We chose the domain and scope of our ontology based on social big data in order to answer the following questions: “How do individuals perceive the factors contributing to obesity?” “What are the sociodemographic characteristics of obese persons?” “How do individuals and healthcare providers evaluate the types and degree of obesity?” “What types of obesity exist?” “What are the results of obesity?” and “How do individuals and healthcare providers treat and prevent obesity?”

We searched publications using the PubMed, EBSCO, and Google Scholar databases with the keyword ‘obesity ontology.’ We then reviewed the purpose of development and the extracted concepts of the retrieved ontologies to consider their utilization.

We identified five obesity-related ontologies [1718192021] whose purposes and domains are presented in Table 1. Two of the ontologies [1718] were developed for the diagnosis and evaluation of obesity, but they did not address its treatment or prevention. The obesity ontologies for intelligent e-therapy [19] and for application in the mobile-device domain [20] included not only evaluations but also interventions for obesity, but they excluded medical treatment. Finally, an ontology for supporting a knowledge-based infrastructure of childhood obesity [21] included generic and customized preventive recommendations and several risk factors of childhood obesity, but did not cover adult obesity.

Since each of the retrieved ontologies dealt with only part of obesity-related concepts, it was necessary to develop a new ontology to integrate the following concepts: factors contributing to obesity, sociodemographic characteristics of obese persons, evaluation methods, types, signs, symptoms, complications, medical treatment or interventions by healthcare providers, and self-management strategies of obesity.

We enumerated the important terms related to the domains and scope of our ontology. The core concepts and terms related to obesity were extracted from the clinical practice guidelines on obesity published by the National Institute for Health and Clinical Excellence (NICE) [22] and the Korean Society for the Study of Obesity (KSSO) [23]. We searched obesity-related social media postings on Internet blogs, Twitter, and Facebook from January 2011 to June 2014 by combining ‘obesity’ and obesity-related keywords representing the following domains: obesity risk factors, sociodemographic characteristics, evaluation methods, types, signs, symptoms, complications, and management strategies. We also extracted additional terms from the retrieved postings that are not included in the guidelines.

The extracted terms were enumerated using a top-down method [11], starting from the most-general concept to more detailed concepts, to define the common concepts as classes and order them hierarchically.

The properties of classes and their facets, such as value types, allowed values, cardinality, and other features of the values, were defined to describe the internal structure of the extracted concepts.

The obesity ontology was developed by defining the relationships of the classes. The terminology was developed by linking terms to the concepts of the obesity ontology and by identifying synonyms for each concept.

The coverage of the ontology developed in this study was evaluated by mapping the concepts and terms of the ontology with those extracted from the obesity-related Twitter postings. We performed a Twitter search using the combinations of keywords related to obesity (e.g., ‘diet,’ ‘exercise,’ ‘sleep,’ ‘smoking,’ and ‘drinking’) from January to March 2015 and selected the postings related to obesity. After converting them into a text file, keywords related to obesity were extracted by natural language processing using the KoNLP package in the R software. The frequencies of the extracted keywords were tallied, and any extracted keywords that were not included in the ontology were added to the revised obesity ontology.

The structure and representative ability of the revised obesity ontology were evaluated by nurse experts using seven items and ten items ontology evaluation criteria on a 5-point Likert-type scale, respectively, developed by Kehagias et al. [24]. We also analyzed the content validity index (CVI) by dividing the number of experts responding to 4 or 5 points by the total number of experts for each evaluation criteria. Three experts who had a master's or higher degree in nursing informatics and one expert who had a doctoral degree in adult health nursing with experience in education and research related to obesity participated in the evaluation. The experts were asked to assess the structure of the entire ontology and the representative ability of the superclasses of the ontology.

We extracted 1,712 terms from the clinical practice guidelines on obesity published by NICE [22] and KSSO [23] and 396 additional terms from social media postings on obesity.

We defined the class hierarchy by creating classes for the general concepts of ‘risk factors,’ ‘sociodemographic characteristics,’ ‘assessment,’ ‘diagnosis,’ ‘types,’ ‘symptoms and complications,’ ‘treatment,’ and ‘prevention.’ We then divided these classes into specialized subclasses. We combined ‘treatment’ and ‘prevention’ concepts into the ‘management strategies’ superclass. To address the concept of the environment where assessments, diagnosis, and management strategies for obesity are conducted, the concept of ‘settings’ was added as a superclass, and ‘sociodemographic characteristics’ was moved down to a subordinate concept for ‘risk factors.’ Table 2 lists the definitions of the superclasses, depths, and number of subordinate classes. The final obesity ontology used for the collecting and analyzing obesity-related social big data comprised 7 superclasses and 148 subclasses. We present the superclasses and subclasses down to the 4th level under each superclass in Figure 2.

We defined the properties and their facets describing the internal structure of each concept, an example of which is presented in Table 3. On average, 4.0 properties and 2.0 relations were defined within each class.

In total, 520 postings were retrieved from Twitter between January and March 2015. After excluding 79 duplicated postings, we selected 441 obesity-related postings for coverage evaluation. We manually reviewed 2,981 words extracted by natural language processing and confirmed the presence of 452 obesity-relevant terms.

Examples of terms that appeared at higher frequencies were ‘obesity’ (n = 191, 43.3%), ‘exercise’ (n = 134, 30.4%), ‘diet’ (n = 50, 11.3%), ‘lower-body obesity’ (n = 41, 9.1%), ‘fat’ (n = 37, 8.4%), and ‘treatment’ (n = 35, 7.9%). The obesity ontology developed in this study included all of these terms.

The 452 extracted terms were mapped to the classes (n = 85), properties (n = 64), and values (n = 380) with a 100% coverage rate for the concepts. However, our ontology included 397 of these 452 obesity-related terms, giving it a coverage rate of 87.8%. The 55 terms not mapped were synonyms for concepts of the classes (n = 3), properties (n = 5), and values (n = 47) of our ontology. These were mostly terms representing values of the ‘management strategies’ superclass in our ontology, such as various types of exercise (e.g., ‘dead lifts,’ ‘burpees,’ and ‘push-ups’). We revised the obesity ontology by adding these terms.

Tables 4 and 5 list the evaluation scores for the structure and representability of the ontology, respectively. The experts rated all seven criteria of the ontology structure at 4.0 or higher (Table 4). The criterion with the lowest score was the relation between concepts, and an expert pointed out that the relationships between the ‘settings’ class and other classes were not clearly represented. We therefore revised the ontology by redefining the relationships between the ‘settings’ superclass with the ‘assessment,’ ‘diagnosis,’ and ‘management strategies’ superclasses.

All ten criteria for representability scored higher than 4.0 (Table 5). The criterion with the lowest score was ‘accuracy’ at 4.3, which indicates how well the ontology included the terms representing obesity-related concepts on social media. The scores for these superclasses were probably lower because these concepts included many medical terms representing drugs, diseases, or medical treatment interventions. In particular, the experts rated the ‘symptom and complications’ superclass as 4.0. One expert suggested dividing the ‘symptom and complications’ superclass into two separate superclasses because these are very different concepts. Based on this suggestion we revised the ontology with 8 superclasses and 124 subclasses (Figure 3). The terminology of the revised ontology comprised 278 terms for classes (including 162 synonyms), 147 terms for properties (including 3 synonyms), and 1,737 terms for values (including 865 synonyms).

The ranks of the frequently used keywords on types of obesity and management strategies for obesity are presented in Table 6. High-frequency terms related to types of obesity were ‘abdominal fat’ (n = 16,830), ‘thigh’ (n = 15,931), and ‘lower-body obesity’ (n = 13,844). ‘Fat’ (n = 36,821) was the most frequently mentioned term related to diet. Terms related diet methods, such as ‘one-food diet,’ ‘Danish diet,’ or ‘Atkins diet’ were also mentioned frequently as diet terms. ‘Stretching’ (n = 8,469), ‘aerobic exercise’ (n = 8,387), ‘walking’ (n = 7,544), and ‘yoga’ (n = 4,779) were the high-frequency exercise terms involving activities that could be performed without any restriction of equipment or place. The terms ‘shakes’ (n = 7,807), ‘Herbalife’ (n = 7,165), and ‘Chinese medicine’ (n = 4,009) were the top-three keywords related to drug or diet supplement therapy.

The results of the sentiment analysis for obesity and diet are shown in Figure 4. Neutral comments were the most common (66%), followed by positive comments (22%) and negative comments (12%) in the 1,441,939 obesity and diet related postings. The examples of the keywords from positive comments were ‘to prevent obesity’ and ‘to manage obesity,’ ‘successful diet’ and ‘healthy diet,’ and ‘helpful exercise’ and ‘effective exercise.’ The examples of the keywords from negative comments were ‘skinny fat’ and ‘super-obesity,’ ‘failed diet’ and ‘excessive diet,’ and ‘exhausted from exercise’ and ‘no time for exercise.’ It was found that the positive perception tend to increase annually during the 3-year analysis period.