Unveiling Dietary Complexity: A Scoping Review of Network Analysis in Dietary Pattern Research and a Methodological Roadmap for Future Research

1. Introduction

Dietary patterns have been associated with a variety of health outcomes. For example, the Mediterranean diet, characterised by a high consumption of fruits, vegetables, whole grains, and healthy fats, has been linked to the prevention of cardiovascular disease [1,2], better cognitive performance [3], and a longer life expectancy [4]. In comparison, the Western diet consisting of a high intake of red and processed meat, refined grains, sugars, fats, and fast food with a low intake of fruit and vegetables has been associated with higher rates of obesity [5,6] and an increase in cancer risk [7].

Currently, most of the research looking into nutrition and its effect on health has focused on analysing foods and nutrients separately from each other [8,9] or has produced ‘a priori’ diet quality scores or ‘data-driven’ composite scores [10]. These traditional dietary assessment methods do not expose food synergies, which may lead to an incomplete understanding of dietary patterns and their health implications [11,12]. Therefore, it is crucial to research not only what foods are consumed but also how foods are consumed in combination. For example, a recent study found that garlic may counteract some of the detrimental effects associated with red meat consumption, including a reduced risk of cardiovascular disease from a high intake of red meat [13]. This finding emphasises the need to examine the interactions between different foods to fully understand their health impacts.

The network approach offers a promising and more holistic way to analyse the co-consumption of foods. It enables the exploration of dietary patterns by using advanced statistical techniques to map and analyse the connections between various dietary components [12]. By capturing the co-consumption patterns and their associated outcomes, the network approach can reveal insights into the relationship between nutrition and health that traditional methods may have previously overlooked.

This review aims to determine whether network approaches can offer a more comprehensive analysis of dietary intake. By evaluating the effectiveness of these methods, this review aims to contribute to the development of more accurate dietary assessment tools, potentially leading to improved dietary recommendations and interventions which are tailored to human eating behaviours. First, the traditional approaches to dietary pattern analysis are examined, highlighting their limitations. Next, a comprehensive scoping review of studies which have applied network analysis to explore dietary patterns is presented, demonstrating its potential to address these challenges. Finally, guiding principles are offered to enhance methodological rigor and advance this innovative field.

1.1. Dietary Patterns and Health

Using dietary analysis in nutritional research is essential for the development of dietary interventions aimed at improving both physical and psychological wellbeing. Accurate dietary analysis allows researchers to identify dietary patterns and nutrients that influence health outcomes, enabling the formulation of targeted interventions. For instance, the Dietary Approach to Stop Hypertension (DASH) diet was developed after observational research found that a carbohydrate rich diet with fruits, vegetables and low-fat dairy products were associated with lowered blood pressure [14,15]. Notably, randomised controlled trials of individual nutrients such as magnesium, potassium, calcium, and fibre, had produced inconsistent results [14]. One explanation was that nutrients from dietary supplements may not benefit health as effectively as those obtained from whole foods due to the synergistic interactions between nutrients and other components present in the diet. Unfortunately, due to the limitations of conventional dietary pattern analysis, most nutritional interactions remain undiscovered. However, recent computational advances may help unveil non-additive and non-linear interactions [16], thereby improving dietary recommendations and the development of multi-component functional foods and supplements that benefit health [8].

1.2. Traditional Dietary Pattern Analysis

The traditional methods used for dietary pattern analysis include principal component analysis (PCA), cluster analysis, and a priori composite scores. These techniques are widely used to investigate the relationships between diet and health outcomes by summarising dietary intake data into meaningful patterns. Details of these traditional methods can be found in Table 1.

PCA is a statistical method which identifies underlying dimensions of food consumption by grouping food items based on the degree to which they correlate [17]. For example, PCA may reveal that certain foods like fruits, vegetables, and whole grains appear together in diets to form a healthy eating component [18]. This method has been used to link dietary patterns to various health outcomes, such as an association between a Western diet and increased prevalence of obesity [5], and between a Mediterranean diet and reduced occurrence of cardiovascular disease [1].

Cluster analysis groups individuals with similar diets into homogenous subgroups [19]. This method groups individuals who consume similar types and amounts of foods, which can be based on food frequency, standardised nutrient intakes, or a combination of dietary and biochemical measures. Further interpretation of these clusters involves comparing dietary profiles and health outcomes across groups, providing the insights into how specific dietary patterns influence health [17].

A priori composite scores, such as the Mediterranean diet score or the DASH score, are based on predefined dietary guidelines and recommendations [20,21]. These scores evaluate adherence to specific dietary patterns known to be beneficial for health, such as assessing the intake of key components of the Mediterranean diet, which has been linked to a reduced risk of cardiovascular disease [20] and improved cognitive function [3].

While these traditional dietary pattern analysis methods provide valuable insights into the associated health outcomes, there are limitations. One significant limitation is that they are unable to fully capture the complex interactions between different dietary components [22]. The beneficial effect of certain nutrients may be enhanced or inhibited by other dietary components, but these synergies are often hidden in these traditional analyses [23]. When dietary patterns are reduced to unidimensional scores, their multidimensional nature is disregarded. While these patterns may capture some synergies, this is only possible when interactions are explicitly recognised and incorporated during score development which is rare [24]. Moreover, in research focusing on individual nutrients or foods, interactions are often implicitly assumed to be non-existent in the model design [24]. Finally, these methods often assume dietary patterns are relatively static, ignoring potential changes in diet over time due to aging, economic changes, or health conditions [25]. These incorrect assumptions about interactions, or assumptions of staticity in model design can result in obscured or false associations and biased effect estimates.

In contrast the methods usually used to quantify dietary patterns, more prescriptive approaches such as linear optimisation have been instrumental in designing diets that meet specific targets for health, sustainability, or cultural appropriateness. Whilst beyond the scope of this review, these methods are fundamentally “knowledge-based,” meaning they can only optimise for the variables they already know about - the handful of macro- and micronutrients that have been well-characterised. This represents the “known knowns” of nutrition, which constitute less than 1% of the thousands of distinct bioactive compounds and phytochemicals present in our food chain. This inherent limitation means that such models are blind to the vast “nutritional dark matter” and the complex food synergies that are crucial for health.

To overcome the limitations of traditional dietary analysis, network approaches have emerged as a promising alternative, representing a superior, bottom-up alternative to knowledge-based prescriptive models like linear optimisation. Unlike the traditional methods which focus on individual nutrients and patterns in isolation [26], network analysis does not require comprehensive prior knowledge of every bioactive compound. Instead, it is a data-driven approach that learns directly from real-world eating behaviours. While also data-driven, this provides a key advantage over methods like PCA or cluster analysis; instead of reducing diet to composite scores or groups, network analysis explicitly maps the web of interactions and conditional dependencies between individual foods [12].

Methods such as Gaussian Graphical Models and Mutual Information Networks enable researchers to visualise and analyse these intricate relationships within a diet [27]. By mapping the connections between foods and nutrients, these methods reveal how they collectively influence health outcomes and allow for the discovery of beneficial food combinations and protective synergies that emerge rather than a pre-defined biochemical model. Furthermore, dynamic or time-varying networks can model how dietary patterns change over time within individuals or populations, turning the complexity of our diet from a limitation into a source of discovery [28].

A variety of network algorithms have been developed, although not all have hitherto been applied to diet (Table 2 and Table 3). Gaussian Graphical Models (GGMs) are probabilistic models that use partial correlations to identify conditional independence between variables. These models are particularly useful for exploring linear relationships in dietary data, offering insights into how one nutrient interacts with others while accounting for the broader dietary context. For example, GGMs can reveal whether the intake of saturated fats and sodium is conditionally independent given calorie consumption. This could help identify whether their relationship is direct or merely a byproduct of consuming high-calorie foods. This makes them valuable for understanding direct and indirect nutrient associations within diets. A limitation is that GGMs assume linear relationships, making them unsuitable for capturing the non-linear interactions that are often present in dietary data. For example, the effect of salt on hypertension may be moderated by the potassium and sugar content of the diet [29]. Additionally, GGMs are sensitive to non-normal distributions, which can distort the results in datasets with significant deviation [30].

Related to GGMs, Mixed Graphical Models (MGMs) accommodate datasets containing both continuous variables (e.g., nutrient intake) and categorical variables (e.g., demographic characteristics) [31]. This versatility is particularly useful for dietary studies that integrate diverse types of information. For example, MGMs can explore how continuous measures of dietary intake correlate with categorical socioeconomic factors such as education or income. By modelling these mixed data types jointly, MGMs expand the applicability of graphical models to more complex nutritional datasets, potentially yielding deeper insights into diet-health relationships. However, MGMs share several limitations with GGMs, including sensitivity to non-normal distributions for continuous variables [30].

Mutual Information (MI) Networks measure the amount of information shared between pairs of dietary components, capturing both linear and nonlinear associations [32]. This can uncover hidden patterns and relationships which may not have been found using traditional correlation-based methods [33]. For instance, by modelling nonlinear patterns, MI Networks can explore how sugar and fat intake interact to disproportionally influence obesity or cardiovascular risk, identifying subtle dependencies such as threshold effects that might be missed by simpler models. However, a significant limitation is that MI algorithms usually give rise to denser networks reducing interpretability and the ability to tease apart direct and indirect dependencies.

Bayesian Networks (BN) are probabilistic graphical models that represent the relationships between variables through directed acyclic graphs, enabling the identification of potential causal pathways [34]. BNs have not yet been applied to dietary pattern analysis however, unlike other traditional correlation-based methods, BNs provide insights into causality [34] which may make them a powerful tool to explore how changes in dietary components may influence one another. One possible application of BNs is to model the fat-sugar seesaw phenomenon, where reducing fat tends to lead to an increase of sugar in diet [35]. One advantage of BNs is their ability to incorporate prior knowledge into the model structure [36], potentially enhancing the interpretability and plausibility of the derived dietary network. There are limitations to BNs, in particular the computational difficulty of exploring a previously unknown network [37].

Dynamic Networks incorporate time-varying dependencies [38], enabling researchers to observe how dietary patterns and meal compositions evolve over time. This approach allows the study of how diets respond to external factors, such as seasonal changes or price increases, and how interventions might alter established dietary habits [39]. One possible application of Dynamic Networks is to understand how patterns of co-consumption may change following dietary interventions. For example, if a person tends to consume meat with vegetables and an intervention reduces meat consumption, Dynamic Networks could reveal whether this intervention also leads to an unintended reduction in vegetable intake. By modelling these changes, the unintended consequences of public health polices can be predicted, such as how promoting plant-based diets may inadvertently decrease the intake of other beneficial food groups. A limitation of Dynamic Networks is the need for detailed longitudinal monitoring of diet over time; this often involves resource intensive data collection methods, such as repeated dietary recalls or food diaries over extended periods of time.

Hypergraphs extend traditional graph theory by allowing edges, known as hyperedges, to connect more than two nodes [40], making it possible to represent group-level interactions or clusters. Standard graph models only consider the pairwise interactions, such as edges between nutrients or foods, while hypergraphs can account for higher-order interactions. This ability to model interactions involving multiple nodes would be particularly useful in dietary pattern research as multiple nutrients often work together with shared function to influence health outcomes. For example, a hyperedge in a hypergraph could represent a meal containing protein, fats and carbohydrates where the combined impact on health emerges from the interplay between these nutrients and cannot be explained by pairwise interactions alone. One limitation of hypergraphs is that their high computational demand can make them resource-intensive [41], especially for large datasets. Another limitation is that the complexity of hypergraphs often makes them hard to interpret [42], reducing their accessibility for researchers who are less familiar with advanced network methods.

Multilayered Graphs represent systems with multiple interconnected layers, each capturing distinct but related interactions [43]. For example, one layer may represent nutrient interactions, another food-level relationships, and another may represent food context. These graphs allow for a comprehensive exploration of both intra-layer and inter-layer connections [43]. Analysing cross-domain relationships is valuable in nutrition research as the impact of food context on nutrient intake can be examined. However, Multilayered Graphs are computationally demanding, especially for large datasets spanning multiple domains. They are also complex and can be challenging to interpret [43].

The choice of method depends on the research question, data type, and available computational resources. By leveraging these diverse tools, researchers can gain deeper insights into dietary patterns, overcoming some of the limitations of traditional dietary pattern analysis (Table 2).

The purpose of this paper is to provide an overview of the existing studies using network approaches for dietary pattern analysis. Specifically, the objectives of this scoping review were to (1) provide a clearer understanding of the strengths and limitations of traditional dietary assessment methods, (2) identify the potential advantages of network approaches, (3) determine whether network approaches can offer a more nuanced analysis of dietary intake data, (4) create guiding principles for the use of network approaches in dietary pattern analysis.

2. Methods

2.1. Search Strategy and Selection Criteria

This scoping review aims to identify and critically evaluate literature on the application of networks to dietary data. Studies were included if they used network analysis to analyse dietary data for human participants. In particular, the network models searched for included Gaussian graphical models, mutual information networks, multilayered networks, hypergraphs, and mixed graphical models. Studies were excluded if network analysis was used for metabolics or systems biology.

A comprehensive search for studies written in the English language and published up to 7th March 2025 was carried out using PubMed, Scopus, and PsycINFO. The following search terms were used: “network approach” or “network analysis” or “network method” or “graphical model” and “gaussian graphical model” or “GGM” or “mutual information network” or “mixed graphical model” and “dietary analysis” or “dietary data” or “nutrition analysis” or “food intake” or “diet” or “nutrition.” Google Scholar was used to obtain additional articles identified by journal hand searching. After deduplication, titles and abstracts were read to assess eligibility based on predefined inclusion criteria. Irrelevant articles or duplicates were excluded. Full texts of the remaining articles were read to verify their suitability. Reference lists from the articles deemed suitable were checked for additional studies. This systematic search was conducted independently by two reviewers (RT and JM); any disagreements were resolved by discussion.

2.2. Data Extraction

Data was extracted by RT, and checked by JM, from the eligible articles using a standardised spreadsheet: author and year of publication, study aims, participant information (age, gender, ethnicity), dietary assessment used, approach to derive dietary patterns, network model used, and appropriateness of the network model.

Consistent with PRISMA-ScR and JBI guidance, we did not appraise risk of bias because our aim was to map methodological characteristics, not to evaluate intervention effects.

3. Results

3.1. Search and Selection of Network Studies

The search conducted in March 2025 identified 171 studies potentially relevant to the review. After deduplication and relevance screening based on language, dietary data collection methods, and network analysis approaches, 19 studies met the eligibility criteria for a full-text review. One study did not have an English translation so was not included in the review [44]. The remaining 18 articles were read, and all were deemed relevant for this scoping review. The flow of studies from identification to final inclusion is represented in Figure 1.

3.2. Study Characteristics of Included Network Studies

The general characteristics of the network studies included in this review are summarised in Table 3. All studies were published between 2018 and 2024. Most (79%) studies used FFQs to collect their dietary data, while two studies used 24-hour dietary recall, and one study used a Mediterranean adequacy questionnaire.

Nine studies analysed data from large, pre-existing cohorts: four from the Cancer Screening cohort in South Korea, three from the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort, one from the Lifelines cohort, and one from the 3C study. The remaining nine studies used smaller, non-cohort-based samples. Sample sizes varied from 230 participants to 74 132 participants.

Regarding network approaches, eleven studies used Gaussian graphical models, with one of these confirming their results with a semiparametric extension – semiparametric Gaussian copular graphical model (SGCGM). Two studies used only SGCGM. Three studies used mutual information matrices. Two studies used mixed graphical models.

A wide range of health outcomes were analysed, highlighting the versatility of network approaches in dietary research. Cancer was the most frequently studied outcome, with five studies varying on various cancer types, including gastric and breast cancers. Other physical health outcomes analysed were incident prediabetes, metabolic syndrome, obesity, anhedonia, non-alcoholic fatty liver disease, and diet quality during pregnancy. Neurological disorders were also explored, with one study on dementia and one on multiple sclerosis. Additionally, one study compared dietary patterns between men and women, and another examined differences between meal-specific and habitual dietary networks.

3.3. Adherence to Methodological Best Practices

The remainder of this results section provides a thematic analysis of the included studies, evaluating their adherence to several key methodological best practices for network analysis (Table 4).

3.3.1. Justifications for Using Network Models

The included studies justified their use of network analysis in several ways. The most common rationale was to counter the known limitations of traditional methods, a justification made by nine of the reviewed studies [45,46,47,48,49,50,51,52,53]. For instance, authors noted that network analysis was chosen because Principal Component Analysis (PCA) can explain only a small proportion of the variability in food intake [52] and fails to demonstrate pairwise correlations between food groups [50].

Another common theme was the use of network analysis to complement existing research and gain additional insights into complex data [33,54,55]. One study framed its use as complementary without explicitly mentioning the limitations of traditional methods [56]. Two studies specifically focused on using network analysis to overcome the limitations of diet scores, allowing for an assessment of diet as a pattern rather than a sum of single food items [57,58].

Furthermore, three studies used network analysis directly alongside traditional dietary pattern analysis methods [59,60,61]. This was often done to compare the dietary patterns identified by each approach, for example by using network analysis to derive dietary networks and then using PCA and reduced rank regression (RRR) for comparison [59], or by comparing new network-derived patterns to PCA patterns from the same cohort in an earlier study [60].

3.3.2. Study Design and Causal Inference

The vast majority of the reviewed studies utilised a cross-sectional design. Acknowledging the limitations of this approach, most of these studies (13 out of 18) did not attempt to make inferences about causality [33,46,47,50,51,52,53,54,55,57,58,60,61]. Among these, three studies explicitly stated that the use of cross-sectional data was the reason for this caution [48,49,56]. One paper acknowledged the cross-sectional design as a limitation but did not specify why [50]. In contrast, only one paper attempted to make inferences about causality from its cross-sectional data [59].

3.3.3. Network Estimation and Regularisation

The reviewed studies employed several techniques to estimate their networks and control for spurious connections, with approaches varying by the chosen network model.

Among the studies using Gaussian Graphical Models (GGMs), LASSO regularisation was the most common approach, used in thirteen of the fourteen GGM-based studies [45,46,47,54,57,59] [49,50,51,52,55,60,61]. Only one GGM study did not employ a regularisation method [58]. However, the transparency in reporting the specific LASSO tuning parameter was inconsistent; just seven of these thirteen studies provided this detail [46,47,54,55,57,59,61], while one study explored network structures across a range of different tuning parameters [45].

For the three studies that used mutual information networks, all applied thresholding to reduce network density [33,53,56], and two of these also used permutation testing to retain only statistically significant connections [33,56].

Regarding the stability and novelty of the findings, two studies explicitly refrained from drawing strong inferences from their results [56,60], while two others, noting they were the first in their specific populations, did make inferences from their findings [50,52].

3.3.4. Use and Interpretation of Centrality Metrics

The application of centrality metrics to identify important foods or nutrients was inconsistent across the reviewed literature. Five of the eighteen studies avoided using centrality metrics in their analysis [46,52,53,55,56]. In contrast, the majority of studies (13 out of 18) did employ centrality metrics, and none of these studies acknowledged or discussed the potential limitations of this approach in the context of dietary network analysis [33,47,48,49,50,51,54,57,58,59,60,61]. For instance, one study examined node centrality using strength, betweenness, and closeness, ultimately opting to use the strength metric for its final analysis [61].

3.3.5. Handing of Non-Normal Data

The studies employing Gaussian Graphical Models (GGMs) used three distinct strategies to address the assumption of normally distributed data. The most robust approach was to use a non-parametric extension of the GGM; two studies used the semiparametric Gaussian copular graphical model (SGCGM) exclusively [51], and a third study used SGCGM to confirm the results of their primary GGM analysis [55].

The most common strategy was data transformation, with six studies applying a log-transformation to their data to improve normality [46,47,54,57,59,60].

Finally, three of the GGM studies did not apply any correction for non-normal data. Of these, two acknowledged the issue as a limitation but did not address it [49,50], while one study did not acknowledge the limitation at all [58].

4. Discussion

This scoping review provides a comprehensive overview of the emerging field of dietary network analysis. The 18 identified studies, all published since 2018, demonstrate the versatility of these methods across a wide range of populations and health outcomes, from cancer to neurological disorders. Our findings show that researchers are actively employing network analysis to overcome the recognised limitations of traditional methods and to gain more nuanced insights into complex food co-consumption patterns. However, our thematic analysis also reveals significant methodological heterogeneity and several recurrent challenges across the literature. These include a heavy reliance on cross-sectional data which limits causal inference, inconsistencies in network estimation and reporting, and the widespread, often uncritical, application of analytical techniques - such as centrality metrics and simple data transformations - that may be inappropriate for complex dietary datasets. These identified challenges highlight the need for a more standardised approach. Therefore, the remainder of this discussion will use these findings to establish a set of guiding principles designed to enhance the rigour and reliability of future research in this promising field.

4.1. Guiding Principles for Future Research

One objective of this review was to create guiding principles for the use of the network approach in dietary pattern analysis. This was achieved by using the critiques of using network analysis to analyse multivariate data [62] and applying them to dietary pattern analysis. A summary of these guiding principles can be found in Figure 2.

Principle 1: Selecting Appropriate Models

Researchers should only use network analysis when it addresses specific limitations of traditional multivariate methods or provides complementary insights that align with the research question [62]. Using network analysis to counter the limitations of widely used dietary pattern analysis methods and to complement the data obtained by these methods is effective in building on the existing knowledge to uncover interactions that have previously been concealed. It is critical, however, to ensure that network analysis is the most appropriate method for the research question at hand.

Principle 2: Aligning Study Designs with Research Questions

Researchers need to avoid making causal inferences from between-person cross-sectional data and ensure that the study design aligns with the research objective. When using between-person cross-sectional data, it limits the inferences that can be made about causation, as correlations among variables do not indicate causality [62]. It is important to refrain from making any inferences about causation, as it is not possible to use between-person statistical associations to draw conclusions about how food groups are related for a given individual [62]. Making inappropriate causal inferences can lead to misleading conclusions about relationships between food groups and health outcomes. For causality-focused research questions, longitudinal data may be better suited.

Principle 3: Best Practices for Reliable Network Estimation

The robustness of estimated networks can be questionable, as networks may appear similar in their global characteristics while their detailed characteristics vary substantially [62,63]. Therefore, inferences should not be drawn before results have been rigorously replicated [62]. To create more reliable and interpretable networks, specific best practices should be followed.

When using GGMs, robust methods such as LASSO regularisation should be used to minimise spurious connections, as this technique limits the number of spurious edges to obtain more interpretable networks [64]. It is also important for researchers to provide details of their tuning parameter choices to ensure reproducibility, as this parameter controls the level of sparsity in the network [64]. Similarly, when using mutual information networks, techniques like thresholding (retaining edges above a certain value) and permutation testing (recalculating mutual information on randomised data to keep only significant connections) should be used to create a more robust network with fewer spurious edges [65,66].

Principle 4: Valid Interpretation of Network Metrics

Researchers should avoid relying on centrality metrics (e.g., degree, closeness, betweenness) for dietary networks. Most network centrality metrics were developed for bounded networks, where the nodes are clearly defined and fixed [62]. Dietary pattern analysis, however, often involves unbounded networks, which are dynamic and do not have a fixed number of nodes or connections. When centrality metrics are applied to unbounded networks, their interpretability is compromised, making them unstable and unreliable [62]. This can lead to inaccurate conclusions. For example, one study concluded that “a low centrality predictability of low-fat milk intake in the networks may indicate that significant associations in the regression analysis could be due to intake levels coinciding with influential risk factors” [61]. However, given the limitations of centrality metrics in unbounded networks, this conclusion may not be valid. While the choice of the ‘strength’ metric in that study was based on its relative stability, this does not negate the broader issue that these metrics can produce misleading information [67]. Therefore, researchers should either avoid centrality metrics entirely or adopt alternative approaches that account for the unbounded nature of dietary networks.

Principle 5: Addressing Non-Normality in Dietary Data

When using GGMs, researchers must address the fact that dietary data is rarely normally distributed, which violates a core assumption of the model [68]. Dietary data is often skewed and zero-inflated (i.e., contains many “no consumption” reports), which presents significant challenges. While transformations such as log-transformation can mitigate some skewness, they are not a robust solution for zero-inflated data [69]. Adding a small constant to avoid errors with zero values can distort the data and complicate the interpretation of the results [70]. More appropriate approaches should be employed, such as using the semiparametric Gaussian copular graphical model (SGCGM), which is a non-parametric extension designed for skewed data. Alternatively, opting for a different method entirely, such as a mutual information network or mixed graphical model, may be a better choice when dealing with non-normally distributed data.

4.2. Strengths and Limitations of This Review

A key strength of this review is its systematic and rigorous methodology. We conducted a comprehensive search across multiple relevant databases and followed the PRISMA-ScR guidelines for scoping reviews. The inclusion of two independent reviewers for the study selection process minimises the risk of bias and enhances the reliability of our findings. Furthermore, this review moves beyond a simple summary of the literature. By critically evaluating the included studies against an established framework of best practices and generating a novel set of guiding principles, this work provides a tangible roadmap for future researchers in this emerging field.

Despite these strengths, several limitations should be acknowledged. First, as a scoping review, our aim was to map the field rather than conduct a formal quality assessment or risk-of-bias analysis for each individual study; therefore, the quality of the primary studies was not formally appraised. Second, our review may be subject to publication bias, as studies with null or non-significant findings may be underrepresented in the published literature. Additionally, our search was limited to English-language publications, however, only one relevant study was excluded because no English translation was available. Finally, dietary network analysis is a rapidly evolving field. While our search was comprehensive up to its cut-off date, new studies will have been published in the interim.

4.3. Conclusion

In conclusion, network analysis offers a powerful and promising alternative to traditional methods for dietary pattern research, allowing for a more holistic understanding of food synergies and co-consumption. However, our scoping review has revealed that the application of these methods is currently marked by significant methodological heterogeneity and a frequent lack of adherence to best practices, particularly concerning study design, network estimation, and the interpretation of metrics. The guiding principles established in this review provide a foundational framework to address these challenges. By adopting these principles, future research can harness the full potential of network analysis to generate more robust, reliable, and interpretable findings, ultimately advancing our understanding of the complex relationship between diet and health.