previous population surveys point out that, after controlling by education, the link between metabolic syndrome and other socio-economic factors such as racial background [59] or income [60] weakens markedly, which is in line with our results.

In terms of which nutrients matter the most, as opposed to unhealthy areas, healthy ones tend to buy more fibres and far less carbohydrates (including sugar). Also, it is less about calorie consumption and more about calorie concentration, which have been previously found to lead to forms of addiction [10]. By combining all the predictors together, one obtains a model that is not only descriptive of how health outcomes are associated with food purchases but also predictive of such outcomes: it turns out that, from food purchases, we can accurately predict whether an urban area will suffer from diabetes, for example.

6.2 Theoretical implications

This study has two main theoretical implications. The first is a call for enlightened nutrition research. The question for food companies is how to continue to make money even as they cut calories. The answer might come from a shift in how food companies have approached the formulation of their products so far. Food product research has focused its attention on taste, not nutrition. That needs to change. The combination of nutritional research with recent advances in biomedical research promises to create foods that are not only delicious but also provide concrete medicinal benefits.

The second theoretical implication has to do with studying how entire neighborhoods eat. Past research has explored two relationships. The first is between “where people live” and their health: income inequality, unemployment rates, and education have all been shown to relate to people’s health. The second is between “what people eat” and their health: nutrition research has long tested associations between eating patterns and health outcomes with survey data. A third relationship transitively follows which has never been tested before: the relationships between “where people live” and “what they eat”. We have now tested it and found that, indeed, healthy neighborhoods eat less and diversify nutrients more than what neighborhoods suffering from chronic diseases tend to do.

6.3 Practical implications

This study suggests practical implications for a variety of stakeholders. We have shown that unhealthy products have a negative impact on community health. The bad news is that many unhealthy products are very popular. The good news is that as many as five stakeholders have incentives for change: food companies do not wish to be seen as the cause of people’s obesity; insurance companies (especially those in life insurance) have our health at heart; technology companies are entering the digital health market; governments want to be seen to act; and local communities increasingly want to be empowered to tackle their own needs.

Food Companies. By simply cutting out bad ingredients, adding good ones or introducing new products, the food industry could reformulate their offering and elaborate plans to improve nutrition.

Insurance Companies. There is at least one corporate sector that benefits from keeping people healthy: insurance companies. Our study encourages new partnerships between insurance firms and large grocery retailers on, for example, data sharing initiatives. Also, retailers could make anonymized purchased data publicly available and launch “hackathons”. These are meetings in which participants are asked to come up with a solution to a problem within a day or two, and some of the teams generally offer effective solutions at little cost (the winning team is typically awarded a prize).

Technology Companies. Predictive analytics and wearable sensors will transform how people manage their health. A smartphone app might be able to warn users that, based on which foods they share on social media and what their wearable sensors measure, they will exacerbate a heart condition. The app could even suggest which foods to eat—foods that are both pleasurable and nutritious.

Local Communities. We have shown that, by mining publicly available prescription data, we are able to identify (un)healthy areas. Mining digital health data reflects concrete health outcomes and might well benefit local communities by enabling residents to hold local authorities to account. Additionally, a city health monitor could help assess the benefits of implementing different policies.

6.4 Limitations

Sample bias. Our data comes from one grocery retailer and concerns only those customers who have opted for the use of a loyalty card. Furthermore, the data is anonymized and does not contain personal information such as age or gender. Future work should use additional geographic data to quantify sample biases.

Limited explanatory power. Our study does not fully explain health outcomes, and rightly so. Our food data does fully cover Londoners’ food consumption, and our study does not include any data on another important predictor of health outcomes: physical activity.

Average product. From our data, we cannot reconstruct the dietary habits of individual customers and, as such, our results reflect the dietary habits of an area in terms of the area’s “average product”.

Causality. Our results do not speak to causality. Though the causal direction is difficult to determine from observational data, one could consider different temporal snapshots of both sets of data (food purchases and medical prescriptions), and perform a cross-lag analysis.

6.5 Conclusion

It was healthy and adaptive for our primate brains to drive us to eat carbohydrates and sugar when only wild grass was at hand. However, carbohydrates (including sugar) are now readily available at every corner. People living in areas of London with higher prevalence of medical conditions linked to the metabolic syndrome seem to surrender to their human instincts and end up buying carbohydrates and sugar to a considerable extent. By contrast, people living in healthy neighborhoods seem to counter their evolutionary adaptation and buy considerable quantities of fibre. This difference in purchases is not explained by socio-economic conditions: income does not matter as much as one expects. By transcending conventional class boundaries, human biases, instead, seem to be the main obstacle to healthy eating. Our study suggests that the “trick” to not being associated with chronic diseases is eating less what we instinctively like (by not listening to the dopamine rushes in our brains), balancing all the nutrients, and avoiding the (big) quantities that are readily available.

In the future, we will explore the impact of two additional factors on health outcomes. The first is the city itself: certain city’s forms are more appealing to pedestrians than others and, as such, one might wonder which forms are “healthier”. The second factor is exercising. We are exploring the possibility of capturing exercising levels across an entire city with wearable devices. Too see why this is important, consider that, by exercising (even a little), an individual boosts his/her immune system, achieves a 20-50 percent reduction in sick days in the short term, and reduces the risk of chronic diseases in the long term [63].

In our cities, food is cheap and exercise discretionary, and health takes its toll. Technology could change that. With modern data analytics, the availability of new open data, recent advances in persuasive computing, and ever increasingly miniaturized health wearables, modern technologies are now best positioned to help people counter the dopamine rushes coming from sugar and fat, eat better, and exercise more.