**Higher Ultra-Processed Food Consumption Is Associated with Greater High-Sensitivity C-Reactive Protein Concentration in Adults: Cross-Sectional Results from the Melbourne Collaborative Cohort Study**

**Melissa M. Lane 1,\*, Mojtaba Lotfaliany 1, Malcolm Forbes 1,2,3, Amy Loughman 1, Tetyana Rocks 1, Adrienne O'Neil 1, Priscila Machado 4,5, Felice N. Jacka 1,6,7, Allison Hodge 8,9 and Wolfgang Marx <sup>1</sup>**


**Abstract:** Background: Few studies have examined associations between ultra-processed food intake and biomarkers of inflammation, and inconsistent results have been reported in the small number of studies that do exist. As such, further investigation is required. Methods: Cross-sectional baseline data from the Melbourne Collaborative Cohort Study (MCCS) were analysed (*n* = 2018). We applied the NOVA food classification system to data from a food frequency questionnaire (FFQ) to determine ultraprocessed food intake (g/day). The outcome was high-sensitivity C-reactive protein concentration (hsCRP; mg/L). We fitted unadjusted and adjusted linear regression analyses, with sociodemographic characteristics and lifestyle- and health-related behaviours as covariates. Supplementary analyses further adjusted for body mass index (kg/m2). Sex was assessed as a possible effect modifier. Ultra-processed food intake was modelled as 100 g increments and the magnitude of associations expressed as estimated relative change in hsCRP concentration with accompanying 95% confidence intervals (95%CIs). Results: After adjustment, every 100 g increase in ultra-processed food intake was associated with a 4.0% increase in hsCRP concentration (95%CIs: 2.1–5.9%, *p* < 0.001). Supplementary analyses showed that part of this association was independent of body mass index (estimated relative change in hsCRP: 2.5%; 95%CIs: 0.8–4.3%, *p* = 0.004). No interaction was observed between sex and ultra-processed food intake. Conclusion: Higher ultra-processed food intake was cross-sectionally associated with elevated hsCRP, which appeared to occur independent of body mass index. Future prospective and intervention studies are necessary to confirm directionality and whether the observed association is causal.

**Keywords:** ultra-processed food; NOVA; diet; inflammation; high-sensitivity C-reactive protein; non-communicable diseases; cross-sectional

**Citation:** Lane, M.M.; Lotfaliany, M.; Forbes, M.; Loughman, A.; Rocks, T.; O'Neil, A.; Machado, P.; Jacka, F.N.; Hodge, A.; Marx, W. Higher Ultra-Processed Food Consumption Is Associated with Greater High-Sensitivity C-Reactive Protein Concentration in Adults: Cross-Sectional Results from the Melbourne Collaborative Cohort Study. *Nutrients* **2022**, *14*, 3309. https://doi.org/10.3390/ nu14163309

Academic Editors: Monica Dinu and Daniela Martini

Received: 15 July 2022 Accepted: 10 August 2022 Published: 12 August 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

#### **1. Introduction**

Nutrition science has long sought to understand the effects of diet on human health. This has largely been done by classifying foods based on their nutrient composition. The impacts on health of individual macro- and micro-nutrients as well as kilojoules have typically been considered independent of different foods and food group sources [1]. Excess intakes of sugar, salt, saturated fat, and kilojoules have been previously linked with increased risk of cardiometabolic conditions [2–4]. Such research has been beneficial for understanding nutritional physiology and has subsequently informed dietary recommendations [5]. However, this nutrient-centric perspective does not capture the effect of complex food matrices. A food matrix is characterised as the molecular interactions between nutrient and non-nutrient components of food [6]. Indeed, emerging experimental [7] and epidemiological [8–10] evidence implicates the extent to which a food has been processed (or undergone food matrix alterations) as a risk factor for chronic non-communicable diseases, morbidity, and mortality.

The NOVA (name, not acronym) food classification system was recently developed to allow for the categorisation of food items based on their level of processing: from unprocessed or minimally processed food, processed culinary ingredients, processed food, to extensively processed food termed "ultra-processed" [11]. Ultra-processed foods, are defined by NOVA as industrial formulations created from compounds extracted, derived, or synthesised from food or food substrates. Ultra-processed foods also typically contain five or more ingredients including artificial food additives (e.g., colours, texturising agents, and olfactory and taste enhancers) and are commonly inexpensive, virtually imperishable, easily consumed, and highly palatable [12]. Time-series country-level sales data from 2006 to 2019 show a substantial growth in the types and quantities of ultra-processed foods sold worldwide, with projected increases to 2024 [13,14]. This suggests a transition away from non-ultra-processed food and toward a more processed global diet [13,14].

Chronic low-grade inflammation, marked by the presence of elevated inflammatory cytokines, is both a driver of chronic diseases and a characteristic of an established diseased state [15]. These diseases include cancers [16], cardiometabolic conditions [17], and mental disorders [18,19]. The shared link between chronic low-grade inflammation and diseased states exists despite the different organs and systems involved in their onset, prognosis, and morbidity [20]. Hence, better understanding and addressing possible drivers of inflammation is of significant public health interest. However, little data are available that have directly linked ultra-processed food intake to inflammation.

In the three epidemiological studies that do exist [21–23], inconsistent results have been observed. These included sex- and cohort-specific differences within and between studies [21,22] as well as associations of ultra-processed food with some inflammatory biomarkers but not others [23]. Importantly, each of these three studies included samples from Brazil, where the concept of avoiding ultra-processed food has received recognition in official dietary guidelines since 2014 [24], and where consumption of ultra-processed food is estimated to be lower than higher-income countries [11,14]. Thus, there is a need for further investigation of the association between ultra-processed food intake and inflammation in other regions, particularly in settings where the substitution of non-ultra-processed foods for those that are ultra-processed is increasingly common. The current study aimed to address this gap by using data from the Melbourne Collaborative Cohort Study (MCCS) to investigate cross-sectional associations between ultra-processed food intake and plasma concentrations of the inflammatory cytokine, high-sensitivity C-reactive protein (hsCRP).

#### **2. Methods**

A full description of methods for data collection in the MCCS was published elsewhere [25]. In brief, the MCCS is a study that aimed to assess prospective associations between diet and lifestyle and chronic non-communicable diseases [25]. Between 1990 and 1994 (baseline), 41,513 people (24,469 women) aged between 27 and 76 (99% 40–69) years were recruited from the Melbourne metropolitan area, with migrants from Southern

Europe included and deliberately recruited to expand the span of diet and lifestyle exposures. At baseline, participants completed surveys and anthropometric measurements, and blood samples were collected. A case-cohort sub-study was undertaken to provide a more economical design within which to perform assays on blood samples collected at baseline (as part of the original MCCS project) and analyse associations of various molecules in the stored plasma with selected disease outcomes; cases were included in the case-cohort if they were identified as having the outcome of interest by 30 June 2002. High-sensitivity C-reactive protein as a biomarker of inflammation was measured in incident cardiovascular death cases together with a random sample (sub-cohort) of all participants in the MCCS.

The current cross-sectional study thus comprised a sample of participants from the case-cohort sub-study for whom valid baseline dietary data and plasma hsCRP measurements were available (see Figure 1). We excluded participants who had missing hsCRP data (*n* = 161), total energy intake (kJ/day) below the 1st and above the 99th percentiles (*n* = 40), or hsCRP concentration above the 99th percentile (*n* = 23). Two thousand and eighteen participants remained for analysis, including both the cardiovascular disease death group as cases (*n* = 567) and the random sample of all participants from the original MCCS project as the sub-cohort (*n* = 1451). Cardiovascular disease death cases were identified from notifications to the Victorian Registry of Births, Deaths and Marriages, and the Australian National Death Index (codes 390–459 and I00–I99 of the International Classification of Diseases, ICD-9 and ICD-10, respectively).

**Figure 1.** Flow-chart of participant selection. MCCS—Melbourne Collaborative Cohort Study, hsCRP—high-sensitivity C-reactive protein.

The current study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement and checklist for crosssectional studies [25]. The current study was prospectively registered with Open Science Framework (OSF) registry (registration DOI: 10.17605/OSF.IO/EHFXD) and was approved for exemption from ethical review in accordance with the National Statement on Ethical Conduct in Human Research (2007, updated 2018) Section 5.1.22 by the Deakin University Human Research Ethics Committee (project number: 2020-413, received 18th of November 2020). The study protocol for the original MCCS project was approved by the Cancer Council Victoria's Human Research Ethics Committee (project number: IEC 9001, received 23rd of August 1990). Participants provided written consent to participate including researcher access to their medical records [26].

#### *2.1. Exposure: Dietary Assessment*

At baseline of the original MCCS project, participants attended clinics where dietary data were collected using a self-administered 121-item food frequency questionnaire (FFQ) specifically developed for use in this multiethnic cohort [27]. This FFQ was based on weighed food records in 810 Melbournians of similar demographics to the cohort [27]. For this study, all FFQ food items were classified according to the NOVA food classification system as ultra-processed foods and non-ultra-processed foods by two experts with Australian food and dietary intake knowledge. Examples of NOVA's classification of ultra-processed food (NOVA group 4) include soft drinks, sweet or savoury packaged snacks, confectionery, packaged breads and buns, margarine, reconstituted meat products, and pre-prepared frozen or shelf-stable dishes. Examples of NOVA's classification of non-ultra-processed food include unprocessed or minimally processed foods (NOVA group 1) such as rice and other cereals, meat, fish, milk, eggs, fruit, roots and tubers, vegetables, and nuts and seeds; processed culinary ingredients (NOVA group 2) such as sugar, plant oils, and butter; and processed foods (NOVA group 3) such as processed breads and cheese, canned fruit and fish, and salted and smoked meats. More information regarding the NOVA food classification system can be found elsewhere (1). When it was not possible to discriminate, (e.g., food items such as 'bread', 'pasta or noodles', 'low fat cheese', 'yoghurt', and 'fruit juice'), cross-sectional data from the National Nutrition Survey 1995–1996 (data not published) and Australian National Nutrition and Physical Activity Survey (NNPAS) 2011–2012 were used for comparison and decision making [28].

As per previous research [27,29,30], the mean daily contribution of ultra-processed foods to intake of total energy (kJ) and weight (g) was calculated by transforming frequencies into grams based on sex-specific portion sizes of each food multiplied by the daily equivalent frequency. Energy was estimated based on the Nutrient Data Table for Use in Australia 1995 (NUTTAB 95). The NUTTAB 95 is a food composition database that contains information for 1800 foods and beverages available in Australia [31].

#### *2.2. Outcome: Inflammatory Cytokine Assessment*

Blood samples were also collected at baseline. Venous blood samples were drawn in lithium-heparin tubes, and plasma was subsequently stored at −180 ◦C (liquid nitrogen) until assayed as part of the case-cohort sub-study. HsCRP concentration expressed as mg/L was measured by a high-sensitivity immunonephelometric assay. The assay was a Dade Behring nephelometric assay done on a BNII Nehphelometric Analyser, Dade-Behring Diagnostics, Lane Cove, NSW, Australia.

#### *2.3. Assessment of Covariates*

Potential covariates were identified based on previous literature [7,21,32–35] and included in a directed acyclic graph to map hypothesised causal relationships between all relevant variables (Supplementary Materials Figure S1). These covariates were assessed through a structured interview that was administered at baseline providing data on sociodemographic characteristics and lifestyle- and health-related factors.

In particular, the sociodemographic characteristics included: age (continuous), sex (men and women), country of birth (Australia/New Zealand, United Kingdom/Malta, Italy, and Greece), marital status (married, de facto, single, divorced, separated, and widow), highest level of education (primary school, high/technical school, and tertiary degree or diploma) and Socio-Economic Indexes for Areas (SEIFA)—Index of Relative Socio-Economic Disadvantage [36]. SEIFA scores are recorded by the Australian Bureau of Statistics and refer to the relative socioeconomic advantage and disadvantage of defined geographical areas such as postal code [37] (we divided these scores into quintiles, with the lowest and highest representing the greatest and least disadvantaged, respectively).

The lifestyle- and health-related factors included: smoking status (never smoked, current smoker, and former smoker), leisure time physical activity over the last 6 months (a score was calculated ranging from 0–16 based on the frequency of walking and less vigorous and vigorous activity multiplied by two, which was then divided into categories, namely: 0 (none), >0 and <4 (low), ≥4 and <6 (moderate), and ≥6 (high) [32,34]), and alcohol intake using beverage-specific quantity frequency questions (lifetime abstainers, ex-drinkers, and current drinkers (further categorised as up to 19, 20–29, 30–39, and 40+ g/day)) [26,32,33]. Height and weight were measured, and body mass index was calculated as kg/m2 [26]. These sociodemographic characteristics and lifestyle- and health-related variables were used as covariates.

#### *2.4. Statistical Analyses*

An inverse probability weighting method was applied to address the case-cohort design and adjust for the possibility of oversampling cases versus participants from the sub-cohort [26]. Characteristics of participants were summarised using mean and standard deviation (SD) or median and interquartile range (IQR) for continuous variables and frequency and percentage for categorical variables.

Linear regression analysis was used to examine associations between the consumption of ultra-processed food and hsCRP concentration. We aimed to better account for ultraprocessed formulations that did not provide energy or were low in energy (e.g., artificially sweetened beverages) as per [38–40]. Thus, the total weight of ultra-processed foods in grams per day (g/day) was adjusted for energy using Willet's residual method [41] and used to model our exposure in 100 g increments. We chose 100 g increments to aid in reporting and interpretation. We assessed our outcome variable, hsCRP concentration, continuously; however, because the variance of residuals for hsCRP concentration was not homogeneous along all of the fitted values, hsCRP was log (to base e) transformed; the exponentiated coefficients represent the percent change from the geometric mean (anti-log). There were no zero values for hsCRP concentration. We verified the assumptions for a linear model with graphical and statistical tests of the associations between ultra-processed food intake and hsCRP concentration as well as between the fitted values and residuals. We further added the exposure value squared in the model to assess whether there was curvature in the association between ultra-processed food intake and hsCRP concentration. We used locally weighted scatterplot smoothing (LOWESS) models to examine whether there was a threshold for the association between ultra-processed food intake and hsCRP concentration. We also tested with graphical and statistical tests the normality of residuals and homoscedasticity (homogeneity of variance). We used the variance inflation factor to assess collinearity between the potential confounders included in our models. Lastly, to allow for comparison with the three previously published ultra-processed inflammation studies [21–23], and given that the assumptions associated with a linear model were not violated, we used linear models. The estimated relative change in hsCRP concentration (mg/L) for each energy-adjusted 100 g increase in ultra-processed food consumption was thus calculated along with robust standard 95% confidence intervals (95%CIs) and *p*-values. We also estimated the variance explained (Pseudo-R2) in hsCRP by ultra-processed food consumption via the Cragg-Uhler method [42].

Four different sequential models were fitted: energy-adjusted ultra-processed food as the exposure variable and otherwise unadjusted (model 1), additionally adjusted for sociodemographic characteristics (model 2), and a fully adjusted model that further controlled for lifestyle- and health-related behaviours (model 3). Since previous studies have highlighted body mass index as a potential mediator in the association between ultra-processed food consumption and inflammation [7,21], supplementary analyses were performed by additionally adjusting for body mass index (model 4). It is important to highlight here that we were interested in assessing the "total effect" of ultra-processed food consumption on hsCRP concentration. As such, and given that body mass index was a prespecified mediator, we assessed its possible impact as part of our supplementary analyses (model 4). However, given the cross-sectional nature of our study, we refrained from referring to and did not formally test for mediation [43]. Because body mass index did not qualify as a confounder (see [44]), model 3 was considered the main model. Other studies have also reported differences between sexes in the association between ultra-processed food consumption and inflammation [21,22]. We thus undertook further supplementary analyses to (a) stratify by sex and (b) assess the potential effect modification of sex with ultra-processed food consumption. To explore sex as a possible effect modifier, we added interaction terms between sex and ultra-processed food consumption into the main effects model.

To ensure the sampling methods did not affect the results, sensitivity analyses were performed across all models with the following exclusions: (1) people with hsCRP > 10 mg/L (*n* = 122), which may indicate acute inflammation [45], although these values can also be seen in cases of chronic inflammation [46]; and (2) cases defined by cardiovascular disease mortality (*n* = 567). We conducted further sensitivity analyses on our main model 3 by excluding people with history of non-communicable diseases, such as hypertension (*n* = 555), stroke (*n* = 44), heart attack (*n* = 129), cancer (*n* = 166), diabetes mellitus (*n* = 100), and body mass index ≥30 (*n* = 520).

Lastly, we conducted post hoc analyses by fitting a logistic regression on our main model 3 (adjusted for sociodemographic characteristics and lifestyle- and health-related behaviours) to assess associations between each energy-adjusted 100 g increase in ultraprocessed food consumption and the odds of hsCRP at or above 3 mg/L, which is considered a risk factor for cardiovascular events [45].

The analyses were undertaken using R version 3.6.3 (R Development Core Team, Vienna, Austria) [47].

#### **3. Results**

The current study included 1261 men and 757 women. Table 1 shows the sociodemographic and lifestyle characteristics of participants. The mean age of participants was 57 years. Most people reported that they were married or in a de facto relationship (75.6%) as well as reporting their country of birth as Australia or New Zealand (64.2%). Approximately one quarter of participants were in the top quintile of SEFIA (least disadvantaged; 25.1%) and reported that they had either some study towards or had completed a tertiary degree or diploma (24.0%). Less than a fifth of participants reported that they were current smokers (14.2%) and over one fifth (21.7%) reported that they had engaged in high physical activity over the last six months. Most participants (40.8%) had an average alcohol intake of less than 19 (g/day). The mean body mass index for all participants was 27.8 (kg/m2), and the mean proportion of ultra-processed food in the overall diet by weight and energy was 26% (g/day) and 40% (kJ/day), respectively. In terms of ultra-processed food intake in absolute weight and energy, the median was 364.4 (g/day) and 2975.1 (kJ/day), respectively. The median hsCRP concentration for participants was 1.6 (mg/L).


**Table 1.** Descriptive characteristics of the study population.

<sup>a</sup> (In)complete tertiary degree or diploma referred to participants who had some study towards a tertiary degree or diploma as well as participants who had completed a tertiary degree or diploma. <sup>b</sup> SEIFA Socio-Economic Indexes for Areas [37]. <sup>c</sup> Ordinal score based on frequency of walking plus frequency of less vigorous activity plus twice the frequency of vigorous activity, and ranging from 0 to 16 [32,34]. MCCS—Melbourne Collaborative Cohort Study, hsCRP—high-sensitivity C-reactive protein, SD—standard deviation.

Table 2 details the results of the multivariable adjusted models. In model 1, every 100 g increase in ultra-processed food intake was associated with a 3.6% increase in hsCRP concentration (95%CIs: 1.7–5.5%, *p* < 0.001). After accounting for sociodemographic characteristics and lifestyle- and health-related behaviours in the main multivariable analysis (model 3), the association remained robust (expected relative change in hsCRP: 4.0%; 95%CIs: 2.1–5.9%, *p* < 0.001). The supplementary analyses including all participants and further adjustment for body mass index are also shown in Table 2 (model 4). Part of the association between ultra-processed food intake and hsCRP concentration was independent of body mass index, where every 100 g increase in ultra-processed food intake was associated with a 2.5% increase in hsCRP concentration (95%CIs: 0.8–4.3%, *p* = 0.004). Results remained relatively stable in our sensitivity analyses that excluded people with hsCRP concentrations above 10 mg/L, cardiovascular disease mortality and history of cardiovascular diseases, cancer, diabetes mellitus, and body mass index ≥30 (see Supplementary Materials Tables S1 and S2).


**Table 2.** Cross-sectional associations between ultra-processed food intake and hsCRP concentration (MCCS, 1990–1994).

Regressions performed with hsCRP on a logarithmic scale. <sup>a</sup> Model 1 = unadjusted. <sup>b</sup> Model 2 = additionally adjusted for sociodemographic characteristics: sex (men and women), age (continuous), education ((in)complete tertiary degree or diploma, completed high/technical school, (in)complete high/technical school, completed primary school, and (in)complete primary school), country of birth (Australia/New Zealand/Other, United Kingdom/Malta, Italy, and Greece), marital status (married, de facto, divorced, separated, and widow), and SEIFA quintiles (Q1–Q5). Change to *n* due missing values for confounders marital status and SEIFA quintiles. <sup>c</sup> \*Model 3 = main model additionally adjusted for lifestyle- and health-related behaviours: smoking status (never smoked, current smoker, and former smoker), physical activity over the last 6 months (0 (none), >0 and <4 (low), ≥4 and <6 (moderate), and ≥6 (high)), and alcohol intake (g/day) (lifetime abstainers, ex-drinkers, and up to 19, 20–29, 30–39, and 40+). Change to *n* due missing values for confounder alcohol intake. <sup>d</sup> \*\*Model 4 = supplementary analyses additionally adjusted for body mass index (kg/m2). Change to *n* due missing values for confounders alcohol intake and body mass index. SEIFA—Socio-Economic Indexes for Areas, 95%CIs—95% confidence intervals.

There was no evidence of sex interactions (all *p*-values > 0.05 and estimates of interaction range: 0.0–2.6%). The supplementary analyses stratified by sex are shown in Supplementary Materials Table S3. After accounting for potential confounders in our main model 3, every 100 g increase in ultra-processed food intake was associated with an increase in hsCRP concentration in both men (estimated relative change in hsCRP: 3.5%; 95%CIs: 1.3–5.7%, *p* = 0.002) and women (estimated relative change in hsCRP: 5.5%; 95%CIs: 0.5–10.5%, *p* = 0.032). However, after further adjustment for body mass index (model 4), the association remained robust in men only (estimated relative change in hsCRP for men: 2.8%; 95%CIs: 0.7–4.9%, *p* = 0.010 versus women: 2.4%, 95%CIs: −2.1–6.8%, *p* = 0.296). Post hoc analyses on our main model 3 showed that each 100 g increase in ultra-processed food consumption was associated with 1.08-fold increased odds of hsCRP concentration at or above 3 mg/L after adjusting for sociodemographic characteristics and lifestyle- and health-related behaviours (odds ratio: 1.080; 95%CIs: 1.034–1.128, *p* < 0.001).

#### **4. Discussion**

This study aimed to examine whether greater ultra-processed food intake was associated with higher hsCRP concentration in a sample of Australian adults. We found evidence of this association, and at least part of this association was independent of body mass index.

Three epidemiological studies have previously tested associations between ultraprocessed food consumption and biomarkers of inflammation [21–23]. Overall associations with men and women combined were not tested in two of these ultra-processed foodinflammation studies [21,22]. This makes it challenging to compare these studies' results with the main results from our study. However, our results are partly consistent with another that assessed overall associations in male and female adolescents aged from 17 to 18 years [23]. That study demonstrated direct cross-sectional associations between the consumption of ultra-processed food and concentration of the inflammatory cytokine, interleukin-8 [23]. It also showed that participants with the highest intake of ultra-processed food had increased concentrations of leptin and C-reactive protein compared to participants with the lowest intake, but these associations were less certain given the 95% confidence intervals that crossed zero in both the unadjusted and fully adjusted linear models [23]. These less certain findings, particularly regarding C-reactive protein, may be partly explained by the included sample of adolescents who were exclusively from public schools in a lower socio-economic region of Brazil [23]. The authors of that study noted that these

sociodemographic characteristics have been associated with lower consumption of ultraprocessed food, with generalisability issues and underestimated effect estimates remaining possible [23]. Indeed, ultra-processed food contributed 26% to total daily energy intake in that Brazilian sample of adolescents compared to 40% in our sample.

While sex was not a significant effect modifier in our study, we conducted sex-stratified supplementary analyses to allow for comparison with previous literature [21,22]. One previous study reported direct prospective associations between the intake of ultra-processed food and interleukin-6 concentrations across two separate cohorts, with one cohort showing an association in women only and the other showing an association in men only [22]. Adiposity did not appear to explain these cohort- and sex-specific findings [22]. Results for the men-only analysis in our study support the data from the second cohort [22], where we also found associations between higher ultra-processed food intake and elevated hsCRP concentration across all models, including additional adjustment for body mass index.

In contrast, for the women in our study, observed associations were not independent of body mass index. These findings are somewhat concordant with another previous ultraprocessed food-inflammation study [21], which reported direct cross-sectional associations between ultra-processed food intake and high-sensitivity C-reactive protein in women only and that appeared to be explained by body mass index [21]. These findings suggest that in women, adiposity is a possible intermediate on the causal pathway from ultraprocessed food consumption to inflammation. This notion may be explained by the greater accumulation of adiposity, on average, in women compared to men; associations between body mass index and C-reactive protein concentration are suggested to be stronger for women than men [48]. However, cross-sectional studies do not allow for formal tests of mediation [43], and this notion requires further investigation in prospective analyses. However, it is important to reiterate that testing for effect modification by sex in our study showed no evidence of interaction. Given the underrepresentation of women compared to men in our study (37.5%), it is possible that we may not have had adequate power to detect this interaction. Further investigation with more appropriately designed studies is needed.

Our study is consistent with recent systematic reviews and meta-analyses [8–10] showing direct associations between intake of ultra-processed food and the prevalence and incidence of common chronic non-communicable diseases, morbidity, and mortality, all of which include inflammation as part of their pathophysiology [49]. Our results are also consistent with a recent systematic review of observational studies and broader whole of diet or dietary pattern analyses [50]. Not unlike the NOVA food classification system, dietary patterns expand beyond isolated nutrients and account for the fact that foods are consumed in complex combinations [50]. This systematic review reported that indices and scores used to assess the inflammatory potential of diets (e.g., Dietary Inflammatory Index) were directly and cross-sectionally associated with inflammatory biomarkers, including C-reactive protein, interleukin-6, tumour necrosis factor-α, and fibrinogen [50]. Pro-inflammatory dietary patterns were characterised by, for example, excess consumption of kilojoule-dense Western-style foods, including red and processed meats, sweets, desserts, fried foods, and refined grains [51]. Similarly, diet scores measuring adherence to healthy or Mediterranean-style diets—rich in fruits, vegetables, fatty fish, poultry, extra virgin olive oil, and whole grains—appeared to be inversely associated with inflammatory biomarkers in cross-sectional analyses [50]. In terms of experimental evidence, our results are also consistent with an earlier meta-analysis of intervention studies showing that Mediterranean diets higher in unprocessed or minimally processed foods were anti-inflammatory [52].

A potential role of the gut microbiota in the link between ultra-processed food intake and inflammation was hypothesised [53]. Preliminary theory posits that extensive food processing leading to the degradation of cell walls within food and acellular compounds (i.e., deconstruction of the food matrix and nutrients not contained within cells, respectively [54]) may impact abnormal absorption and signalling from the gastrointestinal tract as well as its interactions with gastrointestinal microbiota [53–55]. Both may in turn promote microbe encroachment on the gastrointestinal wall and a cascade of inflammatory processes [53]. Though not extensively demonstrated in humans, several pre-clinical rodent studies have indicated an effect of advanced glycation end-products (AGEs) formed during the thermal treatment of food products [56,57] and artificial additives common to ultra-processed food (e.g., carboxymethylcellulose [58,59], polysorbate-80 [59,60], saccharin [61], and sucralose [62]) on the gut microbiota composition and activity together with host physiology including pro-inflammatory states. One recent randomised controlled-feeding study in humans reported a detrimental effect of the emulsifier carboxymethylcellulose on the gut microbiota and metabolome, with the authors of that study surmising that carboxymethylcellulose may be contributing to an array of chronic inflammatory diseases [63]. This emerging evidence is certainly suggestive and warrants further investigation to determine the precise features of ultra-processed food that elicit their unhealthful effects.

#### *4.1. Limitations and Future Research*

Our results should be interpreted with consideration of the following limitations. First, the possible temporality of these associations cannot be established from this single cross-sectional study, and residual confounding cannot be excluded. However, our results were unchanged after adjusting for common confounders and after various sensitivity analyses, which showed that the associations remained relatively stable with or without the inclusion of people with markedly elevated hsCRP concentration. Sensitivity analyses also highlighted that our sampling methods may not have biased results given the stability of effect estimates no matter whether we included cases defined by cardiovascular disease mortality (which occurred after data and sample collection) or participants with a history of non-communicable diseases, such as cardiovascular diseases, cancer, diabetes mellitus, and body mass index ≥30 at baseline.

Second, although the FFQ was not specifically designed to identify ultra-processed food, there is some evidence in certain populations (e.g., New Zealand children [64] and adults from Italy [65] and Mexico [66]) that FFQs have acceptable validity to assess food consumption based on the NOVA food classification system. FFQs have also been reported as the most frequently used dietary data collection tool in reviews investigating ultra-processed food–chronic disease relationships [8]. Nonetheless, some degree of misclassification bias may exist.

Lastly, while the FFQ dietary data used to investigate ultra-processed food intake in the current study were captured over 20 years ago, this may not limit the generalisability of our results since both the participant characteristics and level of ultra-processed food consumption reported in the current study are comparable to current estimates globally. As specified in Table 1, ultra-processed food contributed 40% of total energy intake, which is consistent with the most recent analysis of ultra-processed food intake in a nationally representative sample of Australians taken from the National Nutrition and Physical Activity Survey (2011–2012) [28], where ultra-processed food contributed 42% of total energy. This is also comparable with other estimates from Western countries such as Canada (42%), the United Kingdom (54%), and the United States (56%) [8].

#### *4.2. Implications*

Historically, and as previously noted, nutrition research has focused on the effect of dietary intakes of energy and macro- and micro-nutrients on human physiology and health, including inflammatory processes together with the mechanistic link between inflammation and chronic diseases. The relevance and novelty of the NOVA food classification system becomes prominent, however, when considering emerging evidence for differential health outcomes that depend on the extent and level of food processing [67–70]. Indeed, NOVA largely ignores the nutrient profiles of ultra-processed food, instead focusing on the extent and purpose of food processing [11]. One tightly controlled randomised trial in humans that specifically applied the NOVA food classification system in its design [7] demonstrated a causal effect of an ultra-processed versus unprocessed diet on increased energy intake as well as adiposity (both of which have been associated with pro-inflammatory states [71]). While this landmark study targeted different metabolic outcomes, it also showed a withingroup reduction from baseline to endpoint in hsCRP concentration when participants were allocated to the unprocessed diet [7]. It also underscored the futility of focusing only on nutrient composition given that the two diets were matched for presented energy, sugar, fat, fibre, and macronutrients [7].

Given the association between ultra-processed food intake and morbidity and mortality [8], there were recent calls urging countries to adopt policy interventions that limit the production, distribution, and dietary intake of ultra-processed food [72]. The importance of our study includes its potential generalisability to other Anglo-European populations and ability to inform and encourage future research investigating the possible biological mechanisms of action involved in the observed associations between consumption of ultra-processed food, chronic non-communicable diseases, and all-cause mortality.

#### **5. Conclusions**

The current study showed a cross-sectional association between higher ultra-processed food intake and elevated hsCRP as a biomarker of inflammation. Part of the association between consumption of ultra-processed food and hsCRP was independent of body mass index. Further prospective and experimental studies in humans are needed to examine whether this association is causal. Such information will be key to appropriate health messages in the future.

**Supplementary Materials:** The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/nu14163309/s1, Figure S1: A directed acyclic graph mapping hypothesised relationships between all relevant variables; Table S1: Sensitivity excluding individuals with hsCRP concentrations above 10 mg/L and cardiovascular disease mortality; Table S2: Sensitivity analyses excluding individuals with history of non-communicable diseases; Table S3: Sex-stratified cross-sectional associations between the ultra-processed food intake and hsCRP concentration (MCCS, 1990–1994).

**Author Contributions:** Conceptualization, M.M.L. and W.M.; formal analysis, M.M.L. and M.L.; data curation, M.M.L. and M.L.; writing—original draft preparation, M.M.L.; writing—review and editing, M.M.L., M.L., M.F., A.L., T.R., A.O., P.M., F.N.J., A.H., and W.M.; supervision, W.M., M.L., A.L., T.R., A.O. and F.N.J.; project administration, M.M.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** The current study was approved for exemption from ethical review in accordance with the National Statement on Ethical Conduct in Human Research (2007, updated 2018) Section 5.1.22 by the Deakin University Human Research Ethics Committee (project number: 2020-413, received 18th of November 2020). The study protocol for the original Melbourne Collaborative Cohort Study project was approved by the Cancer Council Victoria's Human Research Ethics Committee (project number: IEC 9001, received 23rd of August 1990).

**Informed Consent Statement:** Participants provided written consent to participate including researcher access to their medical records.

**Acknowledgments:** MCCS cohort recruitment was funded by Cancer Council Victoria and VicHealth. The MCCS was further supported by Australian National Health and Medical Research Council (NHMRC) grants 209057, 396414, and 1074383, and ongoing follow-up and data management has been funded by Cancer Council Victoria since 1995. Cases and their vital status were ascertained through the Victorian Cancer Registry and the Australian Institute of Health and Welfare, including the National Death Index and the Australian Cancer Database.

**Conflicts of Interest:** The Food & Mood Centre has received Grant/Research support from Fernwood Foundation, Wilson Foundation, the A2 Milk Company, and Be Fit Foods. M.M.L. is supported by a Deakin University Scholarship and has received research funding support from Be Fit Foods. M.L. and P.M. are currently funded by an Alfred Deakin Postdoctoral Research Fellowship. M.F. is supported by a Deakin University Scholarship. A.L. has received grants, fellowships and research support from the University of New South Wales, the University of Melbourne, RMIT University, Deakin University, the

National Health and Medical Research Council (NHMRC), Australian Academy of Science, National Institutes of Health (NIH), and The Jack Brockhoff Foundation. A.L. has received honoraria and travel funds from Sydney University, the University of Technology Sydney, American Epilepsy Society, Epilepsy Society of Australia, International Human Microbiome Congress, European Society of Neurogastroenterology, Australian and New Zealand College of Anaesthetists, Falk Foundation and Fonds de la Recherche Scientifique (FNRS). T.R. has received grants, fellowships and research support from University of the Sunshine Coast, Australian Postgraduate Awards, Fernwood Foundation, Roberts Family Foundation, Be Fit Food and Wilson Foundation. T.R. received consultancy, honoraria and travel funds from Oxford University Press, the University of Melbourne, the University of Sydney, Bond University, University of Southern Queensland, Dietitians Association of Australia, Nutrition Society of Australia, The Royal Australian and New Zealand College of Psychiatrists, Academy of Nutrition and Dietetics, Black Dog Institute, Australian Rotary Health, Australian Disease Management Association, Department of Health and Human Services, Primary Health Networks, Barwon Health, West Gippsland Healthcare Group, Central West Gippsland Primary Care Partnership, Parkdale College, City of Greater Geelong and Global Age. A.O. is supported by a Future Leader Fellowship (#101160) from the Heart Foundation Australia and Wilson Foundation. She has received research funding from National Health & Medical Research Council, Australian Research Council, University of Melbourne, Deakin University, Sanofi, Meat and Livestock Australia and Woolworths Limited and Honoraria from Novartis. F.J. has received Grant/Research support from the Brain and Behaviour Research Institute, the National Health and Medical Research Council (NHMRC), Australian Rotary Health, the Geelong Medical Research Foundation, the Ian Potter Foundation, Eli Lilly, Meat and Livestock Australia, Woolworths Limited, the Fernwood Foundation, Wilson Foundation, the A2 Milk Company, Be Fit Foods, and The University of Melbourne, and has received speakers honoraria from Sanofi-Synthelabo, Janssen Cilag, Servier, Pfizer, Health Ed, Network Nutrition, Angelini Farmaceutica, Eli Lilly and Metagenics. F.N.J. has written two books for commercial publication and has a personal belief that good diet quality is important for mental and brain health. W.M. is currently funded by an NHMRC Investigator Grant (#2008971) and a Multiple Sclerosis Research Australia early-career fellowship. Wolfgang has previously received funding from the Cancer Council Queensland and university grants/fellowships from La Trobe University, Deakin University, University of Queensland, and Bond University. Wolfgang has received industry funding and/or has attended events funded by Cobram Estate Pty. Ltd. and Bega Dairy and Drinks Pty Ltd. Wolfgang has received travel funding from Nutrition Society of Australia. Wolfgang has received consultancy funding from Nutrition Research Australia and ParachuteBH. Wolfgang has received speakers honoraria from The Cancer Council Queensland and the Princess Alexandra Research Foundation.

#### **References**


### *Article* **Ultra-Processed Foods as Ingredients of Culinary Recipes Shared on Popular Brazilian YouTube Cooking Channels**

**Anice Milbratz de Camargo 1, Alyne Michelle Botelho 1, Állan Milbratz de Camargo 2, Moira Dean <sup>3</sup> and Giovanna Medeiros Rataichesck Fiates 1,\***


**Abstract:** Social media platforms are readily accessible sources of information about cooking, an activity deemed crucial for the improvement of a population's diet. Previous research focused on the healthiness of the content shared on websites and blogs, but not on social media such as YouTube®. This paper analysed the healthiness of 823 culinary recipes retrieved from 755 videos shared during a six-month period on ten popular Brazilian YouTube® cooking channels. Recipes were categorized by type of preparation. To assess recipes' healthiness, ingredients were classified according to the extension and purpose of industrial processing, in order to identify the use of ultra-processed foods. Additionally, a validated framework developed from criteria established in both editions of the Dietary Guidelines for the Brazilian Population was employed. Recipes for cakes and baked goods, puddings, snacks and homemade fast foods, which were among the most frequently posted, contained the lowest proportion of unprocessed/minimally processed ingredients and the highest proportion of ultra-processed ingredients. Recipes containing whole cereals, fruits, legumes, nuts, and seeds were scarce. Results indicate that users should be critical about the quality of recipes shared on YouTube® videos, also indicating a need for strategies aimed at informing individuals on how to choose healthier recipes or adapt them to become healthier.

**Keywords:** social media; social network site; Internet; cookery channels; recipe quality; cooking instruction; ultra-processed foods

#### **1. Introduction**

Public health initiatives from many countries encourage home cooking as a healthpromoting strategy [1]. This is also true for both editions of the Dietary Guidelines for the Brazilian Population [2,3], which adopt distinct but complementary approaches for the promotion of healthy eating. The first edition of the Guidelines valued the act of eating at home and provided information on how to prepare food in a healthy way. Its directives were based on the intake of adequate amounts of foods, classified into food groups, to prevent nutritional deficiencies and chronic non-communicable diseases [2].

Aside from stressing the importance of home cooking, the Dietary Guidelines for the Brazilian Population published in 2015 focused on categorizing foods according to the extension and purpose of industrial processing [3]. Individuals should base their diets on unprocessed/minimally-processed foods (U/MP) and avoid ingesting ultra-processed foods (UP) as much as possible [3]. UP foods are formulations of ingredients that are usually nutritionally unbalanced, being rich in fats and sugars while poor in fibre and micronutrients [3]. Carbonated soft drinks, packaged snacks, mass-produced breads,

**Citation:** Camargo, A.M.d.; Botelho, A.M.; Camargo, Á.M.d.; Dean, M.; Fiates, G.M.R. Ultra-Processed Foods as Ingredients of Culinary Recipes Shared on Popular Brazilian YouTube Cooking Channels. *Nutrients* **2022**, *14*, 3689. https://doi.org/10.3390/ nu14183689

Academic Editors: Monica Dinu and Daniela Martini

Received: 5 July 2022 Accepted: 7 August 2022 Published: 7 September 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

margarines, candies, cake mixes, and many ready-to-heat frozen products (pies, pizza, sausages, burgers) are examples of UP foods [3,4].

High consumption of UP foods has been associated with chronic non-communicable diseases and all-cause mortality [5–7]. Conversely, cooking at home more often has been associated with a lower risk of developing chronic non-communicable diseases [8,9], possibly as a result of a better diet quality [9,10] due to the use of fresh ingredients. A pattern of healthy cooking practices, where individuals can confidently cook several meals using fresh foods and natural seasonings, and use healthier cooking techniques, was inversely associated with ultra-processed food consumption [11]. A diet composed mostly of U/MP foods, however, can only be achieved if individuals master a certain number of cooking skills [3].

Informal cooking education happens through culinary socialization over the course of a person's life, a process in which individuals acquire patterns of practices and perceptions related to cooking, from socializing agents [12]. The first culinary socializing agents are family members; later in life, different agents start to influence cooking practices, such as friends, partners, cookbooks, culinary television programs, and more recently, the Internet [9,13–15]. Individuals report favouring Internet searches and digital sources when looking for recipes, instead of printed sources such as books, for the convenience of being 'at hand' [13].

Brazilians spend an average of 3.5 h daily on the Internet [16], mainly accessing social media [17]. Social media platforms have become accessible sources of information regarding cooking-related matters—people use Facebook®, Instagram®, Pinterest® and YouTube® to share and search for recipes, and to find meal suggestions and inspiration [13,18–20].

YouTube® was created in 2005 and works as a video sharing platform, which is accessible via personal computers or smartphones through an Internet browser or application [21]. On the platform's homepage, an algorithm suggests videos based on visualization history and the popularity of the content, among other information. Users can also actively search for videos using keywords or browsing channels. A user can interact with a video by watching it, liking, sharing with others, and/or publicly commenting, all of which are important social media features [22–24].

Previous research mentions that YouTube® is one of many people's favourite ways to learn how to cook [20]. Understandably, when compared to just text and images, recipes shared through video technology can favour user engagement, increase the motivation to cook, and reduce the perception of time, skills, and cost barriers [25]. Video recipes also potentially assist with the development of new skills, increase the pleasure of cooking, provide real-time assurance during the cooking process, help people remember the steps, and improve the understanding of the process [26]. In Brazil, YouTube® is the most popular social media platform among individuals aged between 16 and 64 years [16].

Accessing the Internet to search for recipes, learn how to cook, and develop cooking skills is recommended by the Dietary Guidelines for the Brazilian Population [3], but the healthiness of recipes obviously depends on the ingredients and preparation methods employed [11]. In this sense, exploring the sources of knowledge and inspiration to cook is as key as getting people to cook more often. As tools that guide the preparation of dishes [27], culinary recipes can potentially promote health if aligned with recommendations for healthy eating, expanding and encouraging individuals' decision-making autonomy regarding the adoption of healthy eating practices [2,18]. However, in the context of social media, content can be produced and shared by anyone, including lay people not qualified to give nutritional advice or create content that promotes healthy eating.

Previous studies assessed the healthiness of Internet recipes on websites and blogs (which are not social media), and concluded that users tend to interact more often with the least healthy recipes [28]. Authors concluded that even recipes tagged as 'healthy' are often quite unhealthy [28,29]. We identified only one paper on the healthiness of culinary recipes on social media, which used Pinterest® as a data source [30]. The paper reported that recipes using seafood or vegetables as main ingredients had fewer calories, sodium, sugar, and cholesterol than meat- or poultry-based recipes. However, the study's sample

was small due to the adoption of many exclusion criteria [30]. No research investigating the healthiness of culinary recipes shared on other social media was found.

To address this gap, this descriptive and exploratory study analyses the healthiness of culinary recipes shared on popular YouTube® cooking channels from Brazil, using both national dietary guidelines as references. We adopted complementary approaches to assess recipes' healthiness, the first being the analysis of recipes' ingredients according to the extension and purpose of industrial processing, an important and widely used approach to categorize foods. Subsequently, a specially designed qualitative framework was used to characterize recipes according to cooking method, and by the presence of healthy or unhealthy ingredients. We believe this study has the potential to inform the design of public health initiatives that guide individuals and inform dietitians on how to select and critically evaluate sources of cooking information, and improve the quality of homecooked meals.

#### **2. Materials and Methods**

Considering the scarcity of literature on the research topic, a pilot study was carried out to inform the data collection protocol, which included aspects of various channels' eligibility criteria and database layout, introduced different video characterization variables, validated recipes' assessment method, and determined the data collection period, taking into account the temporal feasibility of the study and the amount of content to be analysed [31].

#### *2.1. Selection of YouTube® Cooking Channels*

Cooking channels were purposely selected by taking into consideration that popularity (number of subscribers) can promote a greater reach and be a proxy for users' preference. Channels were selected according to the number of subscribers in February 2020 using The YouTube Channel Crawler page (https://www.channelcrawler.com/, accessed on 10 February 2020), which classifies channels according to criteria established by the researcher (in this study: category, language, country of origin, and number of subscribers). During the pilot study, it was observed that cooking channels belonged to the 'How to and Style' category on YouTube®, thus, all channels of the platform within that category were accessed in decreasing order of subscribers to identify which ones best fit the eligibility criteria. The ten biggest channels which (1) presented audio-visual content in Portuguese and were Brazilian based; (2) were a cooking channel; (3) posted culinary videos at least once a week; and (4) were not an advertising channel or reproduced television cooking programs were selected (Figure 1).

With the aim of having a high number of videos to be analysed during data collection, it was established that channels that posted videos less than once a week would not be included. The pilot study also revealed that some channels which were among the most popular in terms of number of subscribers had suddenly stopped producing content in the weeks preceding the selection of channels. They were not included to avoid the possibility of not having enough content to analyse in the following months. Another reason for adopting this criterion was to try to standardize the number of videos per channel. Novelty was another important factor, as channels need not only to attract, but also maintain users' interest and engagement with content [23].

Eighty-two channels were excluded from the sample because they were not cooking channels, five were excluded because they did not post videos with the desired frequency, and one was excluded for being an advertising channel.

The included channels were mostly presented by women (*n* = 7), two by men, and one by a couple; none of them were popularly known chefs or food celebrities. Subscribers ranged from 514 thousand to 4.25 million; channels' time of existence ranged from 4 to 9 years, and posting frequency varied from 2 to 7 videos per week.

**Figure 1.** Brazilian YouTube® Cooking channels' selection flowchart, February 2020.

#### *2.2. Selection of Recipes*

A sample of 823 recipes presented in 755 videos (104 h and 21 min in total) posted during a six-month period (from February to August 2020) on ten different cooking channels was selected. Considering that this is a recent field of study and there is no specific recommendation in the literature for how long data collection on YouTube® should take place, the pilot study also informed the choice of an appropriate data collection period. With the pilot study, we were able to project that a 6-month data collection period would capture a high number of videos from each channel, carefully accounting for at least three seasons of the year. At the same time, the amount of content collected would meet the temporal and operational feasibility criteria of the study.

All videos with recipes posted within the period were watched in full (first author) to determine if they contained all the ingredients needed, as well as the preparation method. A total of 106 videos were excluded from analysis because they did not meet the eligibility criteria: (1) were recorded live transmissions (*n* = 35), (2) presented a festive recipe (Easter *n* = 20; Mother's day or father's day *n* = 6; Valentine's day = 4; June festivities in Brazil *n* = 12; Channel's subscribers milestone celebration *n* = 4; total *n* = 46), (3) were sponsored by the food industry (*n* = 9), (4) were a repost (*n* = 7), (5) presented recipes linked to the COVID-19 pandemic (with connotations of treatment for the virus, for improving immunity or with tips for food sales during the period of social isolation; *n* = 9).

Reasons for not including recorded live transmissions were: (1) during the pilot study, we observed that those kinds of videos were usually presented as 'extra' content and were produced by only four of the ten channels. They were not included for the sake of standardization (type and number of videos per channel). (2) 'Live' transmissions lasted more than one hour each, as the recipe-related content was diluted among various other content during the video. This affected both the practical relevance of the recipe and the temporal feasibility of the research.

#### *2.3. Data Collection*

Weekly, from February to August 2020, each selected channel was accessed via computer and all videos posted during the previous week were registered. A database in Microsoft Excel 2016® was created to include the following information for each video: title, access link, ID provided by YouTube®, video description, date of posting, date of access, duration in seconds, number of likes, dislikes, and views. To obtain the number of comments posted by users in each video, a command line application was developed in Python 3.0 (third author). Using the public and free Google Data application programming interface (API) service® as the data source, the application generated automated reports from the video ID and the period determined by the researcher (freely available at https://bitbucket.org/amcamargo/healthy-recipe-youtube-br.git, last accessed on 11 August 2020).

Next, the first author watched each video to register the ingredients and the cooking method in the database. If further details about ingredients were needed, the researcher consulted the recipe's ingredient list provided in the video description, or, in case of industrialized products, the packaging, when information was clearly visible on screen. Steps or ingredients mentioned by the youtuber as 'optional' and not shown in the video were not assessed.

#### *2.4. Data Analysis*

#### 2.4.1. Videos' Characteristics

Variables assessed to characterize the videos were duration in minutes, day of the week of posting, and interaction measures including popularity as daily views in the first week, approval as daily likes and dislikes in the first week, direct interaction of users with content through daily comments in the first week, and total comments in the first week and in the first month after the video was posted.

To classify the recipe into a category (e.g., salad, pudding, etc.) a content analysis was carried out based on the video's title, description, and list of ingredients used. This analysis was manually organized in Microsoft Excel 2016® by determining the degree of similarity of the words and phrases used and the characteristics of the recipes, starting at coding recipes' names in videos' titles (first author). After coding, data was categorized until strong or terminal categories appeared [32].

#### 2.4.2. Recipes' Healthiness

Recipes had their ingredients classified according to the extension and purpose of industrial processing, as unprocessed/minimally processed (U/MP), processed culinary ingredient (PCI), processed (P), or ultra-processed (UP) (first author) [3,4,33]. Ingredients that did not have their preparation described in the recipe but are available for purchase as an industrialized version were classified as P or UP (e.g., sweetened condensed milk, mayonnaise), according to the predominant characteristic of products available in Brazilian retail outlets. Whenever agreement about the extension and purpose of industrial processing was not achieved, a conservative criterion was applied, meaning that a lower extension of processing was adopted for the ingredient [34]. Ingredients used twice in the same recipe counted as one (e.g., sugar used in a cake's batter and icing).

Subsequently, the Qualitative Framework for the Assessment of Culinary Recipes' Healthiness [31] was applied to evaluate recipes' cooking methods and presence of key healthy and unhealthy ingredients (first author). The framework was specifically developed and validated to assess culinary recipes' healthiness, and was based on recommendations for healthy eating retrieved from both Dietary Guidelines for the Brazilian Population [2,3].

#### 2.4.3. Data Treatment

To ensure data quality control, the second author independently analysed 10% of the recipes from the dataset. Weighted kappa of agreement between raters for the assessment of ingredients' extension and purpose of industrial processing was 0.96, and ranged between 0.90 and 1.00 (kappa and weighted kappa) for the application of the Qualitative Framework for the Assessment of Culinary Recipes' Healthiness, indicating almost perfect agreement in both analyses [35]. Content analysis for the categorization of recipes was firstly discussed between the first two authors, and divergences were resolved with the participation of the last author.

#### 2.4.4. Statistical Analysis

Qualitative dichotomous and polytomous variables are presented in absolute and relative frequencies. Quantitative variables are presented as median and interquartile range (IQR), considering the non-normality in data distribution when assessed by Shapiro–Wilk test, histogram, kurtosis value, and mean/median proximity.

Variables of videos' characteristics and recipes' healthiness among the categories of recipes were compared. Also, as data collection took place mostly during a social distancing period due to the COVID-19 pandemic, when searches for recipes online increased [36], we also checked for differences in videos' interaction measures (popularity, approval, interaction through comments), and recipes' healthiness in the periods preceding (*n* = 141) vs. during social isolation (*n* = 614) (which, in Brazil, started around 15 March). Mann–Whitney and Kruskall–Wallis tests were used for quantitative variables. For qualitative variables, Pearson's chi-square test was employed. Stata 13.0® (StataCorp LLC, College Station, TX, USA) was used for analysis and a post-hoc power analysis was applied on G\*power 3.1.9.2 whenever necessary, considering a two-tailed test. An alpha of 0.05 was established as the significance level for all analyses.

#### **3. Results**

#### *3.1. Videos' Characteristics*

The videos' durations ranged from 45 s to 27.33 min (*n* = 755). The number of daily likes in the first week was superior to daily dislikes. The option of liking or disliking a video was not enabled by the youtubers for all videos (only *n* = 611), therefore, even if users wanted to give a particular video a thumbs-up or down, they could not. Direct interaction through comments was concentrated in the first week after the videos were posted, as the median of total comments in the first month was close to the median in the first week. Sunday was the day of the week with the lowest number of videos posted, nevertheless, the distribution of videos was similar among the other days (Table 1).

**Table 1.** Videos' characterization variables (*n* = 755).


Footnote: <sup>1</sup> *n* = 611 videos.

The only observed difference between videos collected in the period preceding vs. during social isolation was in the total of comments in the first week, which was higher during the social isolation period (median = 154, IQR = 64; 280) than before the pandemic (median = 140; IQR = 47; 234) (Mann–Whitney's *p* = 0.04, power = 0.19).

More than two thirds of all recipes (68.1%) comprised preparations from only four categories, namely: meat or egg main dishes; cakes and baked goods; snacks and homemade fast foods; and puddings (Table 2). The sixteen different categories of recipes had comparable video characteristics (all Kruskall–Wallis *p* > 0.10; χ<sup>2</sup> = 91.19, *p* = 0.445). The frequency of categories observed in the period preceding vs. during social isolation was statistically the same (χ<sup>2</sup> = 18.25; *p* = 0.07).





U/MP—unprocessed/minimallyprocessedPCI—processedculinaryingredients.P—processedUP—ultra-processedrecipe small cheese bread made of fermented tapioca flour. 2 Traditional Brazilian dish made of manioc flour fried in fat, which can be enriched with other ingredients.

#### *3.2. Recipes' Healthiness*

Of the total 7814 ingredients analysed, the majority were U/MP (54.3%, *n* = 4242) and PCI (23.6%, *n* = 1844). Ingredients classified as P (8.6%, *n* = 676) and UP (13.5%, *n* = 1052) were less frequent. The categories of recipes differed in terms of the ingredients' distinct extension and purpose of industrial processing (χ<sup>2</sup> = 859.22; *p* < 0.001). As Table 2 shows, in many categories, less than half of the ingredients were U/MP, i.e., cakes and baked goods, snacks and homemade fast foods, puddings, breads, sweet and savoury spreads, and pâtés. The ten most frequent U/MP ingredients in the sample were, in decreasing order: water, eggs, onion, all-purpose flour, garlic, milk, black pepper, oregano, spring onions, and tomatoes. Some of the categories with the lowest frequency of U/MP foods also had the highest frequencies of UP foods in the sample, i.e., puddings, cakes and baked goods, snacks and homemade fast foods, sauces, sweet and savoury spreads, and pâtés. The ten most frequent UP ingredients in the sample were, in decreasing order: UHT cream, sweetened condensed milk, Brazilian cheese spread, margarine, ham, industrialized tomato sauce, spicy sausage, vanilla essence, industrialized seasoning mix, and semi-sweet chocolate. The frequency of ingredients with distinct extension and purpose of industrial processing observed in the period preceding vs. during social isolation was not statistically different (χ<sup>2</sup> = 0.68; *p* = 0.877).

Application of the Qualitative Framework for the Assessment of Culinary Recipes' Healthiness (Table 3) identified positive and negative aspects of the recipes. Positively, most recipes that mentioned some type of fat as an ingredient did not suggest the use of margarine (88.4%, *n* = 518). Mentions of tomato sauce with herbs (bottled or freshly made) were more frequent than exclusive mentions of white sauce with mayonnaise or cheese (69.6%, *n* = 131). Exclusive use of industrialized seasonings (1.5%, *n* = 7) and of frying as a cooking method (7.9%, *n* = 60) was also not frequently mentioned. On the other hand, the presence of whole cereals, breads and/or pasta, either exclusively or mixed with refined cereals was low in the recipes (7.1%, *n* = 34), as well as were the presence of fruits (13.7%, *n* = 111), legumes (4.5%, *n* = 37), and nuts and seeds (3.5%, *n* = 28). The categories that presented the most evenly distributed positive and negative criteria were types of meats, presence of foods with high sugar concentration, and presence of vegetables. All results from the framework analysis were statistically the same regarding the period of data collection (preceding vs. during social distancing; all 0.01 < χ<sup>2</sup> > 4.73 and *p* > 0.07).


**Table 3.** Recipes' healthiness according to the Qualitative Framework for the Assessment of Culinary Recipes' Healthiness.


**Table 3.** *Cont.*

Footnote: Fruits, vegetables, and legumes; nuts and seeds, and sugars categories are mandatorily assessed in all recipes. The remaining categories are assessed only when applicable. Criteria: + and − indicate recommended and not recommended components for healthy recipes, respectively [31].

#### **4. Discussion**

This study analysed the healthiness of recipes shared on popular YouTube® cooking channels from Brazil using the Dietary Guidelines for the Brazilian Populationas references. Recipes posted during a six-month period were retrieved and categorized into sixteen different groups. The most frequently posted recipes were of meat/egg-based main dishes; cakes/baked goods; snacks/homemade fast foods; and puddings. This means that recipes for salads and side dishes, which usually contain vegetables, fruits, and legumes, were shared less often than recipes with animal sources of protein, all-purpose flour, fats, and sugar as the main ingredients. This result is not favourable from a health standpoint, as individuals are possibly being led to prepare fewer recipes with fruits, vegetables, and legumes, which are linked to a lower risk of chronic non-communicable diseases, and are largely present in most healthy eating patterns [37,38]. Interestingly, the study by Trattner and Elsweiler (2017) identified different results—in their study, which evaluated content from a recipes' website, the category 'fruits and vegetables' was much more prevalent than 'main dishes,' 'meat and poultry,' 'desserts,' and 'salads.' This disparity may be attributed to differences in the process of categorizing recipes, as in food blogs [29,39] and websites [28], recipes are usually pre-categorized, while we conducted our own categorization. Because YouTube® is a multi-content platform not specifically focused on recipes, our recipe categories were qualitatively and inductively generated from recipes' titles, descriptions, and ingredients. Additionally, several studies only assessed specific categories of recipes [28–30,39], since their aim was not to have an overall picture of what is shared.

Another possible explanation for the low prevalence of fruit- and vegetable-based recipes in our sample may be that content producers expect users to interact with the postings through comments and shares, as interaction is fundamental for a channel's engagement and sustainability [23]. It has been reported by previous studies on a recipes' website [28] and on Pinterest® [30] that interaction is more frequent with posts of highly palatable recipes. In our study conducted on YouTube®, every culinary preparation had statistically equal measures of interaction (popularity, approval, and direct interaction through comments), possibly due to differences between the profiles of users from recipe websites [28] and even between different social media apps [30]. YouTube®, as a video platform, enables a kind of interaction that gives users a feeling of being connected not only to a video, but to a person who shares their beliefs and interests. This feature can promote a certain measure of social bonding in which people feel connected with one another and start following the channel for further communication. For user-created content such as the videos analysed, a sense of community is fundamental; so it is possible that subscribers give the same attention to a recipe, regardless of whether it is a salad or a cake, in order to provide support through constancy of viewership and interaction [21]. The reasoning behind youtubers' choices of categories of recipes for cooking videos, the channels' features that promote connection with users, as well as subscribers' motivations for interaction with content, deserve to be further explored in future research. Nevertheless, health professionals should be aware that, in order to expose individuals to more recipes based on vegetables, fruits, and legumes (such as salads and side dishes), active searching is preferable to just following content from popular cooking channels. Nutritionists and other health professionals can also search for cooking channels whose content is more in line with the healthy eating recommendations of national guidelines to suggest to patients.

Recipes' ingredients were mainly U/MP foods and PCI, so one can argue that from a wide perspective, the recipes could lead individuals to cook recipes that are aligned with the recommendations of the Dietary Guidelines for the Brazilian Population [3]. Nevertheless, this is not true when different categories of culinary recipes are considered. Some categories of recipes had lower frequencies of U/MP foods as ingredients, and a few of them had, in addition to this, higher frequencies of the UP foods in the sample (more than 10%)—i.e., puddings, cakes and baked goods, snacks and homemade fast foods, sauces, sweet and savoury spreads, and pâtés. This result is cause for concern, as some of these were among the most frequently posted recipes. To cook healthily, the Dietary Guidelines for the Brazilian Population recommends the avoidance of UP foods [3], as high consumption of UP foods has been associated with chronic non-communicable diseases and all-cause mortality [5–7]. UP food consumption has been associated with a poor dietary intake (excess calories from free sugars and unhealthy saturated fats, poor in fibre, and an intake of many micronutrients) [40]. Additionally, recent research shows that the majority of the associations between UP food consumption, obesity, and health-related outcomes can be attributed to UP foods on their own, regardless of diet quality or pattern [41].

The presence of UP ingredients in the recipes may be explained by their convenience appeal [4,33]. It is rather common for UP foods to replace U/MP foods in recipes (e.g., sausage vs. U/MP meat seasoned with spices and herbs). Generations have learned to cook using recipes combining UP and U/MP foods through teaching investments by the food industry (leaflets, books, courses, recipes on packaging) [42]. Nowadays, the ongoing increase of options of UP foods and of social media marketing play an important incentivizing role [43]. We observed with our framework analysis (Table 3) that the mixed use of industrialized seasonings with fresh or dried herbs and spices was frequent, indicating an attachment to this type of UP product. To mitigate this effect, strategies involving the promotion of healthy eating through cooking (such as workshops, intervention programs, creation of content for social media, health professionals' advice, etc.) need to consider that people must be taught how to identify UP foods so they can choose recipes in which they are not included. People must also be taught how to substitute UP foods for healthier ingredients, so they can use U/MP foods practically when cooking. For instance, instead

of relying on UP foods as seasoning in puddings, snacks, and homemade fast foods as observed in this sample of recipes, one can substitute such ingredients for fruit zest and juices, fresh or dried herbs, and spices. Another valuable strategy is to rescue and promote the sharing of traditional recipes that do not contain UP foods as ingredients.

While the majority of recipes were healthy with respect to avoiding the use of margarine, avoiding frying, and opting for sauces with lower fat content, other aspects such as incorporating whole cereals, fruits, legumes, nuts, and seeds in preparation were not frequently present. Considering that social media platforms such as YouTube® reach a wide audience, this finding reinforces the need to not only encourage people to look for recipes online [3], but also to teach them how to choose or adapt these recipes by evaluating their healthiness. One strategy is to use the same medium to do this, as video technology can help individuals overcome barriers to cook and incorporate healthier foods in recipes [44], while reducing the perception of barriers to cook with vegetables [25]. We are aware that, for some recipes, whole cereals, fruits, legumes, nuts, and seeds may not be all traditionally present (e.g., a basic homemade bread), but different 'improved' versions of recipes can be proposed and shared. As a matter of fact, many channels assessed in this study adapted recipes to keep producing new content weekly. As a practical implication, we argue that many categories of recipes can be adapted to become healthier—for examples and suggestions, see [31]. Members of academia, health professionals, and social media content creators can also work together and establish partnerships to promote healthier content on the Internet.

#### *Limitations and Strong Points*

The adoption of a conservative criterion for classifying ingredients by the extension and purpose of industrial processing may have led to an underestimation of the number of UP ingredients. Nevertheless, this approach mirrors how information reaches users—they also do not necessarily have access to information on labels when watching videos.

Data collection took place during months of social isolation due to the COVID-19 pandemic, when searches for recipes online increased [36]. This was handled by avoiding the inclusion of videos linked to the COVID-19 pandemic in the sample. Our post-hoc analysis found an underpowered difference in the number of comments in the first week after videos were posted (power = 0.19) [45]; no change in the categories of recipes shared, nor in their healthiness compared to videos from before the pandemic.

YouTube® channels' popularity oscillates constantly. To handle this, we repeated the channel selection step at the end of data collection, and verified that they remained as the ten most popular in the period, despite some outperforming others in the number of subscribers.

As the number of views is validated by YouTube®'s own algorithms, a view might not indicate a user who has watched the content in its entirety. Content approval (likes and dislikes) also does not indicate whether or not an individual fully watched the content before giving a positive or negative rating. Although views made by computer programs rather than by humans are not counted [46], those interaction measures should be cautiously interpreted [47].

In the context of television cooking shows, some researchers argue that the consumption of this content is unlikely to impact habitual dietary intake, because entertainment and leisure are the main reasons people watch those programs [48,49]. Notwithstanding, social media, through its networked nature, provides an additional layer of complexity not experienced by those earlier media scholars [47]. Through observation, people indeed acquire behaviours, knowledge, values, and skills, including those related to cooking [50].

We understand that it is not possible within the confines of the present study to account for variations in the reproduction of recipes at home, such as instances when people do not follow all the steps, or when ingredients are exchanged, which may result in a different assessment of their healthiness.

As positive points, we highlight the investigation of culinary recipes posted in the most used social media in Brazil, by adults [16]; the rigorous quality control; the long period of data collection throughout three seasons of the year, and, therefore, the large sample size. Recipes were also very diverse in terms of categories, video duration, and days of posting, probably reaching different types of audiences. Additionally, collecting measures of interaction (views, comments, likes, and dislikes) reinforces the wide reach that this type of content has. Finally, using a validated framework for the assessment of recipes' healthiness, we were able to deliver a more specific picture of the research problem.

#### **5. Conclusions**

This study provides a comprehensive overview of the healthiness of culinary recipes shared on a social media platform, one of the favoured avenues for the search of cookingrelated content. On a professional practice and health promotion note, although it is praiseworthy that people are cooking and sharing their knowledge on platforms such as YouTube®, users and subscribers to popular cooking channels should be aware that most recipes are based on ingredients such as meats, eggs, all-purpose flour, fats and sugar, and only a few have whole cereals, fruits, legumes, nuts, and seeds. Recipes for puddings, cakes and baked goods, snacks and homemade fast foods, sauces, sweet and savoury spreads, and pâtés had, in addition to low numbers of U/MP food ingredients, higher numbers of UP foods as ingredients. Our findings can inform health professionals and policymakers on how to promote healthier culinary recipes, how to interact with content creators, and how to advise individuals about the quality of the recipes shared on YouTube® videos, and hence, can help them choose healthier recipes or teach them how to modify the recipes into healthier versions. Future research exploring how users from different populational groups interact with culinary content on distinct social media platforms will be relevant for advancing this field of study.

**Author Contributions:** Conceptualization, A.M.d.C.; data curation, A.M.d.C. and Á.M.d.C.; formal analysis, A.M.d.C., A.M.B. and G.M.R.F.; methodology, A.M.d.C. and G.M.R.F.; software, Á.M.d.C.; validation, A.M.d.C., A.M.B. and G.M.R.F.; writing—original draft preparation, A.M.d.C.; writing review and editing, A.M.d.C., A.M.B., Á.M.d.C., M.D. and G.M.R.F.; supervision, G.M.R.F.; funding acquisition, M.D. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by Newton Mobility Grant Scheme 2015 [Award Reference: NG150026. 2 to M.D. and G.M.R.F.], the UK Academies Fellowships Research Mobility, and Young Investigator Awards for UK Researchers in Brazil FAPESC/CONFAP/FUNDO NEWTON [Call N◦ 02/2017; 3 to M.D. and G.M.R.F.]; Coordination for the Improvement of Higher Education Personnel (CAPES) in the form of scholarships [Finance code 001] to [A.M.d.C. and A.M.B.]. Founding sources had no role in study design, in the collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

