Assessing the external household exposome from consumer purchases: The My Purchases cohort

doi:10.21203/rs.3.rs-2996021/v1

Download PDF

Article

Assessing the external household exposome from consumer purchases: The My Purchases cohort

https://doi.org/10.21203/rs.3.rs-2996021/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 07 Dec, 2023

Read the published version in Scientific Reports →

You are reading this latest preprint version

Background: Consumer Purchase Data (CPD) has emerged as a promising instrument to assess the impact of purchases on health. Here we describe a comprehensive prospective collection of CPD, the My Purchases cohort. The cohort included 428 participants as of April 18, 2023. Over eight years of CPD collected, 2,058,689 products were purchased, comprising 210,203 unique products. We matched 87.7% of all products by volume with nutrients, ingredients, or additives from at least one database. We found a median rate difference of 0.26 when we compared kilojoule information from generic and specific product matches. Further, increasing the number of retailers that provide CPD improved the stability of individual CPD profiles. Combined with extensive product databases and access to key health outcomes, CPD could approach a “consumerome” that could provide the basis for extensive investigations of how what we buy affects our health.

Health sciences/Health care/Public health/Epidemiology

Health sciences/Health care/Disease prevention/Lifestyle modification

Health sciences/Health care/Disease prevention/Preventive medicine

Health sciences/Health care/Nutrition

Despite significant advancements in our understanding of human health, uncovering the lifestyle and environmental drivers of disease and aging will require new data resources in the years ahead (1). While large external exposome contributors to lifestyle-related diseases and reduced lifespan, such as tobacco, alcohol, lack of exercise, and high-calorie diets have been identified, less is known about other exposures, such as food products, sanitary products, and chemicals used in our households (2). Indeed, the causes of many inflammatory diseases remain elusive, spurring a need to understand the combined effects of the many exposures we encounter throughout our lifetimes (3).

Household purchases represent a significant proportion of these exposures. Research into consumer purchases, hereafter referred to as consumer purchase data (CPD), has increased dramatically in recent years. Although attempts to standardize reporting of these data have been made, a great disparity in naming conventions remains (4). Additional barriers to using CPD include concerns about data safety and security, personal privacy, data accuracy issues, and consumers not understanding the reason for data linkage or not using the services that produce these data (5).

Though still in its infancy, CPD has also been used to benefit patients directly by aiding outbreak investigations and harbors a large potential for further utilization provided residents and retailers consent to data collection (6, 7). Most often, researchers only have available CPD from one supermarket chain’s loyalty program or scanner data supplier and limited product information beyond item number and name are available for analysis (8). Despite this, the CPD can be analyzed with a limited, but often highly informative purpose, such as providing realistic mixtures of chemicals for chemical risk evaluation (9). To date, research interests have mainly been focused on diet, with recent validation against dietary surveys, aggregated use of time trends, and investigations into local food environments (10–12). A recent review concluded that “electronic sales data have the potential to transform the dietary assessment and worldwide understanding of dietary behavior” and that validation studies are warranted (10). Previous validation studies have focused mainly on data from major retailers and scanner data that require manual scanning by the consumer (8, 11, 13–15). A prerequisite of these validation studies is the ability to sample a sufficiently large portion of purchases made and correctly match purchased products to relevant generic or specific product databases, moving from one type of exposure from one retailer to a broad assessment of many types of exposures from many retailers (16). Such a hypothetical full coverage of all purchases made in a lifetime would approach a “consumerome”, mimicking the phenotypic variation of other “omics” contributing greatly to the assessment of the external exposome. Here, using the My Purchases cohort as an example, we describe a comprehensive prospective collection of CPD. We lay out potential possibilities to match products to generic or product specific databases and address the impact of having access to data from more than one retailer, before providing directions for future research.

Study design and population

The My Purchases cohort is a cohort with ongoing recruitment since June 2021, in which Danish residents 18 years and older can donate information about their purchases for research purposes. In addition, the cohort features the possibility to include embedded cohorts and the option for participants to opt out of specific projects or the project in its entirety (17, 18).

Data sources

Recruitment

Since June 2021, a GDPR-compliant, secure, encrypted web application developed by Human Exposome Assessment Platform (HEAP) researchers has been in operation (19). The solution allows consumers to provide legal consent to the retrieval of their CPD from a range of retailers (including three of the five largest retail chains in Denmark) as well as contact information and basic self-reported disease information at the time of consent (20). The CPD is accessed through an application interface, providing a secure means of data transfer.

Integration with Danish national digital signature solutions ensures the valid consent and secure identification of participants, including retrieval of the Danish national Personal Identification Number (PIN) (21). Using the PIN, the collected CPD can be linked to health outcomes from national registers or specific research projects, as detailed in the consent notice. Participants may choose to limit their consent to one research project or provide broader consent to all health-related research and may further opt-in to be contacted again by the researchers.

Currently, cohort recruitment is mainly passive and still ongoing. Participants are recruited via a link after a short introductory text in the data sharing setting of the digital receipts provider app, which includes a prompt to join, for the second time participants log-on. In addition, recruitment is boosted by embedding links to the cohort landing page in questionnaires from other studies (17, 18). Further recruitment efforts using targeted recruitment by letter are planned for additional projects. The solution is available for further exploration and participant registration with the use of a Danish digital ID (https://mineindkob.ssi.dk/) (20).

Raw data collected from receipts

The CPD collected included information found on receipts such as price, discount, purchase date/time variables, and basic product information for all purchases made with debit/credit cards in connected retail chains. The basic product information included a descriptive product item name e.g. “apples 6-pack” and the item number of the product. The item number corresponded to either a shorter supplier-generated barcode, or more often, the main world standard, the unique Global Trade Item Number (GTIN13) barcode. For a full list of all variables available, see Table S1.

Matching specific product databases to purchased products

Direct matching was possible as available GTIN13 barcodes allowed for the further enrichment of products by matching data to specific product databases. Current agreements allowed us to match CPD with product information from the Global Standardization Organization GS1 (trade-sync), allowing for comparisons between different product sources (22). Further, a collaboration with the Danish Consumer Council enabled matching of products from Kemiluppen (23). Finally, we matched data to the Open Food Facts database using the itemnumber (24).

GS1 is a global standardization company that maintains large country-specific supplier-maintained product databases, with a range of product information, including itemnumber, itemnames, ingredients, nutritional contents, storage information, allergens, and many other variables relevant to health (22).

Kemiluppen is a product database with information on itemnumber GTIN/barcode, product names and ingredients and a three-level risk score from. The risk score spans from A, where the product is free of a number of problematic substances, to B the product contains, for example, perfume that can cause allergies or substances that can harm the environment, to C, the product contains potentially problematic ingredients i.e. hormone disruptors. Danish consumers can scan products in the application and receive information on the possible health effects of personal care products, and cleaning products (23).

Open Food Facts is a crowdsourced database of food products with ingredients, allergens, and nutrition facts (24). The database was started in 2012 and has been more extensively described elsewhere (25).

Matching a generic product database to purchased products

Matching to generic product databases required a preprocessing step that removed words and abbreviations from the product item name that could later interfere with partial word matching/regular expression (regex) matching steps. In addition, general product information that could be derived from the product name i.e., product weight, volume, origin, or type, e.g., organic, was identified and extracted for later use.

After preprocessing, each product item name was matched to an adapted version of the Frida Food Database (hereafter referred to as Frida) (26). The adapted Frida database contained columns of regular expressions representing brands, product type, specific and more generic words, taste/flavors, and fat content, which matches more than 1000 mutually exclusive product types and some extra, mainly non-food categories that were not part of the original Frida database. The regex-matched database is designed to help select the product item names that best matched the product type, using a scoring algorithm to flag unmatched products.

Since not all products were matched directly using the GTIN, misclassification could occur. Therefore, as a quality control measure, greedy matches, ambiguous stems (if any), and products not reaching thresholds (unmatched products) were repeatedly subjected to manual reviews and word stems were altered accordingly and iteratively to improve matching. In addition, for each product with information from more than one data source, i.e. GS1 and Frida, we compared nutrient information before finally performing a manual assessment of a random sample, as detailed in the statistics section.

Statistics

In order to quantify the effectiveness of the applied regular expression matching of products and databases as well as to assess product variation between generic and specific product information, we compared nutrient information from products which were matched to both Frida and GS1. The relative difference in nutrient information for each product and for each nutrient was estimated as the absolute difference between the two nutrient values obtained from Frida and GS1, respectively, divided by maximum of the two values.

In addition, an analysis was performed to evaluate the accuracy in estimating an individual's purchase pattern when using a varying number of retailers. Focusing the analysis on the assessment of having data from only one, two or four of the largest retailers in the database, for each individual we calculated the proportion of purchases from each product group ("bread", "milk products", "vegetables, raw", "red meat", "alcohol" etc.), and compared it to the proportion when using information from all available retailers. The distribution of these differences across all individuals was used to evaluate the variation in estimated purchase pattern by number retailers used and product group. All possible combinations of retailers were used, in the analysis of 1, 2 and 4 among the 4 largest retailers in the database.

R version 4.2.2 was used to conduct simple descriptive statistics (27).

Ethics

Currently, many consumer loyalty programs mainly focus on using data to increase loyalty and revenue, and consumers accept this in order to receive discounts or special offers. Despite this, many consumers hesitate to share purchase data with researchers (5).

Though fundamentally different in purpose with a focus on health, this study addresses the concerns of some consumers who worry about losing control of their data, as the participants in the My Purchases cohort as part of their consent, are given the option to limit data sharing and participation to only one study (see results section).

The My Purchases cohort, obtained consent from all participants and was approved by the Data Protection Agency officer at Statens Serum Institut (SSI) (journal no: 21/00949). The study is compliant with all guidelines and regulations, ethical, and IT-security requirements, and requires no further approval procedures under Danish law.

Data availability

Data on participants are not publicly available but agreements regarding the transfer of code can be made by contacting the author, and a framework enabling data access is underway, either at Denmarks Statitistics or a similar secure platform such as planned by HEAP(28). As several data sources are under license, direct access to these datasets is conditional on agreements with the data provider e.g. GS1 data is available from GS1 but requires a trade sync agreement. The product enrichment pipeline matching products to generic Frida product categories is available at https://github.com/ssi-dk GitHub repository. The raw data is currently not public available in order to secure confidentiality of retailers and participants according to their consents. However a simulated dataset (that do not reproduce the main results, but allows other researcher to run the code) will be made available on the CSC SD connect platform (https://sd-connect.csc.fi), and analysis off data is possible through transfer of code upon reasonable request by contacting the corresponding author.

Between June 22,2021 and April 18, 2023, a total of 428 participants were recruited. Of these, 409 (95.6%) consented to all registered studies and of those, 263 (61.4%) also approved participation in all future studies. 64 (15%) did not consent to the sharing of data in an identifiable form beyond SSI. No participants consented to participation in less than two studies. Eight (1.9%) participants had one or no receipts registered in the data and were excluded. General information regarding the number of participants, products bought, total number of receipts, and self-reported disease are shown in the flowchart in Figure S1.

Characteristics of participants

Table 1 shows basic demographics such as age and sex as well as retailer “loyalty” ranging from shopping in a single retailer to attending a range of retailers. Shopping frequency varied greatly with an interquartile range (IQR) of 5.2–12.5 and an average of 8.3 receipts per month, with receipts being available from about half of the cohort for the past 5 years or more. The median total expenditure of each participant per month was roughly 1,900 Danish Kroner (253 Euros [per March 30, 2023]).

Table 1

Age, sex, purchase behavior, and follow-up distribution for the 420 cohort members.
Characteristics and purchase behavior	N = 420
Age^a	49 [38–59] (19–86)
Sex
Male	189 (45%)
Female	231 (55%)
No. products^a	3,516 [1,462-7,078] (18–29,143)
Average product price (DKK)^a	21.7 [18.6–26.5] (10.9-516.9)
No. receipts^a	350 [155–644] (3 − 1,618)
Receipts pr. month^a	8.3 [5.2–12.5] (1.1–35.6)
Money spent pr. month (DKK)^a	1,867 [1,242-3,111] (158 − 10,010)
Enrollment time^a
< 1 Year	40 (9.5%)
1–2 Years	51 (12%)
2–3 Years	24 (5.7%)
3–4 Years	34 (8.1%)
4–5 Years	35 (8.3%)
+ 5 Years	236 (56%)
Inactive time (months)	8 [0–22] (0–76)
No. merchants used^ab	13 [9–17] (1–27)
^aReports Median [IQR] (Min-Max)
^bDefined as whole months with zero purchases.

Receipt coverage over time

The provider of digital receipts has been present in the Danish market since 2014 and has increased the number of unique retailers from 15 in 2015 to 34 in 2022. In the My Purchases cohort, data from 133 participants was available as far back as 2015 and has since risen to 401 in 2022). Throughout the study period, the annual number of receipts (5,816 to 39,060), products (11,131 to 42,090) and total products bought (77,935 to 432,351) increased, reflecting the rise in the number of participants generating receipts, as well as wider coverage of retailers and the introduction of new products to market. Out of the 210,203 unique products purchased, the distribution of products purchased was skewed towards a few frequently purchased products; thus the 1000 most frequently bought products accounted for 39.1% of all products purchased (Figure S2).

Product enrichment

After collection, all data underwent an enhancement step where each product was matched to specific or generic information coming from different data sources including GS1, Frida, Open Food Facts, and Kemiluppen, which combined, provide information regarding: product type, ingredients, nutrients, intended chemicals, and more. GS1, Open Food Facts, and Kemiluppen are matched using GTIN number whereas Frida uses word-recognition to the product name.

The entire product enrichment pipeline is outlined in Fig. 2 and enables the retrieval of information beyond the product item name and item number for 69.5% of all unique products, and 87.7% of the total amount of products purchased. Of these 17,273 products were directly matched to specific products from either GS1 or Kemiluppen using the GTIN code.

Differences in nutrient values between generic and specific databases

Figure 3 shows the cumulative distribution of the relative difference for all unique products (and total products) for Kilojoule (KJ), fat, protein and carbohydrate content. Table S3 shows the distribution of the relative difference in KJ for all unique products found both in the branded GS1 database and the generic Frida database. The median relative difference was 0.26, but the median varied greatly between product groups.

Differences in proportion of purchases by number of included retailers

Furthermore, Table 2 shows the median (10th; 90th percentiles) change in the estimated proportion of purchases from each product group when using only information from 1, 2 or 4 of the largest retailers compared to using information from all retailers. The median change was-8.5% (-100%; 81.9%) when using data from only one retailer and 2.2% (-14.6%; 19.1%) when using data from four retailers for the most frequently bought product categories. For less frequently bought categories such as tobacco, dental products, and cleaning products, the median changes were − 60.3% (-100%; 150.1%) and 4.3% (-25.2%; 31.1%), respectively, when using data from one retailer versus four retailers.

Table 2

Median (10th; 90th percentiles) change in the estimated proportion of purchases by product group when using only information from 1,2 or 4 retailers compared to using all information
	Number of retailers
Products Group	One	Two	Four	N products
Fruit, raw	-4.5 (-100; 59.9)	2.6 (-46.5; 41.5)	3.4 (-3.9; 18.4)	162,602
Vegetables, raw	-3.6 (-100; 44.8)	2 (-42.2; 32.4)	2.8 (-5.5; 17.6)	156,402
Milk products	-5.5 (-100; 47.1)	1.6 (-54.5; 36.3)	2.8 (-9.7; 16)	135,447
Non-sugary drinks	-5.5 (-100; 89.4)	-2.7 (-51.7; 52.6)	-0.9 (-19.1; 11.5)	102,439
Confecture, chocolate, ice cream, and sugar	-2.9 (-90; 92.3)	-1.8 (-38.8; 55)	-0.8 (-17; 9.7)	94,314
Bread	-7.4 (-100; 75.3)	-0.7 (-60.2; 50.2)	1.5 (-15.5; 14.9)	80,292
Butter, oil and dressings	-4.7 (-100; 69.6)	-0.2 (-53.3; 42.8)	1.2 (-14.8; 12)	58,497
Processed red meat	-6.2 (-100; 68.6)	2.8 (-69.7; 49.7)	4.1 (-5.3; 19)	49,068
Cookies, crackers, biscuits and cakes	-11.5 (-100; 116.9)	-2 (-61.4; 69.6)	0.2 (-20.6; 19.6)	45,784
Poultry	-6.7 (-100; 72)	1.8 (-65.6; 50.1)	3.2 (-7.2; 21.1)	44,826
Cheese	-7.4 (-100; 67)	0.6 (-58.4; 43.3)	2.2 (-12.4; 13.5)	44,323
Red meat	-8.5 (-100; 80.6)	1.3 (-64.9; 56.1)	2.5 (-11.9; 19.1)	41,523
Fish and other aquatic animals	-12.8 (-100; 92.6)	1.1 (-78.1; 65)	2.9 (-12.5; 22.2)	35,345
Ready meals	-6.6 (-100; 107.7)	1.9 (-69.3; 67.5)	3.1 (-9.5; 23.7)	35,302
Fruit products	-11.3 (-100; 89.9)	1.6 (-80.5; 63.6)	3.3 (-10.2; 24.9)	34,781
Vegetable products	-10.1 (-100; 96)	-0.3 (-69.8; 69.2)	2.1 (-15.2; 22.9)	23,095
Alcoholic drinks	-34.1 (-100; 127.8)	-13.2 (-100; 81.3)	-2.5 (-43.9; 11.4)	23,083
Grains	-13.1 (-100; 95)	2.6 (-100; 70.6)	4.5 (-9.1; 26.5)	22,487
Rice and pasta	-16.4 (-100; 102)	1.8 (-100; 76.4)	4.5 (-7.8; 29)	19,167
Eggs	-18.6 (-100; 78.5)	-2.5 (-100; 55)	1.5 (-21.8; 15.7)	16,346
Total of most frequent groups	-8.5 (-100; 81.9)	0.4 (-66.7; 55.1)	2.2 (-14.6; 19.1)	1,225,123
Soap and Hygiene	-66.9 (-100; 156.7)	-19.1 (-100; 122.3)	1.8 (-44.5; 24.4)	16,267
Tobacco	-100 (-100; 178.2)	-7.9 (-100; 136)	4.3 (-25.3; 35.6)	11,978
Cleaning products	-38.5 (-100; 129.7)	-9.5 (-100; 77.8)	-1.9 (-35.8; 17.6)	8,445
Dental products	-63.1 (-100; 141.6)	-20.3 (-100; 95)	-2.8 (-52.1; 15.1)	7,124
Total	-60.3 (-100; 150.1)	-3.5 (-100; 115.7)	4.3 (-25.2; 31.1)	1,548,585

Manual assessment of product matching to generic product database

We also extracted 1000 random products and assessed the product matching to Frida manually. A total of 70.4% of products were exact matches i.e. “skimmed milk” and “skimmed milk”, whereas 80.6% were group matches i.e. matching milk correctly but not the subtype of milk.

Self-reported diseases and conditions

In total, 187 (43.7%) participants reported at least one disease or condition, with 39 (9.1%) reporting two, 25 (5.8%) reporting three, and 27 (6.3%) reporting four or more. Diseases and conditions most frequently reported were back pain (n = 31), hypertension (n = 21), type 2 diabetes (n = 21), irritable bowel disease (n = 20), allergy(any) (n = 18), Crohn's disease (n = 15). Conditions reported by more than five participants are shown in Table S2.

Here we describe one of the most comprehensive prospective collections of consumer purchase data (CPD) to date. More than 400 participants have been registered to date and additional recruitment is ongoing, enabling continuous prospective collection of CPD. Though this number may not seem large, we have collected data spanning up to 8 years of CPD, currently from 34 retailers, including several large supermarket chains, as well as one generic food database and two specific product databases. The various data sources enabled broad exposure assessment of otherwise hard to obtain information on consumption of grocery products including, food, tobacco and alcohol, sanitary products and chemicals.

We found that comprehensive enrichment is possible, which both quantifies known product groups contributing to disease, such as smoking and alcohol consumption, and allows for broad exposure assessment of otherwise hard-to-access exposures, such as indoor pollutants (i.e. candles and sprays), and chemicals in cosmetics (9). Though we report on self-reported diseases here, the continuous CPD collection may be linked to health outcomes, enabling register-based longitudinal analysis of how lifestyle related factors are associated with disease onset, propagation, or cessation.

All 210,203 unique products sold to participants could, in principle, cause disease, as is often the case in foodborne outbreaks (6). Apart from the most commonly bought products and products with high brand recognition, it may be challenging to identify such products using questionnaires as only some participants would remember. If other diseases or changes in disease trajectory are triggered by the purchase of a single product, a raw CPD sample with high coverage could thus be superior to questionnaires, as was the case in a recent simulation study (7). We found high dispersion in proportion of purchases from each product group, when data from only one retailer was used compared to using data from four large retailers. All product groups, including important groups such as tobacco, had a relative fractional difference of 100 at the 10^th percentile when using data from only one retailer, resulting in possible misclassification of purchase behavior, such as classification of an individual as an apparent non-smoker, when in reality, the individual is buying cigarettes at another retail chain. Our results illustrate that having data from only one retailer increases the risk of measurement error, and that this was reduced as more retailers were added, an important finding for future studies comparing consumption patterns to data from i.e. food frequency questionnaires.

Information was retrieved beyond the product item level for 87.7 % of all products sold, enabling analysis beyond the product item name or item number and allowing for the assessment of ingredients, nutrient information, and many other exposures of relevance to health (11, 13, 16). As products were matched to Frida and a range of custom-made, non-food categories using regular expressions, some mismatches were expected. Moreover, product-specific discrepancies, such as in caloric content within the same product group due to differences between average product values in the food database and the specific brand (e.g., different spread products) were expected. We quantified differences between the average product values of generic products in Frida and the GTIN specific product information from the GS1 database. Here, we found the median relative difference for KJ to be 0.26, which is higher than what was previously shown in a large British study where products were manually compared (29). These findings are only partially explained by generic versus branded product differences, and some variation is due to misclassification of products e.g. low versus full-fat products or residual errors in the matching algorithm. Manual assessment of 1000 randomly selected matched products revealed that 70 % of products were good matches to generic products, and 80% matched the groups on the groups level i.e. misclassifying the milk type i.e. skimmed versus full milk, but getting milk right. This highlights the need for continuous improvement of the matching algorithm, as the current framework allows for and encourages further development, as well as the use of other approaches including natural language processing. More promising is direct matching to structured specific product databases such as GS1 Tradesync, Kemiluppen, and Open Food Facts (22-24). The unique GTIN13 maintained by GS1 enables direct matches, but the coverage of these databases may vary by country and for GS1, the coverage is producer-dependent, where some product producers do not allow research-related access. For the foreseeable future, efforts probably need to combine specific and generic product databases, where supplier generated product information and information from specific product databases should be preferred over generic information.

Here we combined name matching and GTIN matching to create a near real-time enrichment pipeline for purchased products. Combined, this framework allows us to follow key health determinants, such as tobacco, alcohol, dental products, and diet over time, and thereby investigate time-varying exposures, possible associations with health outcomes, and resultant targets of interventions (17, 30, 31). Though the outcomes described here were self-reported, the consent given by participants allows researchers further to enrich the cohort with information from Danish registers. The Danish registers provide access to outcomes such as prescription data, microbial and laboratory results, including calprotectin, C-reactive protein, and cholesterol, outpatient and hospital visits, and surgical procedures, in addition to a number of key health and economic outcomes as well as spatial and social data. These registers have been extensively described elsewhere (32-35). In addition to register data, other research projects may add further data to the cohort, enabling the future assessment of onset, flare-ups, alleviation, and cessation of a wide range of diseases (17, 18).

Limitations

This study has a number of limitations. First, the age and sex distributions of the cohort differ from the Danish population as a whole, with the cohort being older and more female, highlighting the risk of selection bias. However, the potential for analytically dealing with this bias is promising due to the large amount of longitudinal data for both the cohort and the total population (14). Other challenges include purchases being made at other retailers, loss to e.g. food waste, the level of detail in product grouping, and the knowledge gaps regarding what is being consumed and by whom (e.g. eating out, at friends’, and within a household) (14, 36). Though identifying individual consumption is a challenge, Danish registers allow for identification of household information including single households where this is less of an issue(14). Further our database did not include all retailers and loss of data due to delays in the participant's updating of the digital receipt app and when credit/debit cards expire and are renewed is a challenge and the most likely cause of the average 7 months without purchases found in this cohort. This is also reflected in the lower-than-expected total expenditure of 1,868 Danish Kroner per participant per month, compared to the mean expenditure of 3,164 for each Dane published in 2019 by Denmark Statistics (37). In addition to the lack of complete consumerome coverage, finding available product data beyond what is reported via receipts is also a challenge. Here, 39% of unique products and 13% of all the products by volume bought by participants were not matched. Further, though a larger amount of information is available for generic products, many other key variables, such as weight or volume, are only available to a limited extent from product names and need structured product specific databases such as GS1 to be assessed (14, 36).

Strengths

Major strengths of the current database are the broad number of retailers providing the sample, the ability to ascertain the impact of having CPD from more than one retailer, the length of follow (up to 8 years of CPD), the ability to enrich and classify the majority of products that combined enable the ability to investigate multiple, hard-to-assess exposures over prolonged time frames with minimal time or effort required from the participants and no social desirability bias or recall bias. Many of the technical limitations mentioned above could be addressed by further improving the product enrichment pipeline, including adding Natural Language Processing approaches, providing services that encourage individuals to provide CPD for the entire household, or by adjusting the data for household composition that could work in tandem with targeted questionnaires to address identified knowledge gaps (14, 36).

Another strength is the ability to use the personal identification number provided by participants to collect key covariates from Danish registries, including household size, income, education, and social data, in addition to other external exposome data and health outcomes, as detailed above.

Implications

The My Purchases cohort combines consumer purchase information with health outcomes. To ensure large-scale collection of CPD, creating services that provide insights to participants, while addressing the need for information, choice, and appropriate safeguards for participants is evident (38). Options such as being able to select/deselect various categories of research and selecting different transaction data/CPD data sources could improve participants trust in the recipient organizations (39). In the future, CPD could enable post-marketing, epidemiological assessment of products and help unravel the commercial determinants of health, including health effects of additives and foods introduced to consumers (40). Such information may then inform residents, politicians, and key institutions, such as European Food Safety Authority (EFSA) and European Chemicals Agency (ECHA) of such effects (41)(42). With time and in combination with other sources, the exposures to biological pathways at different life stages and identification of early signs of health damage caused by environmental factors could be discovered and real-time lifestyle advice ameliorating or preventing the impact of these factors could be directly communicated to consumers.

Increasing the number of retailers that provide CPD improves the stability of assessment of individual CPD profiles. Combined with extensive product databases, this could provide a broad assessment of individual exposures from commercial activities approaching a “consumerome”, that could in time provide the basis for investigations of how what we buy affects our health. Future studies enabling analysis of the impact of the consumerome on health should:

1) Establish large cohort(s) with participants that consent to the continuous transfer of individual/household level CPD, and retrieve person-level information about health outcomes and trajectories as well as enable the analysis of the impact household consumer exposures may have on health.

2) Further develop near real-time pipelines, identify the products sold, and enrich each product to enable broad assessment and identification of product type, ingredients, nutrients, intended or unintended chemicals, and standardized reporting of key metrics concerning product matching efficiency, completeness of data and retailer coverage.

3) Develop analytic methods utilizing the multiple time-varying exposures, taking into account the complex longitudinal data structure, including repeated measurements and missing data.

Thus, additional research and funding are needed to further explore the potential of the consumerome to uncover novel lifestyle disease associations and to translate potential targets of prevention into safer products and actionable lifestyle advice. The authors call on key retailers across the EU (and worldwide) to support research access to CPD, as the number and thus coverage of retailers in the CPD samples are crucial to assess consumer exposures accurately.

Acknowledgments

The ability to categorise products builds on the work of others and we thank Kemiluppen, GS1, Frida food for the datasets provided. We further thank the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 874662. We finally thank Jakob Mathias Melbye for his work on setting up databases, and Liis Staarkopf, for her work on the product data.

Funding statements

Frederik Trier Møller etc. has received funding from European Union’s Horizon 2020 research and innovation programme under grant agreement No. 874662

Tune H Pers was supported by the Novo Nordisk Foundation Center for Basic Metabolic Research, and partially funded by an unconditional donation from the Novo Nordisk Foundation (grant no. NNF18CC0034900).

Conflicts of interest:

All authors: None declared.

Kalia V, Belsky DW, Baccarelli AA, Miller GW. An exposomic framework to uncover environmental drivers of aging. Exposome. 2022;2(1).
Afshin A, Sur PJ, Fay KA, Cornaby L, Ferrara G, Salama JS, et al. Health effects of dietary risks in 195 countries, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. The Lancet. 2019;393(10184):1958-72.
Wild CP. Complementing the genome with an "exposome": the outstanding challenge of environmental exposure measurement in molecular epidemiology. Cancer Epidemiol Biomarkers Prev. 2005;14(8):1847-50.
Jore S, Braae UC, Trier Møller F, Friesema I, Paranthaman K, Jalava K, et al. A common framework for using and reporting consumer purchase data (CPD) in foodborne outbreak investigations in Europe. Infection Ecology & Epidemiology. 2022;12(1):2007828.
Clarke H, Clark S, Birkin M, Iles-Smith H, Glaser A, Morris MA. Understanding Barriers to Novel Data Linkages: Topic Modeling of the Results of the LifeInfo Survey. J Med Internet Res. 2021;23(5):e24236.
Møller FT, Mølbak K, Ethelberg S. Analysis of consumer food purchase data used for outbreak investigations, a review. Eurosurveillance. 2018;23(24):1700503.
Dougherty PE, Møller FT, Ethelberg S, Rø GØI, Jore S. Simulation and identification of foodborne outbreaks in a large supermarket consumer purchase dataset. Scientific Reports. 2022;12(1):11491.
Nevalainen J, Erkkola M, Saarijarvi H, Nappila T, Fogelholm M. Large-scale loyalty card data in health research. Digital health. 2018;4:2055207618816898.
Stanfield Z, Addington CK, Dionisio KL, Lyons D, Tornero-Velez R, Phillips KA, et al. Mining of Consumer Product Ingredient and Purchasing Data to Identify Potential Chemical Coexposures. Environ Health Perspect. 2021;129(6):67006.
Jenneson VL, Pontin F, Greenwood DC, Clarke GP, Morris MA. A systematic review of supermarket automated electronic sales data for population dietary surveillance. Nutrition Reviews. 2022;80(6):1711-22.
Vepsäläinen H, Nevalainen J, Kinnunen S, Itkonen ST, Meinilä J, Männistö S, et al. Do we eat what we buy? Relative validity of grocery purchase data as an indicator of food consumption in the LoCard study. Br J Nutr. 2021:1-24.
Fernandez ID, Johnson BA, Wixom N, Kautz A, Janciuras J, Prevost S, et al. Longitudinal trends in produce purchasing behavior: a descriptive study of transaction level data from loyalty card households. Nutr J. 2022;21(1):67.
Smed S, Tetens I, Bøker Lund T, Holm L, Ljungdalh Nielsen A. The consequences of unemployment on diet composition and purchase behaviour: a longitudinal study from Denmark. Public health nutrition. 2018;21(3):580-92.
Vuorinen AL, Erkkola M, Fogelholm M, Kinnunen S, Saarijärvi H, Uusitalo L, et al. Characterization and Correction of Bias Due to Nonparticipation and the Degree of Loyalty in Large-Scale Finnish Loyalty Card Data on Grocery Purchases: Cohort Study. J Med Internet Res. 2020;22(7):e18059.
Konttinen H, Halmesvaara O, Fogelholm M, Saarijärvi H, Nevalainen J, Erkkola M. Sociodemographic differences in motives for food selection: results from the LoCard cross-sectional survey. International Journal of Behavioral Nutrition and Physical Activity. 2021;18(1):71.
Sørensen KK, Nielsen EP, Møller AL, Andersen MP, Møller FT, Melbye M, et al. Food purchases in households with and without diabetes based on consumer purchase data. Prim Care Diabetes. 2022;16(4):574-80.
Attauabi M, Madsen GR, Bendtsen F, Wewer AV, Wilkens R, Ilvemark J, et al. Influence of Genetics, Immunity and the Microbiome on the Prognosis of Inflammatory Bowel Disease (IBD Prognosis Study): the protocol for a Copenhagen IBD Inception Cohort Study. BMJ Open. 2022;12(6):e055779.
Sørensen AIV, Spiliopoulos L, Bager P, Nielsen NM, Hansen JV, Koch A, et al. A nationwide questionnaire study of post-acute symptoms and health problems after SARS-CoV-2 infection in Denmark. Nature communications. 2022;13(1).
Merino Martinez R, Müller H, Negru S, Ormenisan A, Arroyo Mühr LS, Zhang X, et al. Human exposome assessment platform. Environ Epidemiol. 2021;5(6):e182.
https://mineindkob.ssi.dk/
Pedersen CB. The Danish civil registration system. Scandinavian journal of public health. 2011;39(7_suppl):22-5.
https://www.gs1.dk/
https://taenk.dk/kemi/plejeprodukter-og-kosmetik/kemiluppen-tjek-din-personlige-pleje-uoensket-kemi
https://world.openfoodfacts.org/data
Chazelas E, Deschasaux M, Srour B, Kesse-Guyot E, Julia C, Alles B, et al. Food additives: distribution and co-occurrence in 126,000 food products of the French market. Scientific Reports. 2020;10(1):3980.
Fødevaredata (frida.fooddata.dk), version 4.2, 2022, Fødevareinstituttet, Danmarks Tekniske Universitet
R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.eea.europa.eu/ds_resolveuid/U0MQDN9JRG
Van Veen E, Boeckhout M, Schlünder I, Boiten J, Dias V. Joint controllers in large research consortia: a funnel model to distinguish controllers in the sense of the GDPR from other partners in the consortium [version 1; peer review: 1 approved]. Open Research Europe. 2022;2(80).
Williams G, Cade RD, Beer S, Cade JE. UK food composition data: a comparison between branded and generic products in myfood24. Proceedings of the Nutrition Society. 2022;81(OCE5):E174.
Mönninghoff A, Fuchs K, Wu J, Albert J, Mayer S. The Effect of a Future-Self Avatar Mobile Health Intervention (FutureMe) on Physical Activity and Food Purchases: Randomized Controlled Trial. J Med Internet Res. 2022;24(7):e32487.
Fuchs KL, Lian J, Michels L, Mayer S, Toniato E, Tiefenbeck V. Effects of Digital Food Labels on Healthy Food Choices in Online Grocery Shopping. Nutrients. 2022;14(10).
Bjerregaard B, Larsen OB. The Danish Pathology Register. Scand J Public Health. 2011;39(7 Suppl):72-4.
Kildemoes HW, Sorensen HT, Hallas J. The Danish National Prescription Registry. Scand J Public Health. 2011;39(7 Suppl):38-41.
Schmidt M, Schmidt SA, Sandegaard JL, Ehrenstein V, Pedersen L, Sorensen HT. The Danish National Patient Registry: a review of content, data quality, and research potential. Clin Epidemiol. 2015;7:449-90.
Voldstedlund M, Haarh M, Mølbak K. The Danish microbiology database (MiBa) 2010 to 2013. Eurosurveillance. 2014;19(1):20667.
Merian S, Stöeckli S, Fuchs KL, Natter M. Buy Three to Waste One? How Real-World Purchase Data Predict Groups of Food Wasters. Sustainability. 2022;14(16):10183.
https://www.dst.dk/da/Statistik/nyheder-analyser-publ/nyt/NytHtml?cid=32131
Dolan EH, Shiells K, Goulding J, Skatova A. Public attitudes towards sharing loyalty card data for academic health research: a qualitative study. BMC Med Ethics. 2022;23(1):58.
Skatova A, Shiells K, Boyd A. Attitudes towards transactional data donation and linkage in a longitudinal population study: evidence from the Avon Longitudinal Study of Parents and Children [version 2; peer review: 2 approved]. Wellcome Open Research. 2021;4(192).
Gilmore AB, Fabbri A, Baum F, Bertscher A, Bondy K, Chang H-J, et al. Defining and conceptualising the commercial determinants of health. The Lancet. 2023;401(10383):1194-213.
https://www.efsa.europa.eu/en accessed december 5 2022
https://echa.europa.eu/da/home accessed december 5 2022

No competing interests reported.

SupplementaryAssessingtheexternalhouseholdexposomefromconsumerpurchasesTheMyPurchasescohort.docx

Download PDF

Journal Publication

published 07 Dec, 2023

Read the published version in Scientific Reports →

Editorial decision: Major revision
22 Sep, 2023
Reviews received at journal
04 Sep, 2023
Reviewers agreed at journal
26 Aug, 2023
Reviewers invited by journal
22 Aug, 2023
Editor assigned by journal
22 Aug, 2023
Editor invited by journal
18 Jun, 2023
Submission checks completed at journal
18 Jun, 2023
First submitted to journal
29 May, 2023

You are reading this latest preprint version

Assessing the external household exposome from consumer purchases: The My Purchases cohort

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Study design and population

Data sources

Recruitment

Raw data collected from receipts

Matching specific product databases to purchased products

Matching a generic product database to purchased products

Statistics

Ethics

Data availability

Results

Characteristics of participants

Receipt coverage over time

Product enrichment

Differences in nutrient values between generic and specific databases

Differences in proportion of purchases by number of included retailers

Manual assessment of product matching to generic product database

Self-reported diseases and conditions

Discussion

Conclusion and directions for future research

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1