The zero-calorie paradox

The conceptual premise of the NOVA classification is that health risks may arise from industrial processing – via matrix degradation and additive exposure – rather than nutrient composition alone.^{(Reference Monteiro, Levy and Claro1)} However, the predominant metric used in nutritional epidemiology, %TE, implicitly assumes that the health relevance of a food item scales with its caloric contribution.

This assumption may be poorly suited to a specific class of UPFs: non-caloric or low-energy-dense items, most notably artificially sweetened beverages (ASBs). As illustrated in Figure 2, an individual consuming 500 ml of diet soda receives a dose of additives (sweeteners, acidifiers, colourants) and consumes a non-trivial mass of ultra-processed product. Yet, because the caloric content is near zero, this exposure may be substantially under-represented in energy-based models.

Figure 1.

The evolution of the UPF evidence base (2009–2024). conceptual illustration. the simplified timeline traces the initial development of the NOVA framework, the accumulation of broadly consistent adverse associations in early prospective cohorts, and the later emphasis on subgroup, metric, and contextual heterogeneity.^{(1–Reference Cordova, Viallon and Fontvieille10)} The phases are schematic and overlapping; they do not represent a systematic chronology, pooled estimates, or definitive changes in scientific consensus.

Figure 2.

The proposed metric trap: how exposure metrics may shape UPF risk assessment. Conceptual illustration. The figure depicts the proposed ‘Zero-calorie paradox’ using a 500 ml diet soda as an example. When measured by energy contribution (% energy, left bar), the contribution of this item is negligible because of its near-zero energy content. When measured by weight (% weight, right bar), the mass of the beverage produces a larger exposure signal. The ‘approximately 100%’ label refers only to classification of the beverage itself as ultra-processed, not to its percentage contribution to the individual’s total diet. This framework is hypothesis-generating; values shown are illustrative rather than empirical estimates.

This concern is not purely theoretical: two cohorts that adopted weight-based exposure have stated the rationale explicitly. The NutriNet-Santé mortality analysis used percentage of weight in its primary models in order to capture low- and zero-energy ultra-processed items,^{(Reference Schnabel, Kesse-Guyot and Allès7)} and the SCCS reported defining intake by weight rather than energy proportion ‘considering no energy contribution from some UPF (e.g. artificially sweetened beverages)’, while acknowledging that this choice in turn ignores differences in energy density between ultra-processed items.^{(Reference Wang, Pan and Munro8)} The consequence of the metric is visible within the SCCS itself: under a weight-based denominator, beverages were the largest single ultra-processed contributor to the diet (mean 19.4% of intake), approximately three times the next largest category, and each 10% increment in beverage-derived ultra-processed intake was associated with a 10% higher diabetes mortality (HR 1.10, 95% CI 1.05–1.15).^{(Reference Wang, Pan and Munro8)} One boundary should be stated explicitly: the beverages driving this signal were predominantly sugar-sweetened rather than artificially sweetened. These data therefore support the general proposition that weight-based metrics amplify beverage exposure; they do not by themselves demonstrate the zero-calorie case, for which the argument at present rests on the metric rationale stated by the investigators rather than on a direct empirical comparison.

Divergent outcomes in major cohorts

This methodological divide may account for part of the conflicting evidence observed in Table 1:

• Group A (French cohort studies): Studies from the NutriNet-Santé cohort (e.g. Srour et al.^{(Reference Srour, Fezeu and Kesse-Guyot2)}; Fiolet et al.^{(Reference Fiolet, Srour and Sellem3)} consistently used % Weight metrics. Consequently, these studies captured a broad spectrum of UPF consumption, including ASBs, and reported robust associations with CVD and cancer. In the NutriNet-Santé analyses, sugary and ASBs contributed to the observed CVD association – a contribution more readily detected when the weight-based metric registers these largely non-caloric items.

• Group C (UK energy-based studies): Conversely, studies relying strictly on % Energy reduce the contribution of non-caloric ultra-processed items to the exposure variable: a high intake of ASBs does not substantially increase a participant’s calculated UPF score. This does not, however, imply that energy-based metrics are generally insensitive. In the UK National Diet and Nutrition Survey, Rauber et al.^{(Reference Rauber, Steele and Louzada9)} used percentage of total energy and nevertheless observed strong, dose-responsive associations with adiposity in both sexes: each 10% increment in ultra-processed intake was associated with a 0.38 kg/m² higher BMI, a 0.87 cm larger waist circumference and 18% higher odds of obesity, rising to 1.66 kg/m², 3.56 cm and 90% respectively in the highest versus lowest quartile.^{(Reference Rauber, Steele and Louzada9)} Energy-based metrics therefore appear well suited to outcomes driven by energy-dense UPFs. The proposed Metric Trap is correspondingly narrow in scope: it concerns the specific subset of low- and zero-energy items whose contribution an energy denominator cannot register, and not the sensitivity of energy-based exposure in general.

A further constraint should be stated plainly. Where investigators have compared metrics directly within the same population, the estimates have not always diverged: Fang et al.^{(Reference Fang, Rossato and Hang4)} measured intake as servings per day with energy adjustment by the residual method and reported similar results in a sensitivity analysis using percentage of energy.^{(Reference Fang, Rossato and Hang4)} The proposed Metric Trap therefore predicts divergence only where low- and zero-energy items constitute a material share of intake, and it remains to be tested directly.

We therefore suggest that relying solely on energy metrics may create the pattern described by the proposed ‘Metric Trap’, in which the potential health effects of non-caloric ultra-processed items could be under-represented owing to their lack of caloric density. We emphasise that this remains a hypothesis-generating interpretation: it would need to be tested directly – for example, through comparative modelling of the same cohort under both metrics, sensitivity analyses, or meta-regression – before it can be regarded as demonstrated. To accurately assess the health impacts of the full UPF spectrum, we suggest that future research report weight-based estimates alongside traditional energy metrics.

The bread paradox: context as effect modifier

A frequently discussed divergence in the literature concerns ultra-processed breads and cereals. In European cohorts, such as the EPIC study^{(Reference Cordova, Viallon and Fontvieille10)} or the French NutriNet-Santé, industrial breads are often categorised as neutral or only mildly associated with risk. In the US-based Nurses’ Health Study and HPFS analysis, ultra-processed whole grains were not associated with higher mortality and were in fact removed from the ultra-processed exposure a priori, on the basis of the well-established benefits of whole grains, rather than being shown to be protective.^{(Reference Fang, Rossato and Hang4)} Ultra-processed breakfast foods excluding whole grains, by contrast, remained modestly associated with all-cause mortality (HR 1.04).

We hypothesise that this discrepancy may, in part, reflect Nutritional Displacement:

• In the US Context (Lower Baseline Diet Quality): Given the generally lower quality of the typical US diet, an individual choosing ultra-processed whole-grain bread may be substituting it for refined grains, sugary snacks, or fast food. Despite the industrial processing, the retention of fibre and micronutrients may confer a relative advantage.

• In the European Context (Higher Baseline Diet Quality): In Mediterranean or traditional European dietary patterns, the consumption of packaged industrial bread may displace fresh, artisanal, or minimally processed alternatives (e.g. sourdough, pasta). Here, the shift towards UPF may represent a relative degradation in diet quality.

On this interpretation, the apparent health coefficient of a UPF item may not be intrinsic solely to its formulation, but may also depend on the baseline diet it displaces. We note that this displacement hypothesis is, at present, an inference from cross-cohort comparison rather than a directly tested effect modifier.

Subgroup composition and signal dilution

The second layer of the paradox is compositional. The broad label of ‘Ultra-Processed Food’ groups together disparate food matrices – placing, for example, a pot of fortified fruit yogurt in the same category as a processed-meat sausage.

As visualised in Figure 3, aggregating these distinct subgroups into a single metric may dilute specific risk signals. Fang et al.^{(Reference Fang, Rossato and Hang4)} reported that the modest overall association (HR ≈ 1.04) was concentrated in a few subgroups – principally processed meats and sweetened beverages – while many other ultra-processed categories were not individually associated with mortality, a pattern consistent with dilution of subgroup-specific signals within the aggregate metric.

Figure 3.

Subgroup-specific associations between ultra-processed food and health outcomes in two major cohorts. Hazard ratios (95% CI) are taken from Fang et al.^{(Reference Fang, Rossato and Hang4)} (nurses’ health study/HPFS; all-cause mortality, highest vs. lowest quartile) and Cordova et al.^{(Reference Cordova, Viallon and Fontvieille10)} (EPIC; cancer–cardiometabolic multimorbidity, per 1 SD increment).^{(Reference Fang, Rossato and Hang4,Reference Cordova, Viallon and Fontvieille10)} Because the studies use different outcomes and exposure metrics, absolute values are not directly comparable across panels; the figure displays within-study heterogeneity across subgroups. Red, elevated risk; blue, lower risk; grey, neutral or total exposure. EPIC, European prospective investigation into cancer and nutrition; HPFS, health professionals follow-up study.

This pattern implies that the robust associations observed in earlier ‘Total UPF’ studies may, in part, have been driven by specific high-harm subgroups rather than by the ultra-processing phenotype uniformly. By treating UPF as a single block, analyses risk both masking specific pathogenic drivers (e.g. heme iron, liquid sugars, sodium) and potentially over-penalising nutrient-dense, accessible industrial foods such as fortified dairy or whole-grain cereals. We reiterate, however, that this argument concerns the specificity of risk attribution, not the existence of an overall association.

Heterogeneity in the context of the broader evidence base

The preceding observations should not be read as evidence that UPFs are harmless, or that their documented risks are largely artefacts of classification systems or dilution effects. On the contrary, large umbrella reviews and meta-analyses continue to report broadly consistent associations between higher UPF intake and cardiometabolic disease, obesity, type 2 diabetes, several cancers, common mental disorders, and all-cause mortality.^{(Reference Lane, Gamage and Du5,Reference Pagliai, Dinu and Madarena6)} Subgroup heterogeneity and population-level risk are not mutually exclusive: a robust average association can coexist with meaningful variation in the magnitude – and occasionally the direction – of effects across specific food categories and dietary contexts.

Several further considerations warrant emphasis for balance. First, much of the evidence is observational, so residual confounding, the clustering of ultra-processed intake within broader unhealthy dietary patterns, and correlated lifestyle factors (e.g. smoking, physical inactivity, socioeconomic position) remain plausible partial explanations for the observed associations. Second, the heterogeneity we describe pertains chiefly to the distribution and attribution of risk across subgroups and metrics, not to the presence or absence of an overall signal. Our argument is therefore not that the UPF–health association is spurious, but that its precise interpretation is sensitive to how exposure is operationalised and how subgroups are aggregated.

Physical structure and matrix degradation

A pervasive feature of ultra-processing is the alteration of the original food matrix. Ultra-processed formulations are characteristically hyper-palatable, and diets based on them are energy-dense.^{(Reference Rauber, Steele and Louzada9)} This modification of the matrix has been linked to metabolic consequences, primarily through the proposed mechanism of energy intake rate.

Notably, in the only randomised controlled trial to date directly testing the UPF hypothesis, Hall et al.^{(Reference Hall, Ayuketah and Brychta11)} found that participants on an ultra-processed diet consumed 500 kcal/day more than those on an unprocessed diet, despite diets being matched for presented macronutrients.^{(Reference Hall, Ayuketah and Brychta11)} This was a small, tightly controlled inpatient trial (20 participants, two weeks per arm), so its internal validity is high but its generalisability to free-living populations remains to be established. Two mechanisms have since been proposed:

• Bypassing Satiety Signals: The soft textures characteristic of many UPFs (e.g. industrial breads, snacks) require less mastication, which may allow ingestion rates that outpace physiological satiety signalling (e.g. cholecystokinin [CCK] and peptide YY [PYY]).

• Acellular Nutrients: The disruption of the cellular fibre matrix yields ‘acellular’ nutrients. It has been suggested that this physical state may facilitate rapid absorption and postprandial glycaemic excursions; this remains a mechanistic hypothesis rather than an established pathway.

Additive exposure and neo-formed contaminants

Beyond physical structure, the chemical dimension of ultra-processing introduces exposure to additives and neo-formed compounds. While regulatory safety data for individual additives exist, some epidemiological signals raise questions about chronic, cumulative exposure.

Two strands of evidence bear on this question: cohort analyses that attempt to disentangle the effects of specific additives from the food matrix, and experimental work on emulsifiers.

• Nitrates and Nitrites: In the NutriNet-Santé cohort, food-additive nitrates and nitrites were associated with elevated risks of breast and prostate cancer respectively, whereas nitrates and nitrites from natural sources showed no consistent association.^{(Reference Chazelas, Pierre and Druesne-Pecollo12)} This is consistent with – though does not on its own establish – mechanistic hypotheses regarding the formation of carcinogenic nitrosamines.

• Emulsifiers: Additives such as carboxymethylcellulose (CMC) and polysorbate-80 alter the mucosal barrier and gut microbiota composition in preclinical murine models and in vitro systems.^{(Reference Chassaing, Koren and Goodrich13)} Human evidence has since begun to accumulate but remains limited in scope: a randomised controlled-feeding trial reported that CMC altered the gut microbiota and faecal metabolome of healthy adults, although over a short exposure period and in a small sample,^{(Reference Chassaing, Compher and Bonhomme14)} and prospective cohort analyses have reported associations between food-additive emulsifier intake and CVD risk.^{(Reference Sellem, Srour and Javaux15)} These findings are best regarded as biologically plausible and preliminary rather than as established determinants of chronic disease in humans.

The gut–brain axis and systemic inflammation

Some research suggests that the impact of UPFs may extend systemically via the gut–brain axis. Lane et al. (2022) reported a consistent association between high UPF consumption and adverse mental health outcomes, including depression and anxiety, in a meta-analysis of observational studies.^{(Reference Lane, Gamage and Du16)}

A frequently proposed mechanism involves systemic low-grade inflammation. It has been postulated that microbiome dysbiosis – potentially driven by low fermentable-fibre intake and the presence of non-nutritive components – may increase intestinal permeability, facilitating translocation of bacterial lipopolysaccharides (LPS) and triggering systemic inflammatory responses that may influence the central nervous system, accompanied by increases in C-reactive protein (CRP) and interleukin-6 (IL-6). We stress that this pathway is, at present, largely hypothetical in humans: it is assembled from separate strands of preclinical and associational evidence, and direct human studies using biomarkers of gut permeability and inflammation are required before it can be regarded as established.

Scenario 1: the strict (toxicological) model

Scenario 2: the pragmatic (nutritional) model

Scenario 3: the volumetric (additive) model

A worked example

To make the typology concrete, consider how three illustrative items would be treated under each model (Table 2). The same food can legitimately receive different analytic treatment depending on the research question. Ultra-processed whole-grain bread, for example, would be counted as an exposure in the Toxicological model (it is an industrial formulation), might be reclassified out of the exposure in the Nutritional model (its fibre and micronutrient density are relevant to cardiometabolic endpoints), and would be included by mass in the Volumetric model. Consequently, a study reporting ‘no association between UPF and type 2 diabetes’ under a Nutritional model and a study reporting ‘a positive association’ under a Toxicological model need not be contradictory: they may be answering different questions about the same foods. Making the chosen scenario explicit would therefore improve comparability across studies and reduce apparent conflict.

Feasibility and reproducibility

Operationalising this typology is feasible with existing data infrastructure but carries practical caveats. Scenarios 1 and 2 can be implemented in most cohorts that already classify foods by NOVA group, requiring only a transparent, preregistered list of any items reclassified under Scenario 2. Scenario 3 is more demanding: weight-based metrics are computable in cohorts with quantified intakes, but additive-level analysis (as in Chazelas et al.) requires linkage to food-composition and additive databases that are not yet standardised across countries, which currently limits cross-cohort reproducibility. We therefore recommend that studies (i) state explicitly which scenario is being tested, (ii) report results under at least one alternative scenario as a sensitivity analysis, and (iii) publish the item-level classification used, so that differences between studies can be attributed to data rather than to undocumented definitional choices. We present this typology as a conceptual proposal intended to structure future work; its value will ultimately depend on empirical testing across diverse cohorts.

Exposure misclassification and measurement error

A further, often underappreciated, source of heterogeneity in the UPF literature relates to exposure misclassification and measurement error arising from dietary assessment methods. Most large-scale cohorts rely on either FFQs or repeated 24-h dietary records/recalls, each of which captures different dimensions of UPF consumption. FFQs are optimised for ranking habitual intake of broad food groups but may insufficiently capture formulation-specific details, intermittent consumption, or low-energy items such as ASBs and additive-rich foods. In contrast, repeated 24-h records provide greater granularity but remain vulnerable to day-to-day variability and reporting bias.

Importantly, non-differential misclassification of UPF exposure is generally expected to bias effect estimates towards the null, particularly in energy-based metrics and in cohorts where UPF intake is widely distributed across diverse subcategories. Misclassification at the subgroup level may further contribute to a dilution effect, whereby harmful UPFs (e.g. processed meats or sugar-sweetened beverages) and neutral or potentially protective UPFs (e.g. whole-grain breads or fermented dairy products) are aggregated into a single exposure variable. Such measurement limitations may obscure subgroup-specific associations and contribute to apparent inconsistency across studies. This pattern may partly reflect structural features of UPF ascertainment and classification, although random variation and alternative interpretations remain possible.

Strengths and limitations of this review

This review is intentionally narrative and hypothesis-generating rather than systematic, and therefore does not include a preregistered protocol, formal study selection flow diagram, or quantitative risk-of-bias appraisal. Its primary objective is not to estimate pooled effect sizes, but to synthesise methodological and contextual explanations for heterogeneity across major UPF cohorts. As outlined in the Literature Selection Strategy, study selection was purposive, which improves thematic focus but introduces a risk of selection bias; readers should therefore weigh our interpretations against comprehensive systematic syntheses and umbrella reviews.^{(Reference Lane, Gamage and Du5,Reference Pagliai, Dinu and Madarena6)}

Accordingly, the conceptual frameworks proposed here – the ‘Metric Trap’, the ‘Geographic Paradox’, and the Risk-Stratified Typology – are presented as plausible explanatory models rather than quantitatively validated mechanisms. Alternative interpretations of the available evidence remain possible, including statistical variation, residual confounding, and differences in exposure ascertainment or population composition. The proposed frameworks therefore represent one plausible interpretation intended to stimulate discussion and future empirical validation. Formal meta-regression, head-to-head comparison of metrics within the same cohorts, and preregistered sensitivity analyses would be required to test these hypotheses empirically and represent important directions for future work.

Mechanistic pathways discussed herein – such as additive exposure or food-matrix disruption – should be interpreted as biologically plausible links supported by experimental and observational literature, rather than as causal attributions within the epidemiological associations reviewed. Finally, exposure misclassification and measurement error inherent to FFQ- and 24-h-based UPF ascertainment may contribute to attenuation or dilution of associations, particularly for heterogeneous food subgroups.