To provide easily accessible, evidenced-based food knowledge online, we developed a platform named Food Atlas that could support diversified food information retrieval for different user groups, including specific populations, clinicians, and dietitians. The workflow of its construction is shown in Figure 1. The details are described in the following paragraphs.

To clarify the information needs at the population level, we collaborated with experts from the National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention, and clinicians and dietitians from Peking Union Medical College Hospital to understand the characteristics and needs of different populations. Combined with previous diet-related surveys [41], we summarized the following key insights: (1) There was a need for evidence-based dietary information to address issues such as low data quality and data inconsistency; (2) dietary recommendations are needed to alleviate physical symptoms that impact the individual experience of nutrition-related actions (such as fatigue, physical discomfort, food aversions , nausea, and complications); (3) practical information is needed to help put nutritional guidelines into practice.

Meanwhile, food choices are influenced by a diversity of factors that interact among them to produce a final behavior. Variability in food characteristics, including food constituents and technological processing, as well as variations in population groups’ characteristics, such as nutrient needs and physiology, can greatly affect the final decision [42,43]. Based on the corresponding factors, as well as important aspects with a close connection to the topic of food and population health, we collectively identified the core data content (such as ingredients, nutrients, cooking method, and dietary function) that needed to be covered in our multimodal food knowledge graph. We then analyzed the characteristics of information retrieval inputs for specific populations, clinicians, and dietitians, as well as based on our experience from a previous study [44]. After several rounds of discussion, we determined the system functions, including graph-based interactive visualizations, natural language retrieval, and image-text retrieval.

According to the core data content identified, our food-oriented graph consisted of triples, defined as T = (E, R, E), where E represented entities and R represented relations. Its schema mainly contains 11 entity types of verbal knowledge (including food, food category, ingredients, population, synonym name, dietary function, nutrients, cooking method, cookware, region, and season), as well as visual knowledge (visual perceptions reflected by images). We chose pregnant women as our target population and demonstrated how we used MedKaaS [45] to construct a multimodal Chinese food knowledge graph from evidence-based resources. MedKaaS is a tool set we developed for medical knowledge processing that is equipped with the knowledge schema design tool, knowledge extractor tool, knowledge fusion tool, and quality control tool.

First, a multidisciplinary expert group was established to guide the collection and selection of evidence-based resources. This group included obstetricians, nutritionists, dietitians, pediatricians, and Chinese medicine practitioners, all possessing extensive expertise with over 20 years of experience in prenatal nutrition. We then conducted searches for literature and books from PubMed and CNKI, as well as websites such as Amazon and JD, using predefined keywords including “(‘home-cooked dishes’ OR ‘recipe’) AND ‘pregnant woman’,” “‘pregnancy’ AND (‘nutrition’ OR ‘diet’ OR ‘dietary pattern’).” After manually reviewing the retrieved results, we removed irrelevant publications. Considering that there were few studies on whether specific Chinese foods were beneficial to maternal and infant health, we selected recipe books written by clinical nutritionists from third-class hospitals in China as the primary resources for foods in our knowledge graph. Subsequently, we collaborated with the multidisciplinary expert group to identify crucial nutrients and their associations with maternal and infant health from the retrieved literature. To unify category, we mainly referred to the Food Production License Classification Catalogue, 2017 Classification of National Economic Industries (GBT 4754-2017), and dietary patterns in the literature to define our food categories. Meanwhile, we integrated relevant content from existing food knowledge graphs [35,46]. After that, we used the knowledge graph schema design tool of MedKaaS to establish 11 classes and corresponding relations. A team of 3 annotators was then recruited. All annotators had a dietary research background and annotating experience. Two of the annotators used the knowledge extractor tool to curate all entities and their relations independently for each qualified food from the selected resources. This tool combines large language models and machine learning algorithms to ensure the effective extraction of evidence-based knowledge. Since the quality control tool showed that the consistency rate of annotation was 92%, a senior third-party annotator resolved disagreements. For images corresponding to these entities, we defined a set of image collection rules, including resolution, format, and content, and curated them through 2 search engines, Baidu and Bing. To ensure that the Food Atlas remained up to date, we regularly collected and curated incremental publications through the knowledge fusion tool. An overview of the data sources in Food Atlas is shown in Table S1 in Multimedia Appendix 1. The curation results were finally reviewed by the multidisciplinary expert group. Thereby, this knowledge graph contained a total of 2011 entities, 10,410 triplets, and 23,497 images.

To meet the needs of user information retrieval, Food Atlas mainly supported graph-based interactive visualizations, natural language retrieval, and image-text retrieval.

We evaluated Food Atlas from both the system perspective and the user perspective. From the system perspective, our platform complies with “CESI/TS 021-2020: Certification techniques specifications for knowledge graph construction platform” and “CESI/TS 043-2022: Certification techniques specifications for medical knowledge graph construction platform” issued by the China Electronics Standardization Institute and has obtained certifications. Additionally, we selected 5 representative online dietary platforms for functional comparison.

From the user perspective, we conducted a usability evaluation. We recruited participants from midwifery institutions in Beijing, China, including women preparing for pregnancy (or their spouses), pregnant women (or their spouses), clinicians, and dietitians. Participants were invited to try out Food Atlas, and the System Usability Scale (SUS) questionnaire [71] was used to conduct small-scale user surveys. This questionnaire has a 5-point Likert scale, ranging from “Strongly Agree” to “Strongly Disagree,” for 10 items. A total score of 68 (or higher) is regarded as “above average usability” [71]. An online survey platform, WJX, was used to collect survey data, and R (version 4.4.0; The R Foundation) was used for statistical analysis.

This study was approved by the Ethics Committee of the Institute of Medical Information, Chinese Academy of Medical Sciences (IMICAMS/01/21/HREC). All participants were informed that their responses would be used to inform public-facing research. All procedures were performed in accordance with the Declaration of Helsinki.

The schema of our multimodal food knowledge graph consists of 26 types of semantic relations among 11 entity types. Figure 3 details the correspondence between entities and relations.

The constructed food knowledge graph contains a total of 2011 entities, 10,410 triplets, and 23,497 images. Table 1 lists the descriptive statistics of our multimodal food knowledge graph.

A total of 28 participants used Food Atlas and evaluated its usability. Their demographic characteristics are shown in Table 3. The results showed that the target users of our platform were relatively broad, covering specific populations and their family members. One-half (14/28, 50%) of the participants (including all 3 dietitians) had experience using similar platforms, although the proportion was lower in the “preparing for pregnancy” and “pregnancy” groups (7/28, 35%).

The mean SUS index was 82.5 (SD 9.94; range 40.0-82.5), which indicates an above-average usability score. Table 4 displays the results of the SUS questionnaire. A total of 93% (26/28) of the participants wanted to use Food Atlas frequently, and 96% (27/28) stated that the platform had high consistency. This showed that Food Atlas was tailored to the needs of the target users, and its underlying data were reliable.

In this study, we collaboratively developed an online food data exploration platform, Food Atlas, with a multidisciplinary expert group. To demonstrate its feasibility for Chinese food for pregnant women, a user-friendly and high-quality food knowledge graph was provided; it included 2011 entities, 10,410 triplets, and 23,497 images, as well as diversified food information retrieval for different user groups. Our previous study [16] showed that an online prenatal education curriculum focusing on nutrition has the potential to reduce adverse outcomes in pregnant women. Considering information-seeking behaviors, perspectives, and preferences, multimodal resources are needed to reach people and enhance engagement with evidence-based information and health care [32,44,76]. Therefore, there is an urgent need for an easily accessible, credible, and appropriate evidence-based online food platform to promote healthy diets in specific populations, especially in this digital era.

To guarantee the high quality of Food Atas, we selected, assessed, and curated evidence-based dietary resources and established an update mechanism. It is difficult to determine whether a food, especially a mixed dish, is suitable for a certain population. This requires identifying reliable data sources, combining food information with healthy dietary patterns or recommendations for different populations, and coordinating inconsistencies among different data sources. In particular, the effects of some ingredients or compounds (such as alternative sweeteners) on human health are still unclear [77]. To address this issue, we worked closely with experts from clinical and nutrition-related institutions. Their extensive expertise helped us quickly identify high-quality evidence and standardize our food knowledge curation workflow. Of the participants, 96% (26/28) stated that the platform had high consistency, illustrating the necessity and effectiveness of health professionals participating in online evidence-based resource development.

We graphically represented curated multimodal food data, aiming to optimize the organization, presentation, and interaction of knowledge. A graph is a type of sparse data structure that consists of nodes and edges. Compared with other schemas, it can show knowledge more comprehensively, especially the relations between knowledge nodes [34]. Platforms like Foodwake also offer support for the graphical representation of food knowledge, which in turn shows its necessity. Different from existing food knowledge graphs about recipes [78], food safety [79], nutrient, and health [80], the schema we proposed covers the key content for a healthy diet for specific populations and has the ability to support the description of Chinese food. The constructed knowledge graph includes what the food is, what the dish is, when to eat, where to eat, how to cook, and who can eat. Both verbal knowledge and visual data are involved. Meanwhile, its construction process complied with 2 technical specifications issued by the China Electronics Standardization Institute and has been certified.

Furthermore, we offered various interactions with graph-structured data for easy access, as evidenced in previous studies [81]. Users can choose the appropriate interface according to their personal preferences or information needs, rather than being forced into a limited mode of communication. We used specific use cases to illustrate its “fitness of purpose” for various downstream applications, which has been carried out in previous related work [82,83]. Compared with other online dietary platforms, Food Atlas offers a distinct set of functions that can meet the specific needs of different users. An above-average usability score (82.5) indicates that Food Atlas is tailored to the needs of the target users, which also demonstrates the feasibility of our technical route.

Our online food data exploration platform, Food Atlas, serves as a valuable source of dietary knowledge, enhancing target users’ knowledge regarding nutrition at different life stages, recommended practices, and diets to alleviate physical discomfort symptoms. Given that people increasingly seek information on the web [23], our platform not only meets the rising demand for accessible, credible, and appropriate evidence-based online dietary resources but also enhances the engagement of health professionals. By continually updating and promoting this platform more widely, we aim to encourage healthier diets and improve health.

Our study has several limitations. First, it is a feasibility study based on Chinese food to promote health for pregnant women. The small-scale data sets limit its real impact. In addition, whether Food Atlas can be applied to other cuisines (such as western food and Japanese food) and population groups and to what extent it needs to be localized still need to be clarified and validated. Second, our platform lacks the traceability of evidence, and being equipped with this function will further enhance its reliability. Third, we did not quantitatively evaluate the performance of our natural language retrieval and image-text retrieval. In fact, existing state-of-the-art models [60,67] can be used in the corresponding system modules. With their assistance, the retrieval results would be improved. In the future, we will validate the feasibility and effectiveness of these methods.

This study outlined the development of an easily accessible, credible, and appropriate evidence-based online food platform for specific population health promotion and assessed its fitness for different user groups. The results indicate the necessity and effectiveness of health professionals participating in online evidence-based resource development. Various interactions and navigation strategies of graph-structured food data can meet the information needs of different user groups. The growing prominence of online evidence-based food resources presents an opportunity for healthy diets. To optimize Food Atlas, further research should focus on the potential feasibility for other population groups and cuisine, as well as state-of-the-art models.