Conventional dietary assessment methods

Accurate assessment of dietary intake in daily life and identification of an individual’s representative habitual intake are crucial for developing personalised dietary advice. Conventional dietary assessment methods used in epidemiological studies can largely be categorised into four approaches: dietary records, 24HR, FFQ and dietary biomarkers^{(Reference Bailey10)}. Each method possesses intrinsic strengths and limitations, influencing their distinct applications^{(Reference Moshfegh, Rhodes and Martin11)}. Dietary records involve real-time recording of all foods consumed, while 24HR relies on retrospective recall with the assistance of trained interviewers. Both methods can yield detailed and accurate intake data but may not capture habitual long-term dietary patterns. FFQ, in contrast, are designed to assess habitual intake over extended periods, typically the past year, by querying the frequency of consumption of approximately 100–150 food items. While FFQ are cost-effective and suitable for large-scale epidemiological studies, they have limitations such as recall bias and difficulties in estimating portion sizes – particularly in Asian dietary contexts, where meals often consist of mixed dishes and shared portions. Furthermore, FFQ have limited ability to capture emerging dietary variables, such as meal timing and ultra-processed food consumption, which are increasingly considered relevant to chronic disease risk. Despite these challenges, FFQ remain the primary tool in prospective cohort studies because of their practicality and ability to capture habitual intake. Dietary biomarkers measured in biological specimens, such as blood and urine, can provide objective indicators of dietary exposure and complement self-reported dietary assessment methods. Examples include urinary nitrogen for protein intake and serum carotenoids for fruit and vegetable intake. Although they cannot comprehensively capture overall dietary patterns, dietary biomarkers play an important role in validating self-reported dietary data.

Advances in dietary assessment tools utilising new technology

Research has been focusing on refining and developing new tools to improve accuracy and reduce the burden of both researchers and participants, particularly by integrating digital tools such as smartphones, cameras and sensors into the four existing dietary assessment methodologies^{(Reference Zhao, Xu and Li12,Reference Skinner, Toumpakari and Stone13)} , as depicted in Figure 1.

Fig. 1.

Evolution of dietary assessment tools from past to future. The x-axis represents the progression from manual tools to technology-based tools, while the y-axis differentiates the four dietary assessment methods. 24HR, 24-h recall; DR, dietary record.

Integration of digital technology into dietary assessment can be traced back to the advent of paperless forms of 24HR and FFQ^{(Reference Neuhouser, Prentice and Tinker14)}. Examples of this digital integration include the online Automated Self-Administered 24HR^{(Reference Subar, Kirkpatrick and Mittl15)}, along with various mobile device applications as summarised by Schembre et al.^{(Reference Schembre, Liao and O’Connor16)}. More recently, image-based dietary record tools emerged, which utilise digital cameras to capture food consumption via photographs^{(Reference Stumbo17)}. The objective is to decrease the burden on participants and resolve issues of underestimation and recall bias inherent in traditional methods. Examples of these tools include the Mobile Food Record^{(Reference Boushey, Spoden and Zhu18)} and the Remote Food Photography Method^{(Reference Martin, Correa and Han19)}, where participants capture everything they eat during a defined time period by taking a photo before and after each meal. However, challenges such as high costs, lack of database and the need for enhanced algorithms to analyse the images remain.

Moreover, wearable cameras attached to various body parts, such as the SenseCam and e-Button^{(Reference Sun, Burke and Mao20)}, and sensors worn on the neck or wrist detecting eating motions and the tooth and epidermal sensors detecting metabolites are designed to reduce bias and participant burden while capturing continuous real-time data^{(Reference Skinner, Toumpakari and Stone13)}. Dietary records using wearable cameras and sensors represent a novel approach in dietary assessments, with wearable devices continuously capturing dietary events or specific nutrients or metabolites associated with eating. Recent initiatives, such as the Nutrition for Precision Health powered by the All of Us Research Program, have also incorporated emerging technologies, including wearable devices (e.g. the Automatic Ingestion Monitor-2)^{(Reference Doulah, Ghosh and Hossain21)}, to improve the objective assessment of dietary behaviours.

Although technology-based dietary assessment tools offer several advantages, including reduced participant burden, automated data processing and the ability to capture eating behaviours in real time, they also present important challenges. Most currently available tools still rely on self-reported dietary intake and therefore remain susceptible to measurement error and reporting bias. In addition, rigorous validation, standardised reporting and transparent documentation of tool development are essential to ensure data quality. Accordingly, these technologies should be viewed as complementary to, rather than replacements for, conventional dietary assessment methods^{(Reference Eldridge, Piernas and Illner22)}. No single dietary assessment method is optimal for all research settings, and the choice of method should be guided by the specific research objective while balancing considerations such as accuracy, participant burden, feasibility, cost and scalability. Consequently, although technology-based tools may enhance dietary assessment in specific settings, FFQ continue to play a central role in prospective cohort studies because they provide a practical, cost-effective and scalable approach for assessing habitual dietary intake over extended periods in large populations, despite their inherent limitations in measurement error and recall bias.

Historical context and rationale

Nutritional epidemiology has undergone a paradigm shift over the past several decades as depicted in Figure 2. Early research focused primarily on nutrient deficiencies^{(Reference Willett23)}. With the development of FFQ and the establishment of large cohort studies, attention shifted towards evaluating the effects of single nutrients on chronic disease risk. However, major randomised controlled trials, such as the Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study, yielded null or inconclusive findings on the benefits of vitamin and mineral supplementation⁽²⁴⁾, prompting a transition towards examining individual foods and food groups. Also, O’Connor and colleagues^{(Reference O’Connor, Evans and Ivlev25)} systematically reviewed trial evidence to update the US Preventive Services Task Force recommendation on the efficacy of vitamin supplements for primary prevention. The authors concluded that there was no clear evidence of a beneficial effect of supplements on all-cause mortality, CVD or cancer^{(Reference O’Connor, Evans and Ivlev25)}. More recently, the focus has shifted to dietary pattern analysis, which captures the combined and potentially synergistic effects of multiple foods and better reflects real-world dietary behaviours. This evolution is also evident in major cancer prevention reports. While the 2018 report of the World Cancer Research Fund⁽²⁶⁾ primarily focused on individual foods and nutrients, its 2025 update⁽²⁷⁾ incorporates dietary patterns and lifestyle patterns as a more holistic approach to cancer prevention. This evolution in dietary pattern research has also been accompanied by growing recognition that diet is a multidimensional and dynamic exposure. Building on this perspective, the National Cancer Institute and the National Institutes of Health convened the ‘Extending Methods in Dietary Patterns Research’ workshop, which identified several methodological priorities for advancing dietary pattern research, including improved dietary assessment, standardisation of dietary pattern methods, consideration of dietary dynamism across the life course and better approaches for addressing measurement error^{(Reference Reedy, Subar and George28)}. These recommendations provide an important conceptual foundation for the continued development of dietary pattern research.

Fig. 2.

Shift of the focus of nutritional epidemiology research. ATBC, Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study; DASH, Dietary Approaches to Stop Hypertension; NIH, National Institutes of Health; USPSTF, US Preventive Services Task Force.

Classification of analytical approaches

Dietary patterns are defined as the quantities, proportions, variety and combinations of different foods and beverages in diets and the frequency with which they are habitually consumed^{(29,Reference Reedy, Krebs-Smith and Hammond30)} . Two primary methodological approaches have been established for analysing dietary patterns: a priori methods and a posteriori methods^{(Reference McNaughton, Marriott, Birt, Stallings and Yates31,Reference Zhao, Li and Gao32)} .

A priori methods, such as the Healthy Eating Index^{(Reference Kennedy, Ohls and Carlson33)}, rely on predefined scoring systems based on established dietary guidelines. These methods use fixed criteria to assess adherence to recommended dietary behaviours, producing standardised and interpretable scores. Large collaborative efforts, such as the Dietary Patterns Methods Project, have further demonstrated that major dietary quality indices – including the Healthy Eating Index, Alternative Healthy Eating Index, Mediterranean Diet score and Dietary Approaches to Stop Hypertension score – consistently capture key aspects of healthy dietary patterns and provide a robust basis for dietary guidance^{(Reference Liese, Krebs-Smith and Subar34)}. In contrast, a posteriori methods are data-driven and include techniques such as PCA, factor analysis and CA. PCA and factor analysis extract continuous dietary patterns – such as the ‘Western’, ‘Prudent’ or ‘Traditional’ diet – and assign each person an adherence score^{(Reference Schulz, Oluwagbemigun and Nothlings35)}. Although PCA and factor analysis are often used for similar purposes in nutritional epidemiology, they differ conceptually. PCA is a technique that transforms correlated variables into orthogonal components explaining the maximum variance in the observed data^{(Reference Greenacre, Groenen and Hastie36)}, whereas factor analysis assumes that the observed correlations arise from a smaller number of underlying latent factors^{(Reference Tavakol and Wetzel37)}. While these approaches capture variation in dietary intake, the reliance on continuous scores can limit direct interpretability in some contexts. CA instead produces mutually exclusive groups with distinct dietary profiles^{(Reference Schulz, Oluwagbemigun and Nothlings35)}, facilitating clearer subgroup identification and potentially more actionable results. The commonly used method is the k-means clustering algorithm, which partitions the data into k clusters by assigning each data point to the nearest cluster centroid after specifying the desired number of clusters. Recent work has further demonstrated that CA can complement Healthy Eating Index-based assessment by identifying distinct dietary subgroups within individuals with similarly high diet quality scores, highlighting heterogeneity that may not be captured by a single dietary index^{(Reference Liese, Wambogo and Lerman38)}. Clustering quality is often evaluated using the silhouette coefficient, which compares the average distance between a data point and all other points in the same cluster with the average distance to points in the nearest neighbouring cluster. The coefficient ranges from −1 to 1, with values closer to 1 indicating better-defined and more cohesive clusters^{(Reference Kaufman and Rousseeuw39)}. Although the silhouette coefficient is widely used to assess cluster separation, it has limitations because it is sensitive to cluster shape and density and therefore should not be used as the sole criterion for evaluating clustering performance^{(Reference Lengyel and Botta-Dukát40)}. Hybrid methods, including reduced rank regression^{(Reference Schulz, Oluwagbemigun and Nothlings35)}, combine elements of both approaches by linking dietary patterns to predefined health outcomes or biomarkers. Although this outcome-driven focus can be valuable for elucidating diet-disease relationships, it may be less suited to identifying novel dietary patterns that are not closely aligned with the selected response variables.

Machine learning methods in dietary pattern analysis

Machine learning encompasses a range of algorithms and statistical models designed to automatically learn patterns from data^{(Reference Nichols, Herbert Chan and Baker41)}. In dietary pattern research, machine learning offers powerful tools for handling complex, multidimensional datasets^{(Reference Kirk, Kok and Tufano42)}. Historically, the field of dietary pattern analysis has relied on unsupervised learning, with PCA and CA being the most traditional techniques^{(Reference Wingrove, Lawrence and McNaughton4)}.

However, generating stable and meaningful clusters from dietary data can be challenging. Outside the field of nutritional research, dimensionality reduction techniques – including PCA and UMAP – have been applied prior to clustering to improve clustering quality, enhance visualisation and reduce noise^{(Reference Xiang, Wang and Yang43–Reference Seo, Kim and Woo45)}. Beyond conventional machine learning, methods that perform clustering after dimensionality reduction with deep learning models such as AE⁽⁴⁶⁾ have been proposed since the 2010s^{(Reference Xie, Girshick and Farhadi47)}. According to a recent survey, these approaches, often referred to as ‘deep clustering’, have primarily been applied to image, text, video and graph data^{(Reference Zhou, Xu and Zheng48)}. Also, their use has extended to tabular data as well^{(Reference Svirsky and Lindenbaum49,Reference Rabbani, Medri and Samad50)} .

However, evidence supporting their application to dietary pattern analysis remains limited, and their potential benefits in nutritional epidemiology have yet to be systematically evaluated. Most studies perform clustering directly on dietary intake data; only a few have used PCA as a preprocessing step^{(Reference Engeset, Dyachenko and Ciampi51–Reference Edefonti, Randi and Decarli54)}, and the application of newer methods such as UMAP is limited to a single known example^{(Reference Li, Momma and Chen55)}.

This gap suggests untapped potential for improving dietary pattern research through integrated workflows that combine dimension reduction with clustering. The literature review in the Methodological trends and research gaps in epidemiological studies section will examine these possibilities in greater detail, highlighting both current applications and opportunities for methodological advancement.

Methodological trends and research gaps in epidemiological studies

Numerous epidemiological studies have applied PCA or CA to FFQ-derived dietary data. PCA is more commonly used, as it effectively summarises dietary variation into a smaller set of components. However, unlike PCA, CA offers the advantage of classifying individuals into discrete groups, making the results more interpretable and potentially actionable for public health recommendations. However, ensuring high-quality, well-separated clusters in FFQ data remains challenging.

To date, only a small number of studies have applied CA to derive dietary patterns and examine their associations with cancer risk (Table 1).

Specifically, five studies applied CA alone^{(Reference Li, Liu and Xie56–Reference Brunner, Mosdøl and Witte60)}, and five studies combined dimensionality reduction with CA^{(Reference Engeset, Dyachenko and Ciampi51–Reference Edefonti, Randi and Decarli54,Reference Krieger, Cabaset and Pestoni61)} . Among these, only two – by Wirfalt et al.^{(Reference Wirfalt, Midthune and Reedy59)} and Engeset et al.^{(Reference Engeset, Dyachenko and Ciampi51)} – were prospective cohort studies using cancer incidence as the outcome. Notably, Engeset et al. investigated the association between clusters and cancer incidence in the European Prospective Investigation into Cancer and Nutrition–Norway cohort. One case–control study by Qarmiche et al.^{(Reference Qarmiche, El Kinany, Otmani, El Rhazi and Chaoui52)} reported a silhouette score for clustering quality; however, the score was low (0·12). Apart from this study, no other research reported quantitative metrics of cluster quality. This scarcity of studies underscores the need for further methodological development and application of CA in cancer epidemiology.

Given the limited number of prospective cohort studies that have applied CA to cancer outcomes, it is informative to examine studies investigating other chronic disease outcomes – particularly metabolic syndrome and its components. Table 2 summarises several prospective cohort studies that have applied CA to examine the risk of metabolic syndrome or its components^{(Reference Li, Momma and Chen55,Reference Brunner, Mosdøl and Witte60,Reference Kim, Kim and Shin62–Reference Kimokoti, Gona and Zhu70)} . In most cases, clustering was conducted without prior dimensionality reduction, which may limit cluster separability and obscure meaningful dietary distinctions. The number of clusters was also typically small (3–5), potentially overlooking dietary heterogeneity. A notable exception is a 2024 study by Li et al.^{(Reference Li, Momma and Chen55)}, which applied UMAP prior to clustering to investigate hypertension risk. The reported silhouette score was 0·4, but higher than in previous studies – indicating a methodological improvement. Nonetheless, the study was constrained by a small sample size and short follow-up period. Across these studies, several recurring methodological limitations are evident. First, quantitative metrics to assess clustering quality – such as silhouette scores – were rarely reported, making it difficult to evaluate the robustness and separation of the identified clusters. Second, clustering was often performed without prior dimensionality reduction, a practice that may reduce cluster separability and obscure meaningful distinctions in dietary patterns. Third, the number of clusters was typically restricted to 3–5, which could oversimplify the diversity of dietary behaviours within a population. Finally, the application of advanced machine learning-based dimensionality reduction techniques capable of capturing non-linear relationships in dietary data remains limited. Addressing these limitations through more systematic analytic frameworks and the adoption of machine learning methods has the potential to enhance the identification of meaningful dietary patterns and improve their applicability across diverse populations.

Despite differences in study design and disease outcomes, several common patterns emerged across the reviewed studies. Overall, the available epidemiological evidence suggests a generally consistent association between cluster-derived dietary patterns and chronic disease risk. Dietary patterns characterised by higher consumption of vegetables, fruits and dairy products were generally associated with lower risks of cancer or metabolic diseases. In contrast, dietary patterns characterised by refined grains or unhealthy dietary habits tended to be associated with increased disease risk. Although the magnitude and statistical significance of these associations varied across studies, the overall direction of the findings was broadly consistent. Differences in study populations, dietary assessment methods, clustering algorithms and outcome definitions likely contributed to the observed heterogeneity.