Ultra-Processed Food: Assessing Uncertainty in an Imperfect World

Nutrition Science Involves Uncertainty

Unlike pure mathematics and physics, biology is inherently noisy and evidence is rarely completely certain. Yet, a broad range of perspectives have been positioned at the extreme ends regarding ultra-processed food (UPF), from being a panacea of diet-related disease and that processing, not nutrients, is what matters, to being of no value with there being no agreed consensus definition, that it is not possible to measure ‘actual’ UPF intake, or no direct evidence of ‘processing’ per se. Many arguments take the form of absolute statements, despite the fact that evidence is never perfect and rarely clear cut in the world of diet and nutrition.

To reflect upon this, I draw parallels to the literature on carbohydrates and fiber by the UK Scientific Advisory Committee on Nutrition (SACN), and evidence on Mediterranean diets, which despite a wealth of evidence supporting their value, do not have globally agreed consensus definitions. For example, the UK government estimates the energy content of carbohydrates as monosaccharide equivalents at 3.75 kcal/g, compared with Europe and the US which estimate the energy content from carbohydrates by weight using the difference method, and estimated at 4 kcal/g. ¹ SACN have also highlighted that there is no universal definition for the term ‘dietary fiber’, and a globally standardized definition of ‘wholegrain’ or wholegrain foods has not been established, despite strong evidence to support intake. ² These differences and inconsistencies do not preclude their use and value in nutrition science.

Thus, we must evaluate evidence of validity as a matter of relative degree, rather than in absolutes as a dichotomy. ³ For evidence linking nutrients, foods and diet patterns with health, we must do this by considering the source and extent of uncertainty in the evidence and then evaluating whether the uncertainty may meaningfully impact conclusions. This requires considering evidence from multiple angles.

Defining UPF According to the Nova Classification

The Nova classification considers the drivers shaping food environments (ie, the reason for creating a new food or drink product, or modifying an existing one). In particular, the strong commercial incentives shaping product development that act to optimize favorable food/drink properties that make the product more desirable (taste, texture and mouthfeel, smell, sight), attractive, affordable and available (eg, marketing and labelling strategies). These commercial incentives form reinforcing cycles to maintain and increase profit, by displacing minimally processed foods (MPF), and outcompeting other rival UPF, in a continual process of survival of the fittest. I discuss the parallels of this concept with the ‘junk food’ cycle first highlighted in the UK National Food Strategy.^4,5 The purpose within the Nova classification takes a broader consideration of food systems to understand collectively why and how our food environment has changed.

However, food companies are under no requirement to report product sales data of food and drink products, meaning it is not possible to determine which products may undergo this cycle. Therefore, in order to apply the Nova classification concept to diet assessment tools, nutrient databases and product lists, an operational/practical definition is used, which involves identifying markers of ultra-processing in product ingredients lists. These markers include additives with primarily cosmetic functions, and industrial ingredients rarely or never used in household kitchens. Both markers are used to identify products that have been optimized for desirability, attractiveness, affordability and availability. The use of an operational definition to define an overarching concept (ie, markers of ultra-processing to define UPF) is not novel; and is analogous to defining products as ’less healthy’ using a nutrient profiling model (NPM). However, this marker-based approach can be complicated by the fact that some additives can serve multiple functions (cosmetic and not cosmetic) and different functions, depending on the product in question. Importantly, markers should not be misinterpreted as being the key causal mechanisms between any studies purportedly linking UPF intake and any health outcomee. ⁶

Importantly, the definition of UPF according to the Nova classification does not discriminate against industrial food processing per se (ie, is not defined by it), nor uses additives that serve to improve the safety and durability of food, nor micronutrient fortification as markers to identify UPFF.^6,7 Industrial processing techniques span all Nova groups (eg, milk pasteurization in MPF, cheese production in PF), additives that serve to protect original properties, prevent microorganism proliferation or extend shelf life span all Nova groups (eg, antioxidants and preservatives in PF), and micronutrient fortification spans all Nova groups (eg, fortified flours in MPF, iodised salt in PCI).

Key Areas of Uncertainty When Examining Health Impacts Linked With UPF Intake

Uncertainty regarding examining associations between UPF intake using the Nova classification and health outcomes can be broadly split into 2 groups. The first being uncertainty that is specific to UPF and the Nova classification. This includes (1) issues in estimating UPF intake, such as identifying UPF or agreement in inter-rater coding of diet assessment tools; (2) UPF does not add any new additional value beyond already known links between nutrients or food groups and health; (3) heterogeneity across different UPF based on nutrient content, and not all UPF being ‘unhealthy’; and (4) a lack of biological plausibility linking UPF with some health outcomes.

The second group relates to uncertainties that are not specific to UPF and the Nova classification, and which apply across nutrition science and further afield. This includes the largely observational evidence base, and with it, unmeasured confounding and determining causality (or ruling out reverse causality).

Indeed, both the UK SACN and US Dietary Guidelines for Americans Committee (DGAC) highlighted a number of these issues when evaluating current evidence of ultra-processing and health.^8,9 However, the UK SACN reports did not formally examine the extent to which the suggested limitations were present in the body of evidence, nor formally considered the extent to which their proposed research recommendations had already been met. In the case of US DGAC, the analysis had only been conducted for certain methodological aspects, and for evidence examining UPF intake and obesity only. Thus, more detailed and comprehensive evidence appraisals across the body of UPF evidence to date are limited.

Evaluating Uncertainty in Estimating UPF Intake

When applying the operational definition of UPF (identifying markers of ultra-processing) to diet assessment tools used in cohort studies and trials, measurement error in estimating UPF intake may arise. Generally, all diet assessment tools do not provide or capture all of the information required for exact coding of items into the Nova classification. Food frequency questionnaires (FFQs) are more frequently used in large-scale cohort studies, which offer less granular detail to code into UPF than food diaries or 24-hour recalls. This impacts on the certainty of coding items according to the Nova classification, which can (1) lead to misclassifying food items as UPF or not UPF; (2) generate variation in coder agreement on what is UPF or not UPF; and (3) subsequently require decisions on how to code, such as whether to code each item in the diet assessment tool, disaggregate items and code constituent ingredients, or to weight items based on typical consumption of UPF or non-UPF versions in the population.

Several studies have evaluated the consistency and uncertainty in coding diet assessment tools between raters. One widely cited study suggested that there was poor agreement amongst raters for coding a small subset of food items into the Nova classification. ¹⁰ The availability of an ingredients list should greatly improve confidence in coding given the importance for identifying markers of ultra-processing. However, in this study, access to ingredients lists to identify UPF according to the operational definition did not improve evaluator consistency, nor did it affect evaluator confidence levels. The authors concluded that “evaluators relied on their own knowledge or subjective feelings about the foods when making their assignments” This is evident with products such as ‘Fermented soy pudding with fruit B’ containing thickeners and flavors were being coded as MPF, processed foods (PF), and even as processed culinary ingredients (PCI) by some raters (see Supplemental Table A). The study also focused on a specific subset of 3 groups of items (fresh dairy products, bread products, and mixed dishes), ¹⁰ and thus does not reflect agreement across the overall available food and drink supply. The study does, however, highlight the need for coders to be sufficiently trained for the correct application of the Nova classification.

Where sufficient training has been undertaken, and the whole nutrient database or diet assessment tool has been coded, agreement between trained coders is generally high. For example, in the UK National Diet and Nutrition Survey (NDNS), 2 independent coders reached 97% agreement in coding 4784 items into the Nova classification, and 99.8% agreement on coding UPF. ¹¹ In 3 large-scale cohort studies in the US, 3 independent coders reached 70.2% agreement on coding several hundred items from food frequency questionnaires, reaching 95.6% agreement after further expert and dietitian discussion. ¹² Using 24-hour recall data, 88% agreement in coding was achieved for 3100 items between 6 pairs of independent coders. ¹³ In another study constructing UPF vs non-UPF menus, coders reached greater agreement for classifying foods into the Nova classification than agreement on whether such foods could fit into a diet meeting the US Dietary Guidelines for Americans (coder agreement for DGA-compliant foods was low (W = 0.48, P = 0.678), whereas coder agreement for the Nova classification was high (W = 0.82, P < 0.001)). ¹⁴

Thus, coding of diet assessment tools into Nova is not perfect between raters, but high agreement is achieved with appropriate training, whereby uncertain cases/ambiguous items tend to make up 5-10% of all items.^11,12,15,16 Researchers looking to code diet assessment tools or product lists should refer to published guidance to identify/code UPF with different measurement tools, including best practices for applying the Nova classification, ¹⁷ and approaches used to code FFQs ¹² and 24-hour recalls,^13,15 or identifying UPF in nutrient databases. ¹⁸ These guides provide stepwise processes, considerations of auxiliary data to support decision making, and appropriate sensitivity analyses to reflect uncertainty in coding. In the future, coding approaches may be further aided with machine-learning assistance. ¹⁹

Validation Assessments

A common misconception in observational studies that examine diet-disease relationships is that the diet assessment tool used needs to identify the exact exposure for each participant (ie, measure UPF intake exactly). In fact, this is not the case. To examine diet-disease relationships in nutritional epidemiology, the diet assessment tool instead needs to be able to reasonably rank participants into low, medium or high exposure levels, to enable comparison of disease risk across groups or levels of the exposure. ²⁰ In this regard, a valid diet assessment tool is one that consistently ranks participants, such that low or infrequent consumers are ranked below high or frequent consumers. ²⁰ In other words, a valid diet assessment tool for this purpose does not necessarily need to identify exact true intakes.

In this respect, diet assessment tools used to capture nutrient intakes are not perfect. For example, correlations of energy-adjusted nutrient intake estimates between FFQs and multi-day 24-hr recalls or diet diaries are in the order of 0.45-0.70.^21,22 The reproducibility of nutrient intakes estimated from FFQs over the course of 1 to 10 years is also around 0.50-0.70. ²¹ Likewise, estimating UPF intake is not perfect either. But, validation assessments conducted to date suggest that FFQs are no worse at estimating UPF intake than estimating nutrient intakes. Estimates of UPF intake from generic FFQs (ie, not designed to capture UPF) show moderate correlations and acceptable ranking of participants with multiple recalls or weighted food diaries.^23–25 Results are similar for diet assessment tools designed to capture UPF, with similar correlations with dietary reference methods (0.47-0.72), ²⁶ and moderate-high reproducibility (0.46-0.94). ²⁷ There are also moderate correlations between generic FFQs and Nova-specific FFQs, with 84% of participants ranked in the highest tertile across both tools. ²⁸ This suggests generic FFQs may perform reasonably well in the absence of Nova-specific tools.

Does the coding approach for estimating UPF intake using an FFQ matter, whether coding each FFQ item, or disaggregating items into constituent ingredients? In the NIH-AARP Diet and Health study, the correlation between UPF estimated from the FFQ by each item or by disaggregation were very high (0.97). ²⁹ Furthermore, both methods of coding items or disaggregating items showed moderate to high correlations with UPF intake estimated from two 24-hour recalls (0.43-0.66). In fact, the FFQ performed similarly or better for estimating intake of Nova classification groups than it did for the 26 dietary components in the original validation study. As with nutrient intake estimates, correlations of UPF intake estimates with reference dietary methods tend to increase when adjusting for total energy intake. ²⁹

Overall, current evaluations suggest that estimates of UPF intake from FFQs are reasonable to rank participants across groups compared with reference dietary methods. However, caution should be taken in assuming this translates into their ability to identify exact intakes. As highlighted by UK SACN, ² errors in dietary assessment measurement such as those above tend to obscure, rather than generate diet-disease relationships in observational studies. ² In other words, significant diet-disease associations are observed in spite of dietary assessment measurement errors, rather than because of it. ² This is because measurement errors tend to increase variance in a random manner, rather than to systematically deviate it (unlike unmeasured confounding). Thus, the random non-differential measurement error in estimating UPF intake would tend to bias effect estimates towards the null.^30,31

Evaluating Uncertainty Across UPF-Outcome Associations

The majority of evidence linking UPF with health outcomes arises from observational studies^32,33 which, whilst they provide value in comparison to experimental studies, also carry limitations. We can consider a series of considerations to help assess the uncertainty and confidence in conclusions to date, and whether observational studies on UPF and health can provide a reliable evidence base.

Firstly, earlier meta-analyses pooled all observational studies, including cross-sectional and case control designs, at higher risk of bias. ³² More recent meta-analyses consider prospective cohort studies only ³³ and are consistent with wider meta-analyses of all observational studies. ³² These meta-analyses demonstrate that higher UPF intake is primarily associated with higher risks of obesity, cardiometabolic disease, cancer and all-cause mortality.

Prospective cohort studies are a key backbone of evidence supporting dietary guidelines, given the challenges in conducting long-term, large-scale clinical trials. ² The quality of evidence they may offer can depend on the rating systems used to evaluate them (eg, GRADE vs NutriGrade), which in turn can impact on perceived confidence of results.^34,35 Nevertheless, meta-epidemiological studies find that when populations, intervention/exposures, comparators and outcomes are closely matched, there is high agreement in the findings between nutrition randomized controlled trials and cohort studies. ³⁶ This similarity has been replicated across other nutrition meta-research studies.^37,38

In sensitivity analyses addressing uncertainty in coding diet assessment tools into the Nova classification, effect estimates remain significant when recoding ambiguous items, as seen in risks of weight gain, ³⁹ type 2 diabetes ⁴⁰ in the European Investigation into Cancer and Nutrition (EPIC), and type 2 diabetes ⁴¹ in the US Nurses’ Health Study (NHS) I and II and Health Professional Follow-Up Study (HPFS). Thus, uncertainty in coding does not significantly affect estimates. However, few studies to date have conducted such analyses.

Different diet assessment tools may introduce different levels of measurement error. Effect estimates from meta-analyses of cohort studies using FFQs or 24-hour recalls are directionally consistent in demonstrating significant associations between higher UPF intake and higher risk of adverse health outcomes, but the effect estimates differ in size for type 2 diabetes and obesity. ⁴²

Appropriate adjustment for confounding variables is necessary to obtain unbiased effect estimates. In recent meta-analyses of prospective cohort studies (n = 104), ³³ all studies controlled for sociodemographic factors, and most control for lifestyle factors (smoking and physical activity (n = 96), baseline body mass index (n = 79), alcohol intake (n = 57)). However, there is inconsistency between studies, such as the range of sociodemographic variables used (eg, age, gender, ethnicity, income, education level). Importantly though, many of the cohort studies examining associations between UPF and health are the same cohorts used to inform dietary guidelines and recommendations for nutrients and foods ² (eg, NHS I and II, HPFS, EPIC, UK Biobank, UK NDNS). As such, wider nutritional epidemiological studies with these cohorts contain the same range of variables and same ability to adjust for confounding, and therefore are similarly affected by unmeasured confounding. Nonetheless, directed acyclic graphs ⁴³ can be used to identify appropriate confounder adjustment sets for the effect estimand of interest, and E-values ⁴⁴ used to assess the potential impact of unmeasured confounders on effect estimates, as used in studies of UPF and type 2 diabetes. ⁴⁰ Additionally, negative control analyses can be used to assess the extent of residual confounding. For example, UPF intake was significantly associated with accidental death (as a negative control when assessing cancer outcomes) in the EPIC cohort. ⁴⁵ This would reflect some level of residual confounding. The effect estimate for accidental death was half that for overall cancer risk, thus residual bias, whilst likely present, would broadly explain away half of association.

Given the overlap between high UPF intakes and poor diet quality as already captured in dietary guidance, ⁸ it is plausible that the associations are driven by nutrients or food groups. However, evidence to date suggests UPF-disease associations are not explained by diet or nutrient quality.^27,46 Some studies include these dietary adjustments in their fully adjusted models, and some do not, instead performing dietary adjustments as secondary or sensitivity analyses. Particularly when adjusting for energy intake or other dietary components, dietary adjustments can introduce inadvertent unspecified or specified substitution models, and variation in dietary adjustments in fully adjusted models can generate heterogeneity across studies in meta-analyses. ⁴⁷

Furthermore, whilst most UPF tend to be nutrient poor, ⁴⁸ the nutrient profile of UPF is heterogeneous with some being high in nutrient quality. ⁴⁸ Thus, subgroups of UPF may be expected to have varying associations with health outcomes, of which to date this appears evident.^40,49 The interpretation of these results however has been contested, and debates continue regarding the relative importance of UPF subgroups analyses with health outcomes. ⁵⁰ From a measurement error approach, current diet assessment tools in epidemiology are limited in capturing intakes of UPF and similar non-UPF comparison foods (ie, in order to conduct like-for-like specified food substitution models), and currently, UPF subgroup analyses tend to use an unspecified substitution approach (ie, the association of consuming more of that UPF subgroup and less of the average weighted diet). ⁵⁰ One must also consider that there are likely to be multiple ‘UPF dietary patterns’; high consumers of plant-based UPF meat/milk alternatives are unlikely to be the same individuals who are high consumers of animal-based UPF.

Lastly, whilst plausible mechanisms exist between higher UPF intake and a number of health outcomes, ³³ spurious associations may exist for others, such as accidental death. Thus, consideration of biological plausibility is needed.

Trials

Briefly, a range of free-living and lab-based ad libitum trials suggest that UPF diets lead to greater energy intake and less favorable weight change compared with matched non-UPF diets for diet/nutrient quality,^51–53 in alignment with evidence from PCS. Potential mechanisms linking UPF with health outcomes include changes in eating behavior/physiology, and changes in energy/nutrient intake, with product characteristics including energy density, texture, hyper-palatability and other sensory characteristics viewed as confounders or mediators by proponents and opponents. Studies need to consider how such sensory properties and factors vary across matched UPF, non-UPF, such as evidence of higher energy density in UK UPF vs MPF matched foods.^48,54

Research Recommendations and Conclusions

The confidence in conclusions and recommendations must align with the certainty in the evidence, by considering biases and error at each stage of research.

Not all UPF are equal, with heterogeneity in associated outcomes. Evidence would suggest an ‘effect’ of ultra-processing across matched diet patterns, but there is uncertainty in the effect size. Any assessment of health impacts must consider the baseline diet or food substitution in question. Notably, most cohort studies to date use diet assessment tools that have not undergone validation assessment. ²⁷ However, those that have show moderate-high consistency with reference methods equal to or better than nutrient intake estimates, and such tools are similar to those more widely used in other cohort studies. ²⁹ There is a strong need for validated diet assessment tools or validation assessments of existing diet assessment tools to estimate intake based on the Nova classification, and sensitivity analyses are key.

Whilst health associations are not fully explained by existing dietary guidance or nutrient intake, the Nova classification is also not a panacea to explain all diet-disease relationships. Food-based dietary guidance combined with consideration of the Nova classification may provide most informative value.

This paper considered the health impacts of UPF, and does not address or consider the wider sociopolitical, economic or environment aspects related to UPF as addressed in other studies. ⁵⁵

Ongoing work will formally establish the level of uncertainty and confidence in prospective cohort studies to date, ⁵⁶ evaluating UK SACN and US DGAC limitations and research recommendations on UPF and health. This work will evaluate the quality and standard of prospective cohort studies identified in the systematic review by Monteiro et al ³³ plus additional studies since the systematic review up until March 2026, mapping each PCS to UK SACN and US DGAC suggested limitations and research recommendations.

Supplemental Material