Health & Place 85 (2024) 103146 Available online 5 December 2023 1353-8292/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Assessing the healthiness of menus of all out-of-home food outlets and its socioeconomic patterns in Great Britain Yuru Huang *, Thomas Burgoine, Tom R.P. Bishop, Jean Adams MRC Epidemiology Unit, University of Cambridge School of Clinical Medicine, Box 285 Institute of Metabolic Science, Cambridge Biomedical Campus, Cambridge, CB2 0QQ, UK A R T I C L E I N F O Keywords: Menu healthiness Out-of-home food environment Deep learning Consumer nutrition environment Area deprivation A B S T R A C T Food environment research predominantly focuses on the spatial distribution of out-of-home food outlets. However, the healthiness of food choices available within these outlets has been understudied, largely due to resource constraints. In this study, we propose an innovative, low-resource approach to characterise the healthiness of out-of-home food outlets at scale. Menu healthiness scores were calculated for food outlets on JustEat, and a deep learning model was trained to predict these scores for all physical out-of-home outlets in Great Britain, based on outlet names. Our findings highlight the “double burden” of the unhealthy food envi- ronment in deprived areas where there tend to be more out-of-home food outlets, and these outlets tend to be less healthy. This methodological advancement provides a nuanced understanding of out-of-home food environ- ments, with potential for automation and broad geographic application. 1. Introduction Foods prepared out-of-home form an increasingly important part of the global diet (Wellard-Cole et al., 2021; Lachat et al., 2012). However, these foods are typically high in energy, saturated fat, and salt, and low in micronutrients (Muc et al., 2019; Roberts et al., 2018; Jaworowska et al., 2014; Davies et al., 2016). Studies have consistently shown that more frequent out-of-home food consumption, especially fast food, is associated with poorer diet quality and higher body weight (Well- ard-Cole et al., 2021; Nago et al., 2014). Exposure to out-of-home food outlets in neighbourhoods may have an important influence on food consumption. However, studies inves- tigating the relationship between neighbourhood exposure to out-of- home food outlets, diet quality, and body weight have reported mixed results (Burgoine et al., 2014, 2018; Harbers et al., 2021; Cobb et al., 2015; Wilkins et al., 2019). Many studies to date have focused on the location and type of food outlets within neighbourhoods; the “commu- nity nutrition environment” (Lytle and Sokol, 2017; Engler-Stringer et al., 2014). However, other factors, such as the types of food available within outlets, the price of this food, and any promotions offered, may also be key elements of the out-of-home food environment (Glanz et al., 2005). Conceptualised together as the “consumer nutrition environ- ment”, these aspects of food environments have seldom been studied (Glanz et al., 2005). Those studies that do exist are mostly limited in their geographical scale, which hinders the generalisability of study findings (Holsten, 2009; Lytle and Sokol, 2017; Glanz et al., 2005). Failure to account for the consumer nutrition environment in the context of neighbourhood exposure to out-of-home food outlets may partly explain the mixed findings (Glanz et al., 2005). Studies have found that access to out-of-home food outlets—part of the “community nutrition environment”—is socioeconomically patterned in some countries, including the United Kingdom (UK) (Maguire et al., 2015; Keeble et al., 2021a). For example, there are more takeaways (“fast-food” outlets) in more deprived neighbourhoods, and the same was seen for online access to takeaway outlets in England (Maguire et al., 2015; Keeble et al., 2021a). As noted, exposure to takeaways may have implications for takeaway consumption and body mass index (BMI) (Burgoine et al., 2014). Socioeconomic patterning in access may, therefore, contribute to known socioeconomic inequalities in diet quality and health (Keeble et al., 2021b; Burgoine et al., 2014). While access to out-of-home food outlets is socioeconomically patterned in countries such as the UK, little is known about neighbourhood-level inequalities in the healthiness of out-of-home food outlets. Healthiness of menu items within an out-of-home food outlet constitutes part of the “consumer nutrition environment”. Individuals living in deprived areas have greater access to out-of-home food outlets * Corresponding author. E-mail address: Yuru.Huang@mrc-epid.cam.ac.uk (Y. Huang). Contents lists available at ScienceDirect Health and Place journal homepage: www.elsevier.com/locate/healthplace https://doi.org/10.1016/j.healthplace.2023.103146 Received 27 February 2023; Received in revised form 14 November 2023; Accepted 15 November 2023 mailto:Yuru.Huang@mrc-epid.cam.ac.uk www.sciencedirect.com/science/journal/13538292 https://www.elsevier.com/locate/healthplace https://doi.org/10.1016/j.healthplace.2023.103146 https://doi.org/10.1016/j.healthplace.2023.103146 https://doi.org/10.1016/j.healthplace.2023.103146 http://creativecommons.org/licenses/by/4.0/ Health and Place 85 (2024) 103146 2 (Maguire et al., 2015), and if these food outlets are predominantly un- healthy, it is a double burden of risk that may further contribute to health inequalities. Studies have also shown that the nutritional profile of out-of-home foods varies greatly both within and between cuisine types (Jaworowska et al., 2014; Roberts et al., 2018). As such, classi- fying out-of-home food outlets based on cuisine types may not fully capture the nuances in the healthiness of their offerings. Existing methods of assessing out-of-home food outlet healthiness are time and resource consuming, which might explain the relative lack of consumer nutrition environment research (compared to community nutrition environment work) at scale. Traditional survey methods, such as the Nutrition Environment Measures Study in restaurants (NEMS-R), require in-person visits to outlets and extensive manual categorisation of menu items (Saelens et al., 2007). While previous studies have used such methods, their application has been predominantly limited to smaller geographical areas (Neckerman et al., 2014; DuBreck et al., 2019; Wang et al., 2016). Evaluating healthiness using nutritional information is also challenging as such information is rarely available for food served by small independent retailers and laboratory analysis of menu items at scale is prohibitively expensive (Department Of Health & Social Care, 2021). In a recent study, Goffe et al. proposed a novel universal health rating system for online takeaway food outlets on JustEat – the market-leading food delivery platform in the UK (Goffe et al., 2020). They identified menu attributes, such as the availability of water and the diversity of vegetables on offer, which could be used to calculate a health rating for out-of-home food outlets in Newcastle (UK), and were predictive of experts’ healthiness ratings. In this study, we modified the universal health rating system devel- oped by Goffe et al. and developed a new model to predict menu healthiness of all out-of-home food outlets in Great Britain (GB). We then investigated whether the healthiness of all out-of-home food outlets in England, along with area-level access to these outlets, is socioeco- nomically patterned. 2. Methodology 2.1. Overview We calculated menu healthiness scores for food outlets on JustEat in GB based on menu attributes. We then predicted menu healthiness scores for all physical out-of-home food outlets in GB based on food outlet names. Fig. 1 shows our study steps and how we used different data sources and methods to calculate and predict menu healthiness scores of all out-of-home food outlets across GB. Using the predicted scores, we examined the cross-sectional relationship between area deprivation and out-of-home food outlet menu healthiness in England, adjusting for urban/rural status. 2.2. Data source 2.2.1. Menus of food outlets on JustEat The location and menus of online out-of-home food outlets in GB were scraped from the JustEat online delivery platform using a web crawler developed in Python Scrapy framework (Step 1, Fig. 1). Data were collected over two days on the 18th and 19th August 2021. 2.2.2. Ordnance Survey Points-of-interest data We obtained food outlet data from Ordnance Survey (OS) Points-of- Interest (POI), which is a comprehensive and robust dataset of all physical out-of-home food outlets across GB (not including Northern Ireland) in March 2021 (Burgoine and Harrison, 2013). OS POI data contain around four million geographic features in the UK, including commercial services, retail, accommodations, eating, and drinking. To capture all out-of-home outlets, we included all POIs under eating and drinking, except for banqueting and function rooms and Internet cafes (Step 6, Fig. 1). We grouped the subcategories fast food and takeaway outlets, fish and chip shops, and fast food delivery services as “fast food and take- aways” in our analysis. This resulted in four types of outlets in the POI dataset: (1) cafes, snack bars, and tea rooms, (2) fast food and take- aways, (3) pubs, bars, and inns, and (4) restaurants. 2.2.3. Area deprivation and urban/rural status To examine the relationship between area deprivation and out-of- home food outlet menu healthiness, we used English Indices of Multi- ple Deprivation (IMD) data from 2019 (Ministry of Housing, 2019). IMD is provided for Lower Super Output Areas (LSOAs), which are small areas with 1000 to 3000 residents (Office for National Statistics). IMD is the UK Government’s preferred measure of relative deprivation for neighbourhoods in England, combining indicators of income, employ- ment, education, skills and training, health and disability, crime, bar- riers to housing services, and living environment. Similar measures in Fig. 1. Menu healthiness scores calculation and prediction diagram. Y. Huang et al. Health and Place 85 (2024) 103146 3 Scotland and Wales are different enough to prevent straightforward combination in a single study, thus analyses including deprivation were restricted to England (Step 8, Fig. 1). We used the National Statistics Postcode Lookup (NSPL) for the UK to link food outlets with their corresponding LSOAs and attributes (e.g., IMD) (Office For National Statistics, 2021). This table was updated in August 2021, the same month as the JustEat data collection. We were able to successfully map the majority of out-of-home food outlets (99.92%) using this lookup. We also obtained urban/rural status from the NSPL, which was based on 2011 rural-urban classification of output areas (OAs). OAs are smaller units than LSOAs (Office for National Statistics). We binary coded this variable as urban (i.e., the output area has a resident popu- lation above 10,000 people) or rural (i.e., the output area has a resident population less or equal to 10,000 people) (Department for Environ- ment, 2016). 2.3. Menu healthiness models 2.3.1. Universal health rating system to calculate healthiness of online menus We modified the universal health rating system (Step 2, Fig. 1) developed by Goffe et al. (2020). Briefly, Goffe et al. proposed a health rating system for online food outlets through menu analysis. Invited experts rated outlets from 0 to 12 (12 being the healthiest) through manual review of menus. The research team then identified 15 menu metrics with known associations with health outcomes (e.g., diversity of vegetables, number of chip mentions) to predict expert scores. Among these metrics, we were unable to automatically quantify the number of multiple size meals due to our data limitations, thus it was excluded from our model. Additionally, the number of menu items available differs greatly in different out-of-home food outlets. To mini- mise the impact of a large number of items in some food outlets, we assumed the effect of some metrics plateaus at certain breakpoints. Breakpoints were derived from the distribution of expert ratings data: 10 for the number of special offers, 15 for chip mentions, 20 for dietary requirements, and 20 for salads. In cases where metrics were above these values, we recoded them as the values. We then used stepwise model selection by Akaike Information Criterion (AIC) to select the best model (Akaike, 1974). In our final universal health rating model, the healthi- ness score of a menu was calculated using the following menu attributes: number of special offers, desserts, salads, chips, milk, water, and the diversity of vegetables. We programmatically extracted data for each metric from data scraped from JustEat and applied the modified model uniformly to all out-of-home food outlets on JustEat (Step 3, Fig. 1). Details of model coefficients and how data were extracted can be found in Supplementary Appendix File S1. If a calculated menu healthiness score was greater than 12, we set the score to 12. Similarly, we set the score to 0 for a calculated menu healthiness score less than 0 to be consistent with the expert rating range (0–12). 2.3.2. Model experiments to predict menu healthiness of all out-of-home food outlets As menus for food outlets without a presence on JustEat were not available, we could not apply the universal health rating system, which relies on menu content. Instead, we built a model to predict menu healthiness of all physical out-of-home food outlets in GB (Step 4, Fig. 1). The overall approach of developing and applying this model is described in Fig. 2. Briefly, we used calculated menu healthiness scores of food outlets on JustEat, derived from the universal health rating system, as target values for our prediction task. We then defined available features and prepared data for training, which included data cleaning and resampling. We broadly followed the steps of developing deep learning models for natural language processing tasks as detailed by Bourke (2022). This included establishing a baseline model, creating various deep learning text models, and comparing their performance, as detailed below. Features: In order to apply this model uniformly to POI data, we could only utilise features available for all out-of-home food outlets in GB. The only readily available features for both JustEat and POI data were food outlet names and hygiene ratings. Hygiene ratings are determined by local authorities after in-person inspection (Food Stan- dards Agency, 2022a). In our experiments, we tested both features. For experiments involving hygiene ratings, however, we only used English data, as each UK nation has different hygiene rating schemes. We ob- tained hygiene ratings of food outlets in England through a public-facing application programming interface (“API”) provided by the Food Stan- dards Agency (FSA) (Food Standards Agency, 2022b). The hygiene rat- ings were normalised before feeding into the models. Data cleaning: We followed the data cleaning steps described by Bishop et al. to clean the names of food outlets (Bishop et al., 2021). We converted non-ASCII characters (e.g., é) to ASCII equivalents, and all characters to lowercase. Special characters (e.g., ~, %, ’) were replaced by a space. However, for our task of predicting menu healthiness, we neither removed out-of-home food outlets with the same name, nor removed the location of chain outlets. We assumed locations carried meaning (e.g., The Burger King in Cambridge may serve a healthier range than the Burger King in Liverpool; and this is reflected in our results), and that food outlets with the same name but in different lo- cations were different data points that could be fed into the models to improve prediction accuracy. Data splitting and balancing: In our data, food outlets with high or low healthiness ratings were underrepresented. We used the SMOGN method, an approach for imbalanced data for regression tasks, to resample these food outlets (Branco et al., 2017). The SMOGN algorithm combines random under-sampling with two over-sampling techniques. We first split our data into training (90%), test (5%), and validation (5%) sets, and then applied the SMOGN method to the training set only. The test set was used to assess the performance of the model, while the validation set was used to tune model parameters during model experiments. Model experiments and rationale: Our available input features were out-of-home food outlet names and their hygiene ratings. Input features were used to predict the output, a healthiness score for each outlet’s menu. Given the complexity of text data, it was not straightforward to predict the best modelling approach. As a result, we conducted a variety of experiments using models suitable for natural language processing tasks and selected the best-performing model. Baseline model: We used a traditional machine learning model to benchmark performance, serving as a baseline for comparison. A base- line is a simple model that provides reasonable results for a task and typically does not require extensive resources to build (Li et al., 2020). Traditional machine learning models, being generally simpler than deep learning models, are suitable as a baseline. Guided by the scikit-learn Fig. 2. Key steps in developing the deep learning model for predicting menu healthiness. Y. Huang et al. Health and Place 85 (2024) 103146 4 model selector flowchart, we chose support vector regression (SVR) as a baseline machine learning model (Drucker et al., 1996, scikit-learn Machine Learning in Python). SVR is a supervised machine learning algorithm used to predict continuous outcomes such as menu healthi- ness scores. To use this model, we generated the term frequency-inverse document frequency of food outlet names, which was subsequently fed into the SVR model with default parameters (Ramos, 2003). Deep learning models: Deep learning is a subset of machine learning that uses deep neural networks. Deep learning models generally outperform traditional machine learning models (Janiesch et al., 2021). We developed deep learning models with the aim of improving the performance of the baseline model. As shown in Fig. 2, we built deep learning models from scratch and also experimented with transfer learning approaches. Transfer learning involves leveraging a pre-trained model, or parts of it, to solve new problems. Often, transfer learning requires less training data to achieve better performance compared to models trained from scratch (Niu et al., 2020). In total, we created 10 distinct deep learning models, each comprising several layers with specific functions. Detailed model specification, architecture, and per- formance can be found in Supplementary Appendix File S2. For deep learning models built from scratch (n = 8), the following key components are included: 1. A tokenisation layer that splits text into smaller units. Tokenisation is often the first step for natural language processing tasks (Webster and Kit, 1992). Since the names of out-of-home food outlets are in text format, they need to be broken down into smaller units for processing. Bishop et al. suggest that as business names are short with only a few words, a character-level tokenisation may work better than word-level tokenisation (Bishop et al., 2021). In this study, we experimented with both character- and word-level tokenisation. 2. An embedding layer for creating a numerical representation of the text. This layer transforms the tokens into numeric values and cap- tures the semantic properties of these tokens (TensorFlow, 2023). 3. A multi-layer artificial neural network model. Suitable models for natural language processing (NLP) tasks include Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Recurrent Neural Networks (RNN) (Bourke, 2022). In this study, we evaluated the performance of these three types of neural network architectures. These models were chosen due to their ability to capture the sequential natural of text data (Cheng et al., 2016; Schmidt, 2019; Chung et al., 2014). 4. A flatten or concatenation layer to reshape the output. Outputs from LSTM, GRU, and RNN are often multi-dimensional. The flatten layer reshapes these outputs and makes them compatible with sub- sequent layers. In experiments involving hygiene ratings, we also concatenated the numeric input with the output from the multi-layer artificial neural network model designed for the NLP task. 5. A dropout layer to prevent overfitting. This layer randomly sets 20% of the input units to zero during training, as a rule of thumb, to mitigate the risk of overfitting. 6. An output layer for numeric output. For transfer learning experiments (n = 2), we first tested a pre- trained embedding layer, universal sentence encoding, which has been trained on diverse data sources and is applicable to a variety of natural language understanding tasks (Cer et al., 2018). Our second transfer learning approach used the Universal Language Model Fine-tuning (ULMFiT), which offers a novel way to fine-tune pre-trained language models (Howard and Ruder, 2018). ULMFiT has shown state-of-art performance in a number of natural language tasks (Bishop et al., 2021; Howard and Ruder, 2018). Briefly, we first fine-tuned a pre-trained language model, and then fine-tuned it for our regression task. Details of the ULMFiT implementation can be found in the fast.ai (version 2) documentation (Fastai, 2022). We developed these models using the Python libraries TensorFlow 2.8.0 and fast.ai 2.7.2(fastai). The best model was selected based on its performance judged by mean absolute error (MAE). MAE is the mean absolute difference between the predicted and “true” values. The lower the MAE, the closer the predicted and “true” values are. Final Model: We selected the best performing model from our model experiments based on the MAE of our test dataset (Step 4, Fig. 1). Models including only the out-of-home food outlet names outperformed models including hygiene ratings. The final model, which only used the out-of- home food outlet name, showed a MAE of 0.82. This is an improvement over our baseline model’s MAE of 0.87, indicating better performance than the default baseline model. We applied this model to OS POI data to predict healthiness scores of all out-of-home food outlets in GB (Step 7, Fig. 1). The best performing model used the ULMFiT approach proposed by Howard et al. (Howard and Ruder, 2018). 2.4. Validation study Although the performance of the deep learning model was validated against the test data set of JustEat data, there may be systematic dif- ferences in out-of-home food outlets that are and are not on JustEat that could affect prediction accuracy. We thus conducted a small validation study in Peterborough and Cambridge to verify if our algorithm could reliably predict the menu healthiness of food outlets not on JustEat (Step 5, Fig. 1). Details of our validation study can be found in Supplementary Appendix File S3. 2.5. Statistical methods We applied the best performing model uniformly to the OS POI data and obtained predicted menu healthiness for all out-of-home food out- lets in GB. We then summarised and visualised the average predicted menu healthiness score at local authority level. Whilst menu healthiness scores were predicted for all physical out- of-home food outlets in GB using the deep learning model, IMD scores were only available for England. We divided English LSOAs into deciles based on the IMD ranking, with 1 being the least deprived. We used linear mixed models with random intercepts to examine the relationship between area deprivation and menu healthiness, accounting for clus- tering at the LSOA level. Random slopes were not added as the variations of slopes within each LSOA was not our primary focus. Additionally, we investigated the association between area deprivation and out-of-home food outlet access at the LSOA level using generalised linear models. Area deprivation was coded as IMD decile, menu healthiness was modelled as a continuous variable on the outlet level, and out-of-home food outlet access was measured as the number of food outlets at the LSOA level. We performed the analysis on all types of food outlets, and also stratified analysis by food outlet type (i.e., restaurants, cafes/snack bars/tea rooms, pubs/bars/inns, and fast food and takeaways), as pre- vious analyses indicate that eating from different types of out-of-home outlet has different potential health impacts (Roberts et al., 2018; Pen- ney et al., 2017). In all models, we adjusted for urban/rural status. All statistical an- alyses were conducted in R version 4.0.2. 2.6. Ethical and legal considerations This study, focusing on the healthiness of out-of-home outlet menus, did not require ethical approval as it did not involve human subjects or tissues. When using web scraping to obtain menus of online takeaways, we only obtained publicly accessible information for non-commercial research purposes. We followed the web scraping policy published by the Office for National Statistics (ONS) to minimise any impact on the website’s server (Office For National Statistics, 2020). This included deploying strategies such as delayed page access and data scraping Y. Huang et al. Health and Place 85 (2024) 103146 5 during low-traffic periods. Our methods complied with UK copyright law, which allows lawful access material to be copied for non-commercial research (Intellectual Property Office, 2021). 2.7. Data/code availability The code for collecting JustEat menu data, conducting machine learning and deep learning experiments, applying menu healthiness ratings, and performing statistical analysis is available on GitHub (https://github.com/YuruHuang/Menu-Healthiness-Algorithm/). More details can be found in the README file in the repository. We are making all the code publicly available to facilitate further research development regarding menu healthiness. The repository contains a sample of JustEat menu data, but we are unable to share the full JustEat dataset due to copyright restrictions (full code to obtain the data has been provided). We are also unable to share the Ordnance Survey Points of Interest data as it was obtained from a third party. 3. Results In August 2021, we identified 54,575 food outlets available in GB on JustEat and calculated their menu healthiness. Using these data, we developed a deep learning model to predict menu healthiness of all out- of-home food outlets, including those not on JustEat, in GB in 2021 (N = 177,926). The deep learning model achieved an MAE score of 0.82 on the test set, and 1.12 in the validation study. 3.1. Menu healthiness of out-of-home food outlets in GB 3.1.1. Predicted menu healthiness scores, by type of food outlet The average predicted menu healthiness score was 6.72 (95% Con- fidence Interval (CI) = 6.72, 6.73). Overall, 90.9% of food outlets (N = 161,761) had a menu healthiness score between 4 and 8 on the 0–12 scale. Predicted menu healthiness scores differed by the type of out-of- home food outlet (p < 0.001). Across different types of out-of-home food outlets (Fig. 3), the average predicted menu healthiness score was highest for restaurants (7.36, 95%CI = 7.35, 7.37), followed by cafes, snack bars, and tea rooms (6.92, 95%CI = 6.92, 6.93), pubs, bars, and inns (6.88, 95%CI = 6.87, 6.88), and fast food and takeaways (6.20, 95%CI = 6.19, 6.20). As shown in Fig. 3, healthiness scores of restau- rants as well as fast food and takeaways also had larger variations compared to other types of food outlets. 3.1.2. Predicted menu healthiness, by local authority The average menu healthiness of all out-of-home food outlets (Fig. 4) was summarised and divided into quintiles at the local authority level. Local authority districts with the highest menu healthiness scores include City of London, Kensington and Chelsea, and Westminster. The local authority districts with the lowest menu healthiness scores were Northeast Lincolnshire, Luton, and Kingston upon Hull. In Supplementary Appendix File S4, we also visualised the average healthiness of food outlets on JustEat at the local authority level. 3.2. Deprivation and out-of-home food outlets’ menu healthiness 3.2.1. Deprivation and the overall menu healthiness The results were suggestive of an inverse dose-response association between area deprivation and menu healthiness in all out-of-home food outlets (Fig. 5), adjusting for urban/rural status and clustering at the LSOA level. In more deprived areas, out-of-home food outlets tend to have lower scores, indicating they are less healthy. For example, on average, the menu healthiness score was 6.50 (95% CI = 6.48, 6.53) for a food outlet in a LSOA that was in the most deprived decile, and 6.91 (95% CI = 6.88, 6.93) for a food outlet in a LSOA that was in the least deprived decile (p < 0.001 for trend). Out-of-home food outlets also tended to cluster in more deprived areas. The most deprived decline of LSOAs on average had 8.39 food outlets per LSOA (95% CI = 7.82, 8.95), while the least deprived LSOAs had 3.85 food outlets (95% CI = 3.29, 4.42). 3.2.2. Deprivation and menu healthiness, by type of food outlet We also found a non-linear inverse association between menu healthiness and deprivation in all four types of food outlet (Fig. 6, p < 0.001). All four types of food outlets tended to be less healthy in more deprived areas. The difference between the least and most deprived deciles of neighbourhoods, in terms of the average menu healthiness, was highest for fast food and takeaways (Q10 vs Q1: 0.25 95% CI = 0.20, 0.29), followed by restaurants (Q10 vs Q1: 0.21, 95% CI = 0.15, 0.27), cafes, snack bars, and tea rooms (Q10 vs Q1: 0.19 95% CI = 0.14, 0.24), and pubs, bars and inns (Q10 vs Q1:0.07, 95% CI = 0.04, 0.10). Across all four types of food outlets, we found that more were present in more deprived LSOAs (deciles, p < 0.001). The biggest difference between the least and most deprived LSOAs was for numbers of fast food and take- aways (Q10 vs Q1: 2.58, 95% CI = 2.24, 2.91). 4. Discussion 4.1. Summary of findings Our study demonstrates the feasibility of automating the character- isation of menu healthiness of food outlets at scale. We predicted the menu healthiness of over 170,000 out-of-home food outlets in GB in 2021. Two models were developed to achieve this: (1) using the modi- fied Universal Health Rating model Goffe et al. proposed (Goffe et al., 2020), we calculated healthiness based on menus; (2) we developed a deep learning model to predict healthiness score based on outlet name alone. These models have high face validity and the derived scores can be used to explore and identify area level variation in out-of-home food outlet menu healthiness. Using predicted menu healthiness scores, we found that in more deprived neighbourhoods, there tended to be both more outlets and these tend to be less healthy. This was seen both overall and for all four types of outlets studied. The difference in the average menu healthiness between the most and least deprived areas was most pronounced for takeaway and fast-food outlets. 4.2. Interpretation of findings 4.2.1. Menu healthiness scores Characterising the healthiness of all food outlets in GB could help identify target areas for local authority action to improve the consumer food environment. Similar work to predict food hygiene rating and compliance was carried out by the FSA and was helpful for local au- thorities to prioritise which businesses to inspect (Central Digital And Data Office, 2022). These automated approaches have the potential to Fig. 3. Distribution of predicted menu healthiness score, by type of food outlets. Y. Huang et al. https://github.com/YuruHuang/Menu-Healthiness-Algorithm/ Health and Place 85 (2024) 103146 6 maximise efficiency in local government. Our characterisation of out-of-home food outlet healthiness relied on menus and outlet names. However, alternate methods to assess the community food environment exist. For instance, the AUDITNOVA tool assesses the healthiness of the consumer food environment based on the NOVA food classification system for ultra-processed food (Borges et al., 2021). However, such tools require manual categorisation and might not be feasible for large-scale applications. Elsewhere, a novel method was used to quantify the healthiness of restaurants in Ohio based on online food review images (Chen et al., 2022). The advantages of our approach, compared to using online im- ages, are that it requires minimal data storage and human and computing resources, and is easy to apply to a large number of outlets based only on food outlet names. Future research should clarify the validity of different food outlet healthiness metrics. 4.2.2. Socioeconomic inequalities in neighbourhood food outlets’ menu healthiness Previous studies have shown that out-of-home food outlet access (an aspect of the “community food environment”) is socioeconomically patterned (Cummins et al., 2005; Maguire et al., 2015; Keeble et al., 2021a). Our study found that the menu healthiness of out-of-home food outlets (an aspect of the “consumer food environment”) is also socio- economically patterned in England. This may lead to a “double burden” of risk for populations living in deprived neighbourhoods where there are more outlets and these tend to be less healthy, relative to less deprived neighbourhoods. People of lower socioeconomic status are also more susceptible to unhealthy food environments (Burgoine et al., 2016, Fig. 4. Average Menu Healthiness of Out-of-Home Food Outlets at the LA level in GB. Fig. 5. Menu healthiness and out-of-home food outlet access, across areas with varying levels of deprivation. Y. Huang et al. Health and Place 85 (2024) 103146 7 2018). For example, a study found participants with the lowest income were more likely to be obese when living in areas with a high proportion of fast-food outlets (Burgoine et al., 2018). This could further result in a “triple burden”. To the best of our knowledge, our study is the first to examine so- cioeconomic inequalities in menu healthiness of out-of-home food out- lets on a national scale. Studies on socioeconomic inequalities in healthy food availability in the UK have focused on supermarkets and conve- nience stores, as opposed to out-of-home food outlets (Black et al., 2014; Williamson et al., 2017). These studies found few meaningful associa- tions between store healthiness and neighbourhood deprivation, sug- gesting socioeconomic inequalities in healthiness might be specific to out-of-home outlets (Black et al., 2014). We also found differences be- tween the most and the least deprived areas, in terms of food outlet access and healthiness, were most pronounced for fast food and take- aways relative to other types of food outlet. Given that the use of fast-food outlets, but not other types of food outlet, was associated with higher odds of obesity in one study, this may be a particular concern for public health (Penney et al., 2017). Previous studies have used other assessment tools, such as NEMS-R, to investigate socioeconomic patterning of outlet healthiness on a smaller scale, but their findings are likely to be context specific (Wang et al., 2016; DuBreck et al., 2019). In Saskatoon, Canada, consistent with our findings, researchers found restaurants in less deprived neighbour- hoods were rated healthier (Wang et al., 2016). Elsewhere, DuBreck et al. found that children in more disadvantaged areas had access to healthier restaurants in Rochester, USA, but less healthy outlets in London, Canada (DuBreck et al., 2019). Consistent with previous studies in the UK, we found that the number of out-of-home food outlets is greatest in the most deprived areas in England (Keeble et al., 2021a). 4.3. Limitations and future directions The universal health rating system was developed using a small sample size of takeaways (n = 149) from JustEat, which may not be representative of all out-of-home food outlets. Validating this model presents a challenge due to the lack of a universally accepted standard for menu healthiness and the lack of consensus regarding the definition of ’healthiness’. Compared to the model Goffe et al. proposed, we were unable to obtain the number of multi-size items, due to the complexity of harvesting and extracting this information. Furthermore, our menu healthiness score does not capture the intricate nuances of the menu, such as portion size, cooking methods, and levels of food processing. This could be important, as interventions such as healthy catering awards introduced by local government focus on aspects like smaller portion sizes, reducing salt, and switching cooking oils (Lancashire Government, 2022). As these menu item details were not accessible online, they could not easily be incorporated into the calculation. In our deep learning model, we predicted outlet menu healthiness based on outlet name alone. Although similar work has predicted cuisine type from outlet name, it is likely that menu healthiness is more complex and may not be accurately captured by name only (Bishop et al., 2021). Furthermore, our POI data was gathered in March and JustEat data in August 2021, which may introduce biases that were not accounted for in our study. However, our model performed well with an MAE of 1.12 in our validation study. Due to the nature of the metric we selected (“MAE”), predicted healthiness scores centre around 6–7 and have a relatively small variation even after resampling. However, the model is able to distinguish healthier and less healthy food outlets despite the small variation. It should also be considered that the purpose of our study is not to present perfect algorithms. Researchers can improve these models by incorporating additional information on food outlets or designing more efficient model architectures in the future. Our proposed menu healthiness score is a relative measure and ab- solute values are difficult to interpret. For instance, the difference be- tween menu healthiness scores of 6 and 7 — calculated based on the universal health rating system — could be due to more salad items, fewer chips/wedges/fries listed, and/or many other factors. The differ- ence in ‘healthiness’ between scores of 6 and 7 and 7 and 8 could also be different, yet when modelling the healthiness scores, we assumed a linear scale. Additionally, we captured what was available on menus, rather than what people purchased. Both healthier and less healthy purchases can be made from many outlets. Nonetheless, the refined categorisation of out- of-home food outlets based on menu healthiness constitutes a pivotal first step in understanding the association between menu offerings and dietary behaviour and hence whether public health efforts to improve the healthiness of menu offerings are likely to be valuable. Further research is needed to understand whether, and to what extent, the healthiness of menus may influence dietary behaviour and health outcomes. 5. Conclusion For the first time in the internationally published literature, we used online menus and outlet names to characterise the healthiness of all out- Fig. 6. Menu healthiness across areas with varying levels of deprivation, by type of food outlet. Y. Huang et al. Health and Place 85 (2024) 103146 8 of-home food outlets on a national scale. Our findings highlight the “double burden” of unhealthy consumer and community food environ- ments in deprived areas – where more, and less healthy outlets, are found. The automated method developed has the potential to enrich food environment research, as well as to aid policymaking, by better characterising out-of-home food outlets. Funding This work was supported by the Medical Research Council (grant number MC_UU_00006/7). YH is supported through a Gates Cambridge Scholarship. No funders had any role in the study design; collection, analysis and interpretation of data; the writing of the manuscript; or the decision to submit the manuscript for publication. Authors’ contributions YH, JA, and TB designed the study. YH curated data, conducted data analysis and drafted the original draft, with input from JA and TB. TRPB assisted with machine learning experiments. JA, TB, and TRPB reviewed and edited the manuscript. All authors approved the final manuscript. Rights retention statement For the purpose of Open Access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manu- script version arising. Declaration of competing interest The authors have no conflict of interest to declare. Data availability The code and algorithm used in this research are publicly available on GitHub. However, we are unable to share the full JustEat menu data (code provided) and Ordnance Survey Points of Interest data. Abbreviations API Application Programming Interface BMI Body Mass Index FSA Food Standards Agency IMD Indices of Material Deprivation LSOA Lower-layer Super Output Area MAE Mean Absolute Error NSPL National Statistics Postcode Lookup NEMS-R Nutrition Environment Measures Study in Restaurants OS Ordnance Survey POI Point of Interest SVR Support Vector Regression ULMFiT Universal Language Model Fine-tuning Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.healthplace.2023.103146. References Akaike, H., 1974. A new look at the statistical model identification. IEEE Trans. Automat. Control 19, 716–723. Bishop, T.R.P., Von Hinke, S., Hollingsworth, B., Lake, A.A., Brown, H., Burgoine, T., 2021. Automatic classification of takeaway food outlet cuisine type using machine (deep) learning. Mach. Learn Appl. 6. None. Black, C., Ntani, G., Inskip, H., Cooper, C., Cummins, S., Moon, G., Baird, J., 2014. Measuring the healthfulness of food retail stores: variations by store type and neighbourhood deprivation. Int. J. Behav. Nutr. Phys. Activ. 11, 69. Borges, C.A., Gabe, K.T., Jaime, P.C., 2021. Consumer food environment healthiness score: development, validation, and testing between different types of food retailers. Int. J. Environ. Res. Publ. Health 18. Bourke, D., 2022. Zero to Mastery TensorFlow for Deep Learning [Online]. Available: https://github.com/mrdbourke/tensorflow-deep-learning. (Accessed 23 September 2022). Branco, P., Torgo, L., Ribeiro, R.P., 2017. SMOGN: a pre-processing approach for imbalanced regression. In: PAULA BRANCO LUÍS, T., NUNO, M. (Eds.), Proceedings of the First International Workshop on Learning with Imbalanced Domains: Theory and Applications. Proceedings of Machine Learning Research: PMLR. Burgoine, T., Forouhi, N.G., Griffin, S.J., Brage, S., Wareham, N.J., Monsivais, P., 2016. Does neighborhood fast-food outlet exposure amplify inequalities in diet and obesity? A cross-sectional study. Am. J. Clin. Nutr. 103, 1540–1547. Burgoine, T., Forouhi, N.G., Griffin, S.J., Wareham, N.J., Monsivais, P., 2014. Associations between exposure to takeaway food outlets, takeaway food consumption, and body weight in Cambridgeshire, UK: population based, cross sectional study. BMJ Br. Med. J. (Clin. Res. Ed.) 348, g1464. Burgoine, T., Harrison, F., 2013. Comparing the accuracy of two secondary food environment data sources in the UK across socio-economic and urban/rural divides. Int. J. Health Geogr. 12, 2. Burgoine, T., Sarkar, C., Webster, C.J., Monsivais, P., 2018. Examining the interaction of fast-food outlet exposure and income on diet and obesity: evidence from 51,361 UK Biobank participants. Int. J. Behav. Nutr. Phys. Activ. 15, 71. Central Digital And Data Office, 2022. Food Standards Agency: Food Hygiene Rating Scheme – AI [Online]. Available: https://www.gov.uk/government/publications/food-stan dards-agency-food-hygiene-rating-scheme-ai. (Accessed 13 October 2022). Cer, D., Yang, Y., Kong, S.-Y., Hua, N., Limtiaco, N., John, R.S., Constant, N., Guajardo- Cespedes, M., Yuan, S., Tar, C., 2018. Universal Sentence Encoder. arXiv preprint arXiv:1803.11175. Chen, X., Zhao, B., Yang, X., 2022. The obesogenity of restaurant food: mapping the nutritional foodscape of Franklin County, Ohio using food review images. Appl. Geogr. 144, 102717. Cheng, J., Dong, L., Lapata, M., 2016. Long Short-Term Memory-Networks for Machine Reading. arXiv preprint arXiv:1601.06733. Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2014. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv preprint arXiv:1412.3555. Cobb, L.K., Appel, L.J., Franco, M., Jones-Smith, J.C., Nur, A., Anderson, C.A., 2015. The relationship of the local food environment with obesity: a systematic review of methods, study quality, and results. Obesity 23, 1331–1344. Cummins, S.C., Mckay, L., Macintyre, S., 2005. McDonald’s restaurants and neighborhood deprivation in Scotland and England. Am. J. Prev. Med. 29, 308–310. Davies, I.G., Blackham, T., Jaworowska, A., Taylor, C., Ashton, M., Stevenson, L., 2016. Saturated and trans-fatty acids in UK takeaway food. Int. J. Food Sci. Nutr. 67, 217–224. Department For Environment, F. R. A, 2016. In: Rural Urban Classification [Online]. Available. https://www.gov.uk/government/collections/rural-urban-classification. (Accessed 30 December 2022). Department Of Health & Social Care, 2021. Calorie Labelling in the Out of Home Sector: Implementation Guidance [Online]. Available: https://www.gov.uk/government /publications/calorie-labelling-in-the-out-of-home-sector/calorie-labelling-in-the-o ut-of-home-sector-implementation-guidance. (Accessed 6 January 2021). Drucker, H., Burges, C.J.C., Kaufman, L., Smola, A., Vapnik, V., 1996. Support vector regression machines. In: Proceedings of the 9th International Conference on Neural Information Processing Systems. MIT Press, Denver, Colorado. Dubreck, C.M., Sadler, R.C., Arku, G., Seabrook, J., Gilliland, J., 2019. A comparative analysis of the restaurant consumer food environment in Rochester (NY, USA) and London (ON, Canada): assessing children’s menus by neighbourhood socio-economic characteristics. Publ. Health Nutr. 22, 1654–1666. Engler-Stringer, R., Le, H., Gerrard, A., Muhajarine, N., 2014. The community and consumer food environment and children’s diet: a systematic review. BMC Publ. Health 14, 522. Fastai, 2022. In: Wikitext Data Tutorial, Fastai [Online]. Available. https://docs.fast.ai/. Aug 17. FOOD STANDARDS AGENCY, 2022a. Food Hygiene Rating Scheme [Online]. Available: https://www.food.gov.uk/safety-hygiene/food-hygiene-rating-scheme. (Accessed 10 October 2022). FOOD STANDARDS AGENCY, 2022b. UK food hygiene rating data API [Online]. Available: https://ratings.food.gov.uk/open-data/en-GB. (Accessed 10 October 2022). Glanz, K., Sallis, J.F., Saelens, B.E., Frank, L.D., 2005. Healthy nutrition environments: concepts and measures. Am. J. Health Promot. 19, 330–333 ii. Goffe, L., Uwamahoro, N.S., Dixon, C.J., Blain, A.P., Danielsen, J., Kirk, D., Adamson, A. J., 2020. Supporting a healthier takeaway meal choice: creating a universal health rating for online takeaway fast-food outlets. Int. J. Environ. Res. Publ. Health 17. Harbers, M.C., Beulens, J.W.J., Boer, J.M., Karssenberg, D., Mackenbach, J.D., Rutters, F., Vaartjes, I., Verschuren, W.M.M., Van Der Schouw, Y.T., 2021. Residential exposure to fast-food restaurants and its association with diet quality, overweight and obesity in The Netherlands: a cross-sectional analysis in the EPIC-NL cohort. Nutr. J. 20, 56. Holsten, J.E., 2009. Obesity and the community food environment: a systematic review. Publ. Health Nutr. 12, 397–405. Y. Huang et al. https://doi.org/10.1016/j.healthplace.2023.103146 https://doi.org/10.1016/j.healthplace.2023.103146 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref1 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref1 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref2 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref2 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref2 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref3 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref3 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref3 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref4 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref4 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref4 https://github.com/mrdbourke/tensorflow-deep-learning http://refhub.elsevier.com/S1353-8292(23)00183-1/sref6 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref6 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref6 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref6 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref7 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref7 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref7 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref8 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref8 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref8 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref8 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref9 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref9 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref9 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref10 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref10 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref10 https://www.gov.uk/government/publications/food-standards-agency-food-hygiene-rating-scheme-ai https://www.gov.uk/government/publications/food-standards-agency-food-hygiene-rating-scheme-ai http://refhub.elsevier.com/S1353-8292(23)00183-1/sref12 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref12 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref12 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref13 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref13 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref13 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref14 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref14 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref15 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref15 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref16 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref16 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref16 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref17 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref17 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref18 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref18 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref18 https://www.gov.uk/government/collections/rural-urban-classification https://www.gov.uk/government/publications/calorie-labelling-in-the-out-of-home-sector/calorie-labelling-in-the-out-of-home-sector-implementation-guidance https://www.gov.uk/government/publications/calorie-labelling-in-the-out-of-home-sector/calorie-labelling-in-the-out-of-home-sector-implementation-guidance https://www.gov.uk/government/publications/calorie-labelling-in-the-out-of-home-sector/calorie-labelling-in-the-out-of-home-sector-implementation-guidance http://refhub.elsevier.com/S1353-8292(23)00183-1/sref21 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref21 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref21 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref22 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref22 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref22 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref22 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref23 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref23 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref23 https://docs.fast.ai/ https://www.food.gov.uk/safety-hygiene/food-hygiene-rating-scheme https://ratings.food.gov.uk/open-data/en-GB http://refhub.elsevier.com/S1353-8292(23)00183-1/sref27 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref27 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref28 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref28 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref28 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref29 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref29 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref29 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref29 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref29 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref30 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref30 Health and Place 85 (2024) 103146 9 Howard, J., Ruder, S., 2018. Universal Language model fine-tuning for text classification. Available:. https://ui.adsabs.harvard.edu/abs/2018arXiv180106146H. January 01, 2022. Intellectual Property Office, 2021. Exceptions to Copyright. gov, uk. Janiesch, C., Zschech, P., Heinrich, K., 2021. Machine learning and deep learning. Electron. Mark. 31, 685–695. Jaworowska, A., Blackham, T.M., Long, R., Taylor, C., Ashton, M., Stevenson, L., Glynn Davies, I., 2014. Nutritional composition of takeaway food in the UK. Nutr. Food Sci. 44. Keeble, M., Adams, J., Bishop, T.R.P., Burgoine, T., 2021a. Socioeconomic inequalities in food outlet access through an online food delivery service in England: a cross- sectional descriptive analysis. Appl. Geogr. 133, 102498. Keeble, M., Adams, J., Vanderlee, L., Hammond, D., Burgoine, T., 2021b. Associations between online food outlet access and online food delivery service use amongst adults in the UK: a cross-sectional analysis of linked data. BMC Publ. Health 21, 1968. Lachat, C., Nago, E., Verstraeten, R., Roberfroid, D., Van Camp, J., Kolsteren, P., 2012. Eating out of home and its association with dietary intake: a systematic review of the Evidence. Obes. Rev. 13, 329–346. Lancashire Government, 2022. Recipe 4 Health Food business support pack [Online]. Available: https://www.lancashire.gov.uk/media/934221/recipe-4-health-food- business-support-pack.pdf. (Accessed 24 August 2023). Li, D., Hasanaj, E., Li, S., 2020. Baselines - Machine Learning Blog [Online]. Carnegie Mellon University. Available: https://blog.ml.cmu.edu/2020/08/31/3-baselines/. (Accessed 24 August 2023). Lytle, L.A., Sokol, R.L., 2017. Measures of the food environment: a systematic review of the field. Health Place 44, 18–34, 2007-2015. Maguire, E.R., Burgoine, T., Monsivais, P., 2015. Area Deprivation and the Food Environment over Time: A Repeated Cross-Sectional Study on Takeaway Outlet Density and Supermarket Presence in Norfolk, UK, 1990–2008, vol. 33. Health & Place, pp. 142–147. Ministry Of Housing, C.L.G., 2019. English indices of deprivation 2019 [Online]. Available: https://www.gov.uk/government/statistics/english-indices-of-deprivat ion-2019. (Accessed 28 June 2022). Muc, M., Jones, A., Roberts, C., Sheen, F., Haynes, A., Robinson, E., 2019. A bit or a lot on the side? Observational study of the energy content of starters, sides and desserts in major UK restaurant chains. BMJ Open 9, e029679. Nago, E.S., Lachat, C.K., Dossa, R.A., Kolsteren, P.W., 2014. Association of out-of-home eating with anthropometric changes: a systematic review of prospective studies. Crit. Rev. Food Sci. Nutr. 54, 1103–1116. Neckerman, K.M., Lovasi, L., Yousefzadeh, P., Sheehan, D., Milinkovic, K., Baecker, A., Bader, M.D., Weiss, C., Lovasi, G.S., Rundle, A., 2014. Comparing nutrition environments in bodegas and fast-food restaurants. J. Acad. Nutr. Diet. 114, 595–602. Niu, S., Liu, Y., Wang, J., Song, H., 2020. A decade survey of transfer learning (2010–2020). IEEE Trans. Artificial Intell. 1, 151–166. Office For National Statistics. Census geography [Online]. Available: https://www.ons. gov.uk/methodology/geography/ukgeographies/censusgeography#super-output -area-soa. (Accessed 28 June 2022). Office For National Statistics, 2020. Web Scraping Policy. ons.gov, uk. Office For National Statistics, 2021. In: National Statistics Postcode Lookup (August 2021) [Online]. Available. https://geoportal.statistics.gov.uk/datasets/aef0a4ef0 dfb49749fe4f80724477687/about. Sep 2021 2021. Penney, T.L., Jones, N.R.V., Adams, J., Maguire, E.R., Burgoine, T., Monsivais, P., 2017. Utilization of away-from-home food establishments, dietary approaches to stop hypertension dietary pattern, and obesity. Am. J. Prev. Med. 53, e155–e163. Ramos, J., 2003. Using tf-idf to determine word relevance in document queries. In: Proceedings of the First Instructional Conference on Machine Learning, pp. 29–48. Citeseer. Roberts, S.B., Das, S.K., Suen, V.M.M., Pihlajamäki, J., Kuriyan, R., Steiner-Asiedu, M., Taetzsch, A., Anderson, A.K., Silver, R.E., Barger, K., Krauss, A., Karhunen, L., Zhang, X., Hambly, C., Schwab, U., Triffoni-Melo, A.D.T., Fassini, P.G., Taylor, S.F., Economos, C., Kurpad, A.V., Speakman, J.R., 2018. Measured energy content of frequently purchased restaurant meals: multi-country cross sectional study. BMJ 363, k4864. Saelens, B.E., Glanz, K., Sallis, J.F., Frank, L.D., 2007. Nutrition environment measures study in restaurants (NEMS-R): development and evaluation. Am. J. Prev. Med. 32, 273–281. Schmidt, R.M., 2019. Recurrent Neural Networks (Rnns): A Gentle Introduction and Overview. arXiv preprint arXiv:1912.05911. Scikit-Learn Machine Learning In Python, 2023. In: Choosing the Right Estimator¶ [Online]. Available. https://scikit-learn.org/stable/tutorial/machine_learning_ma p/index.html. Jun 22. TENSORFLOW, 2023. Word embeddings [Online]. Available: https://www.tensorflow. org/text/guide/word_embeddings. (Accessed 25 August 2023). Wang, J., Engler-Stringer, R., Muhajarine, N., 2016. Assessing the consumer food environment in restaurants by neighbourhood distress level across saskatoon, saskatchewan. Can. J. Diet Pract. Res. 77, 9–16. Webster, J.J., Kit, C., 1992. Tokenization as the initial phase in NLP. COLING 1992 volume 4. The 14th international conference on computational linguistics. Wellard-Cole, L., Davies, A., Allman-Farinelli, M., 2021. Contribution of foods prepared away from home to intakes of energy and nutrients of public health concern in adults: a systematic review. Crit. Rev. Food Sci. Nutr. 1–12. Wilkins, E., Radley, D., Morris, M., Hobbs, M., Christensen, A., Marwa, W.L., Morrin, A., Griffiths, C., 2019. A systematic review employing the GeoFERN framework to examine methods, reporting quality and associations between the retail food environment and obesity. Health Place 57, 186–199. Williamson, S., Mcgregor-Shenton, M., Brumble, B., Wright, B., Pettinger, C., 2017. Deprivation and healthy food access, cost and availability: a cross-sectional study. J. Hum. Nutr. Diet. 30, 791–799. Y. Huang et al. https://ui.adsabs.harvard.edu/abs/2018arXiv180106146H http://refhub.elsevier.com/S1353-8292(23)00183-1/sref32 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref33 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref33 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref34 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref34 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref34 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref35 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref35 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref35 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref36 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref36 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref36 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref36 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref37 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref37 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref37 https://www.lancashire.gov.uk/media/934221/recipe-4-health-food-business-support-pack.pdf https://www.lancashire.gov.uk/media/934221/recipe-4-health-food-business-support-pack.pdf https://blog.ml.cmu.edu/2020/08/31/3-baselines/ http://refhub.elsevier.com/S1353-8292(23)00183-1/sref40 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref40 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref41 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref41 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref41 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref41 https://www.gov.uk/government/statistics/english-indices-of-deprivation-2019 https://www.gov.uk/government/statistics/english-indices-of-deprivation-2019 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref43 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref43 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref43 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref44 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref44 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref44 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref45 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref45 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref45 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref45 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref46 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref46 https://www.ons.gov.uk/methodology/geography/ukgeographies/censusgeography#super-output-area-soa https://www.ons.gov.uk/methodology/geography/ukgeographies/censusgeography#super-output-area-soa https://www.ons.gov.uk/methodology/geography/ukgeographies/censusgeography#super-output-area-soa http://refhub.elsevier.com/S1353-8292(23)00183-1/sref48 https://geoportal.statistics.gov.uk/datasets/aef0a4ef0dfb49749fe4f80724477687/about https://geoportal.statistics.gov.uk/datasets/aef0a4ef0dfb49749fe4f80724477687/about http://refhub.elsevier.com/S1353-8292(23)00183-1/sref50 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref50 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref50 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref51 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref51 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref51 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref52 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref53 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref53 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref53 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref54 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref54 https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html https://www.tensorflow.org/text/guide/word_embeddings https://www.tensorflow.org/text/guide/word_embeddings http://refhub.elsevier.com/S1353-8292(23)00183-1/sref57 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref57 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref57 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref58 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref58 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref59 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref59 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref59 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref60 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref60 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref60 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref60 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref61 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref61 http://refhub.elsevier.com/S1353-8292(23)00183-1/sref61 Assessing the healthiness of menus of all out-of-home food outlets and its socioeconomic patterns in Great Britain 1 Introduction 2 Methodology 2.1 Overview 2.2 Data source 2.2.1 Menus of food outlets on JustEat 2.2.2 Ordnance Survey Points-of-interest data 2.2.3 Area deprivation and urban/rural status 2.3 Menu healthiness models 2.3.1 Universal health rating system to calculate healthiness of online menus 2.3.2 Model experiments to predict menu healthiness of all out-of-home food outlets 2.4 Validation study 2.5 Statistical methods 2.6 Ethical and legal considerations 2.7 Data/code availability 3 Results 3.1 Menu healthiness of out-of-home food outlets in GB 3.1.1 Predicted menu healthiness scores, by type of food outlet 3.1.2 Predicted menu healthiness, by local authority 3.2 Deprivation and out-of-home food outlets’ menu healthiness 3.2.1 Deprivation and the overall menu healthiness 3.2.2 Deprivation and menu healthiness, by type of food outlet 4 Discussion 4.1 Summary of findings 4.2 Interpretation of findings 4.2.1 Menu healthiness scores 4.2.2 Socioeconomic inequalities in neighbourhood food outlets’ menu healthiness 4.3 Limitations and future directions 5 Conclusion Funding Authors’ contributions Rights retention statement Declaration of competing interest Data availability Abbreviations Appendix A Supplementary data References