Systematic Review and Comparison of Classification Frameworks Describing the Degree of Food Processing

Background: Many available foods have been processed such that they differ from their original harvested state. There are several different classification frameworks used to distinguish between the levels of processing applied to a food but the comparability of the frameworks is unknown. Methods: We sought to identify all relevant food classification frameworks and determine the level of agreement between them. This was done by applying each framework to the 135 food categories described in the Euromonitor 2011 ‘Who Eats What’ report and comparing how each framework defined each food category. Results: We searched PubMed, ScienceDirect and Google Scholar databases, reviewed 185 abstracts and 25 full text reports. There were 6 eligible classification frameworks identified – FSANZ, IARC, IFPRI, NOVA, IFIC and an unnamed framework described by Poti. FSANZ simply dichotomised foods into unprocessed or processed while the remainder divided foods into three (IFPRI), four (IARC-EPIC), five (IFIC and NOVA) or seven (Poti) categories based upon the degree of processing. For the primary analysis of determining the agreement of unprocessed versus processed across classification frameworks, there were 810 comparator points between the 6 frameworks, for which 651matched (81%). The highest level of agreement was between IFPRI and IARC (134/135, 99%) and the lowest level of agreement was between FSANZ and NOVA (92/135, 68%).The NOVA framework stood out as having between 68 and 81% agreement with the other 5 frameworks while agreement was between 89 and 99% if NOVA was excluded. Interpretation: Most frameworks made similar basic classifications of food categories as processed or unprocessed. Systems with more levels of classification may prove more useful for tracking changes in the food supply and in defining associations of food processing with health concerns.


Introduction
Most foods now available have been processed such that they differ from their original harvested state [1][2][3][4][5].Processing was initially developed as method for reducing spoilage and increasing food safety but is now often done to deliver time-saving, convenient and low cost products to consumers.Food processing varies from basic techniques such as salting and drying to complex manufacturing methods that completely transform the foods' characteristics [1,[6][7][8].nutrition and food science, and Master's Degrees in public health (MC and TB) with disagreements resolved by consultation.

Included classification frameworks
Classification frameworks were eligible for inclusion if they were: (i) generalizable and could be applied to all 135 Euromonitor food categories; (ii) able to be applied to foods available at the point of purchase (this excluded frameworks designed specifically for use in dietary assessment settings such as dietary recalls, household food surveys, and weight food records); (iii) specific to detailing industrial food processing; (iv) able to allocate foods into two or more different groups depending upon the level of processing applied; and (v) could be applied to a broad range of different food types.

Food categories used for making the assessment
We used the 135 food categories defined in the Euromonitor 2011 'Who Eats What Report' (Appendix 1) to compare the identified frameworks.These 135 food categories cover the main sources of packaged and unpackaged foods sold across 80 countries [22].For the purpose of assessing the level of processing, unless otherwise specified, food groups were considered plain or unflavored, canned fruit was assumed to be in natural juices and not syrup, canned vegetables were assumed to be in water not brine, canned meat and fish were assumed to be in oil not sauce or marinade, and nuts were assumed to be unsalted.

Assignment of food categories within each classification framework
Each of the 135 food categories was assigned to the appropriate level of processing in each of the different frameworks.This was done independently by two authors (MC and TB).Discrepancies between the two were recorded and a final assignment was agreed through consultation.In addition to the assignment of each food category to the most appropriate level of processing within the framework, the processing levels within each framework were also grouped to provide a dichotomisation of categories into unprocessed or processed.Once again this assignment was done independently by two authors (MC and TB) with disagreements recorded and a final assignment agreed by discussion.

Analysis
The primary analysis was a comparison of the proportion of products categorised as processed versus unprocessed for each of the 810 (6 frameworks x 135 food categories) food category pairs that could be compared between the 6 classification frameworks.The agreement between a pair of frameworks was estimated by counting the number of food categories for which the assignment of processed or unprocessed was congruent, dividing that number by 135 (the total number of food categories) and multiplying by 100 to obtain a percentage.The number of times each food category was differently assigned as processed or unprocessed across the six frameworks was also recorded.In addition, to understand the ease of applying the frameworks, we tabulated the number of discrepancies between assignments made by the two authors for the allocation of food categories within each framework and also recorded initial inconsistencies in the allocation of processing levels within each of the non-dichotomous frameworks as unprocessed or processed.

Characteristics of the frameworks
Of the 6 frameworks, FSANZ simply dichotomised foods into unprocessed or processed [23], IFPRI divided foods into three levels of processing [13], IARC-EPIC into four levels of processing [16], IFIC and NOVA into five levels of processing [12,25] and Poti into seven levels of processing [15] (Figure 2).Although the NOVA classification has 3 major classifications [25], with the third group split in to (a) processed ready-toeat and (b) ultra-processed, creating four classifications, the framework was considered to have 5 levels of processing as it also separately defined unprocessed and minimally processed [18].For the IARC-EPIC framework we excluded the 'moderately processed at home' categorisation option (because it described processing after purchase) and placed each food category into one of the other 4 levels.
All frameworks defined the lowest level of processing as 'unprocessed' or 'non-processed' except for IFIC which used the term, 'minimally processed' .The NOVA framework specified a category of 'ultra-processed foods' above 'highly processed foods' which was the highest level of processing defined for all other frameworks except IFIC, where the descriptor of the highest level of processing was 'prepared foods/meals' .
Division of the levels of the five frameworks that were not already a dichotomy into unprocessed or processed resulted in processing levels described as 'minimal or no processing' , 'unprocessed' , 'not processed' , 'minimally processed' , 'processed basic ingredients' and 'processed for preservation' being assigned as unprocessed.The descriptors of the levels assigned as processed were 'foods that have been substantially changed from their original form' , 'ready-to-eat processed' , 'mixtures of combined ingredients' , 'processed staple/basic' , 'prepared foods/meals' , ' moderately processed' , 'highly processed' , and 'ultra-processed' .There were 24 different descriptors of processing across the 5 non-dichotomous frameworks that had to be assigned as unprocessed or processed by the authors.The independent assignments made by the authors (MC and TB) were initially the same for 18/24 with 6 requiring discussion to make a final assignment as unprocessed or processed.

Assignment of foods categories within the frameworks
Two authors (MC and TB) independently assigned each food category to a processing level within each framework for a total of (6 × 135=810) allocations each.There were 43/810 (5.3%) assignments that were discrepant between the two authors that had to be resolved by consensus.There were no disagreements for the Poti framework, 2/135 (1.5%) for the IFIC framework, 11/135 (8.1%) for the NOVA framework, 9/135 (6.7%) for the IARC-EPIC and IFPRI frameworks and 12/135 (8.9%) for the FSANZ framework.Examples of food categories that were more difficult to assign and were initially assigned discrepantly across several frameworks were prepared salads, honey and canned/preserved sea food (Appendix 1).
The proportion of food categories assigned as processed ranged between 46% and 67% for the six frameworks (Figure 2).The NOVA framework stood out as assigning less food categories as processed at 46% compared to the other five frameworks where the proportion ranged from 57% to 67% (Figure 2).There was between 87% and 99% agreement in the assignment of food categories as unprocessed or processed within the 5 frameworks most coherent with each other but that figure fell to between 68% and 81% when the NOVA framework was compared against those five categories (Table 2).Almost two thirds of food categories were consistently assigned to either unprocessed or processed across all 6 frameworks.The food categories that varied most in their assignment to unprocessed or processed across frameworks were chilled/fresh pasta, cooking fats, dried pasta, other dairy and plain noodles which were each assigned as unprocessed by three and processed by three.

Discussion
The six identified frameworks took different approaches to the classification of foods but ultimately defined major categories of products as unprocessed or processed in fairly similar proportions and a fairly similar way.Accordingly, while research findings based upon one framework compared to another are likely to vary, it is unlikely that the variation would be large.If one categorisation system was an outlier it was the NOVA framework which was more discrepant from the others in terms of the overall proportion of foods classified as processed and consequently also had lower levels of agreement with the other frameworks across the full set of 135 food categories evaluated.Whether the NOVA framework is a substantially better or worse framework than the others will only be determined by directly comparing research findings derived using the NOVA framework and one or more of the alternatives.It is noted that, to-date, the NOVA classification framework is the most widely used Processed for basic preservation or precooking 15 11 Moderately processed for flavor 6 4 Processed 77 57 Moderately processed grain products 1 1 Highly processed ingredients 18 13 Highly processed stand-alone 52 39 classification framework of those that we assessed, and has been used on numerous occasions in nutrition research to assess the healthiness of diets and food environments where food composition data has not been readily available [7,14,21] The relative utility of these frameworks will depend somewhat upon the research question being asked and the setting in which the research is being done.The capacity for different users to consistently interpret and apply a framework will, however, be important in most settings because systematic or random errors in the use of a framework will adversely impact upon the ability to address the research question under investigation [15].In this regard the Poti framework performed best with independent assignments by the two reviewers identical for all 135 food categories.By contrast, almost one in ten of the assignments made for the FSANZ and NOVA frameworks were discrepant.In practice, when it is individual food products, rather than broad food categories that need to be assigned a level of processing the errors are likely to be greater because the foods within some categories can be quite varied.
Prior work has identified no clear association between the number of different levels of processing described by a framework and the ease of assigning foods to processing categories [15,21].The same was true in this study.While the FSANZ framework with just two levels of processing proved the hardest for which to make consistent assignments and the Poti framework with seven levels the easiest, the NOVA framework with five described levels of processing had the second highest rate of disagreement between the assignments made.The rather vague definition of processed foods used by FSANZ -"substantial change to the original state of the food" [23] -proved to be open to significant interpretation while the more technical definitions used by the NOVA framework were difficult to apply to some food categories [25].More detailed data such as individual product data (product description and ingredients), a more detailed food categorisation system and familiarity with industrial food processes may aid with the use of all systems, but more so the NOVA system [21,25].Alternatively, the use of less technical definitions and the provision of more food product examples, as has been done for the Poti framework which is a modification of the NOVA framework [15], can make application of the system much more straightforward (Appendix 2).
An advantage of frameworks that describe multiple levels of processing is that they should provide for a better description of the nature of the food supply and the nutritive value of foods within it [15,21,25].For example, in many research settings, it will be important to distinguish between canned fruits that are in syrup versus natural juices, something that cannot be achieved by a dichotomous framework such as FSANZ but that would be distinguished by all the other frameworks that comprise three or more different levels.The use of a framework that provides a multilevel description of the degree of processing of foods will be important for research seeking to demonstrate associations between food processing and health outcomes because a more detailed description of the exposure of individuals or populations to different types of food should enhance the capacity to detect associations.It will also make it possible to explore possible dose-response associations between not just the portions of food that is processed but also the extent of processing that food has undergone.
Ultimately classification frameworks such as these are an effort to make a high level evaluation of a food product or food category that infers something about the nutritive value of the food.This is important because several studies have suggested links between processed foods, obesity Open Access

5
and NCDs that may be particularly pronounced for more highly or ultraprocessed products [1,2,[10][11][12][13][14][15][16][17][18][19]21,26,27].An objective and easy to apply framework that makes robust assignments of foods on the basis of their degree of processing therefore has significant potential to better define the way in which the food supply impacts upon health outcomes.
A key strength of this research is that we sought to systematically identify and summaries all applicable frameworks.Further, we moved beyond a prior qualitative assessment of selected frameworks [21] to provide a quantitative evaluation of the various systems and the ease with which they can be applied.A more robust evaluation of the comparative ease with which each framework can be used could be achieved by having more people repeat the assignments made by the authors of this paper.A further evaluation of the frameworks could be made by comparing the levels of 'positive' and risk-associated nutrients in the food categories assigned as unprocessed or processed by each framework and its association with nutrient profiling and association with NCDs, but was beyond the scope of this research.While a detailed nutrient-based evaluation of foods is likely to be preferred in most research scenarios, it can be hard to obtain those data and simple classification frameworks such as those evaluated here may offer a valid alternative.A final limitation to the evaluation made here is that the Euromonitor 'Who Eats What' 2011 report food categories as obtained employed for the main comparisons includes only dairy-based beverages and no other beverage categories [22].‡ Food groups used for assessment were the 135 food groups defined in the Euromonitor 2011 'Who Eats What Report' which covers the main sources of packaged and unpackaged foods across 80 countries [22].Lightest grey is least processed circling clockwise to black which is most processed with correspondingly ordered bulleted list descriptors.

Conclusion
There are a number of research questions for which the application of a framework that classifies foods according to their level of processing are likely to be of value.The available frameworks were mostly easy to apply and resulted in broadly comparable findings.The primary exception was the NOVA framework but whether this was performing better or worse is unclear.This means that future research that relies upon the classification of foods as unprocessed or processed should ideally test their hypotheses using at least two frameworks, one of which should be NOVA.A piece of research that applied nutrient profiling to a large standard set of foods and compared the average nutrient profile of processed versus unprocessed foods between frameworks would be of value.Application and interpretation of classification systems were found to be highly variable.If utilised properly, with a minimum 2-stage reviewer process, classification frameworks can provide useful insight to the types of foodstuffs available in a food environment of interest and/or the proportion of an individual's or nation's diet that is comprised of processed foodstuffs.Classification frameworks with more levels of food processing provided more insight to the (non-nutritional) composition of food environments than the dichotomous frameworks, which were observed to be limited in their use.It is recommended that classification frameworks are used in conjunction with a quantitative-based system, such as nutrient profiling, to robustly assess the healthiness of food environments.

Processed ready-to-consume products
Manufactured by adding substances like oil, sugar or salt to whole foods to make them durable and more palatable and attractive.Directly derived from foods and recognisable as versions of the original foods.Generally produced to be consumed as part of meals or dishes, or may be used, together with ultra-processed products to replace foodbased freshly prepared and meals.Processes include canning and using oils, sugars, syrups or salt, and smoking , saltpickling Canned or bottled vegetables and legumes preserved in brine, peeled or sliced fruits preserved in syrup; salted nuts; canned fish in oil, cheese, un-reconstituted processed meat and fish such as ham, bacon, smoked fish Ultra-processed ready-toconsume products Formulated mostly or entirely from substances derived from foods.

Figure 1 :
Figure 1: Flowchart describing search for classification frameworks

Figure 2 :
Figure 2: Classification of foods by level of processing for 6 different classification frameworks ‡‡ Food groups used for assessment were the 135 food groups defined in the Euromonitor 2011 'Who Eats What Report' which covers the main sources of packaged and unpackaged foods across 80 countries[22].Lightest grey is least processed circling clockwise to black which is most processed with correspondingly ordered bulleted list descriptors.

Table 1 :
[22]er (%) of foods with different levels of processing for each classification framework according to original definitions and dichotomised as unprocessed or processed* *Food groups used for assessment were the 135 food groups defined in the Euromonitor 2011 'Who Eats What Report' which covers the main sources of packaged and unpackaged foods across 80 countries[22]

Table 2 :
[22]ortion of 135 food categories with the same designation (processed or unprocessed) for each pair of classification frameworks** **Food groups used for assessment were the 135 food groups defined in the Euromonitor 2011 'Who Eats What Report' which covers the main sources of packaged and unpackaged foods across 80 countries[22] Typically contain little or no whole foods.Durable, convenient, accessible, highly or ultra-palatable, often habit-forming.Typically not recognizable as versions of foods, although may imitate the appearance, shape and sensory qualities of foods.Many ingredients not available in retail outlets.Some ingredients directly derived from foods, such as oils, fats, flours, starches, and sugar.Others obtained by further processing of food constituents.Numerically the majority of ingredients are preservatives; stabilizers, emulsifiers, solvents, binders, bulkers; sweeteners, sensory enhancers, colours and flavours; processing aids and other additives.Bulk may come from added air or water.Micronutrients may 'fortify' the products.Most are designed to be consumed by themselves or in combination as snacks.Raw fruits, non-processed nuts, fresh raw vegetables, fresh grated vegetables, raw crustaceans/molluscs, fresh milk, fresh whole cream, raw meat, eggs, honey Processed staple/basic Foods industrially prepared involving high degree of processing such as drying, flaking, hydrogenation, heat treatment, use of industrial ingredients and industrial deep frying.Includes foods from bakeries and catering outlets requiring no or minimal domestic preparation apart from heating and cooking.Bread, pasta, rice, milk, butter, vegetable oils Moderately processed Industrial and commercial foods involving relatively modest processing and consumed with no further cooking Dried or semi-dried fruits, nuts, seeds, vacuum packed meat/salads, canned fruit (natural juices), canned vegetables (natural juices), green andchamomile tea, extra virgin olive oil, frozen meat Highly processed Foods industrially prepared involving high degree of processing such as drying, flaking, hydrogenation, heat treatment, use of industrial ingredients and industrial deep frying.Includes foods from bakeries and catering outlets requiring no or minimal domestic preparation apart from heating and cooking.Highly processed Foods that have undergone secondary processing into readily edible form, likely to contain high levels of added sugars, fats or salt Pastries, biscuits, cookies, sausages, prepared meats, ice cream, frozen desserts, breakfast cereals, confectionery, shortening, pasta products, prepared meals, complementary foods, crackers