Ontology development was performed based on previous reports by Noy and McGuinness [16] and Jung et al [17].

First, the domain and scope of the ontology were determined using the following competency questions: (1) What types of cancer are mentioned in posts on social media? (2) Which care delivery processes (eg, prevention, diagnosis, treatment) of cancer in general and specific cancer types are mentioned in social media posts? (3) What cancer-related topics are mentioned in posts for cancer in general and specific cancer types? (4) What emotions, which mean a range of feelings a patient with cancer can experience when dealing with cancer, are mentioned in posts for cancer in general and specific cancer types?

The purpose of the ontology was determined as collecting and analyzing social media data to identify cancer information needs and emotions related to cancer.

Second, existing ontologies on cancer were identified: the Liver Cancer Ontology [13], Breast Cancer Ontology [14], and Gastric Cancer Ontology [15]. Each is limited to a specific type of cancer, and none includes consumer terms. Therefore, a new ontology was developed to include various types of cancer and consumer terms.

An existing ontology on emotion was also identified—the Sentiment Ontology for Social Web [18]. This ontology has top-level classes of emotion as positive, neutral, or negative. These top-level classes were too broad, and the second-level class was too detailed to describe a range of feelings a patient with cancer can experience when dealing with cancer. Therefore, a new ontology reflecting emotions accompanying cancer was deemed necessary.

Third, terms extracted from the 3 existing cancer-related ontologies, the Sentiment Ontology for Social Web, cancer information portals, and social media posts related to cancer were enumerated.

The cancer information portals that were reviewed to extract terms were 2 US websites (the National Cancer Institute [19] and American Cancer Society [20]), 1 UK website (Cancer Research UK [21]), and 2 Korean websites (the National Cancer Information Center [22] and National Health Information Portal by Korea Centers for Disease Control and Prevention [23]).

These portals included information on emotions accompanying cancer and how to manage them. Emotions included across portals were anger, guilt, and depression [19-22]. In addition, the Cancer Research UK website [21] included overwhelmed, denial, anxiety, fear, and sadness; the National Cancer Institute [19] included overwhelmed, denial, anxiety, fear, and sadness, loneliness, hope, and gratitude; and the Korean National Cancer Information Center [22] included overwhelmed, denial, anxiety, hope, and gratitude.

Natural language processing (NLP) was used to extract consumer terms from social media data on cancer. Terms with the same meaning as those extracted from existing cancer-related ontologies and cancer information portals or terms with a new meaning related to cancer were collected as consumer terms. These included (1) heteronyms such as “jol-eob” (graduation in English, meaning complete cure); (2) abbreviations such as “jaegeom” (an abbreviation of jaegeomsa, meaning re-test in English), “chompa” (an abbreviation of cho-eumpa, meaning ultrasound in English), “holmon” (an abbreviation of holeumon, meaning hormone in English); and (3) terms used for herbal medicine and complementary therapies.

Fourth, the classes and their hierarchy and relationships were defined. The collected terms were grouped according to semantic meaning and determined concepts as classes with an independent existence. Hierarchies of the classes were designed based on the relationships of the concepts. The superclass and subclass concepts of the ontology were determined by analyzing the structures of the ontology. A list of synonyms was compiled for each class concept as a terminology presenting the relationship between the concept and synonyms in the ontology.

Nine domains of cancer-related emotions were identified: overwhelmed, denial, anger, fear and anxiety, sadness and depression, guilt, loneliness, hope, and gratitude. Each domain was defined as a class concept, and a list of synonyms of the class concepts was mapped as a terminology.

Fifth, the structure, correctness, and quality of the ontology were evaluated using the evaluation tool described below and by interviewing 3 domain experts: 2 professors of family medicine and 1 professor of bioinformatics.

The tool consisted of 13 items selected from the studies of Hlomani and Stacey [24] and Kehagias et al [25]. The items were scored on a 5-point scale for structure (size, depth of hierarchy, breadth of hierarchy, balance, overall complexity, and connectivity between concepts), correctness (accuracy, completeness, conciseness, and consistency), and quality (computational efficiency, adaptability, and clarity). Interviews included open-ended questions that allowed experts to recommend revisions of the ontology. The ontology was revised based on the results of the evaluation.

Based on the existing cancer-related ontologies and cancer information portals, 10 superclasses were identified: cancer type, prevention, diagnosis, treatment, prognosis (including recurrence and cure), risk factor, symptom, side effect, dealing with cancer, and emotion. The ontology consisted of concepts that represented care delivery processes and patient outcomes, such as prevention, diagnosis, treatment, and prognosis, and another set of concepts that represented how consumers managed, felt, perceived, and acted in their personal lives, such as risk factor, symptom, side effect, dealing with cancer, and emotion.

The average scores by the 3 experts of 13 items designed to evaluate the structure, correctness, and quality of the ontology ranged from 4.33 to 5 on the 5-point scale. The score for correctness was the highest (mean 4.83), followed by structure (mean 4.67) and quality (mean 4.67). A suggestion made by one of the experts was to combine the superclass of side effect with that of symptom, moving cancer symptoms and treatment side effects into subclasses of the symptom superclass, because it is difficult to distinguish side effects from symptoms without context. Another suggestion was to add the national cancer support system as a subclass of the dealing with cancer superclass. The ontology was revised to reflect the comments made by the experts.

The revised ontology had 9 superclasses, 213 class concepts, and 4061 synonyms. It had 3 to 4 levels of hierarchy, with 36 first-level subclasses and 41 second-level subclasses (Figure 1).

In this study, an ontology containing consumer terms was developed to collect and analyze social media data to identify health information needs and emotions related to cancer. The ontology has the following characteristics.

First, compared to other cancer ontologies, this ontology covers more cancer types and comprehensive topics, including the care process describing the interactions of health care providers and patients, risk factors that consumers have, emotions that consumers feel related to cancer, symptoms that consumers perceive, and the lifestyles that consumers lead.

Second, the ontology has a terminology component that presents consumer terms to analyze social media posts about cancer, including synonyms, heteronyms, and abbreviated expressions.

A total of 754,744 cancer-related posts on social media were collected. The 6-month collection of 2017 posts was less than half that of the full-year collection of the previous year. Monthly analysis showed cancer-related social media posts were more frequent in the second half of the year than in the first half of the year. In particular, the number of posts in February 2014 was almost twice the monthly average, which was the same as the search results of Google Trends. According to Google Trends, the cancer diagnosis and death from cancer of 2 Korean celebrities had been announced at this time.

Social media data were analyzed using NLP with ontology concepts and terms to identify consumers’ information needs. The frequencies of topics in social media data may indirectly reflect consumers’ information needs, as they frequently post on topics in which they are interested or about which they have concerns [8].

Thyroid cancer, which ranked in the top 3 in the national cancer statistics, was not included in the top-ranked cancer types on social media. This finding is similar to that of Buis and Whitten [26], showing that the information needs related to cancers with a low survival rate are higher than those for cancers with a high survival rate. The 5-year survival rate for thyroid cancer in Korea was about 100% in 2013-2017, compared to 70.4% for all types of cancer [22].

Frequency analysis by superclass revealed the highest frequency of posts on social media related to risk factor (320,568/754,744, 42.47%), followed by emotion (254,920/754,744, 33.78%), symptom (243,010/754,744, 32.20%), and treatment (227,942/754,744, 30.20%). These findings can be compared to those of 2 previous studies: Lu et al [7] clustered posts of patients with lung and breast cancer and extracted symptoms, examinations, and treatments (drugs, procedures) as dominant topics. Cho et al [1] performed a qualitative content analysis of Q & A posts of patients with breast cancer, and extracted treatment, physical condition, and lifestyle/self-care as dominant topics. They also found that 75% of the information requests included expressions of emotion [1]. Although these previous studies [1,7] focused mainly on the stages after diagnosis, such as treatment, examinations, and physical condition, our results revealed that the majority of consumers’ information needs involved risk factors. Also, previous studies [1,7] used data collected from cancer community posts by patients with specific cancer types, whereas this study used data collected from social media posts by the general public. Social media data may include concerns of the public about cancer risk factors. In addition, as the ontology developed in this study covered comprehensive topics, including risk factors and emotions, this study had a more diverse focus than those of previous studies [1,7].

Consistent with the findings of the 2 previous studies [1,9], the most common emotions related to the top 10 cancers in this study were fear/anxiety, hope, and sadness/depression. Cho et al [1] reported that anxiety/worry, gratitude, fear, and sadness were frequent in posts by Korean women with breast cancer, and Freedman et al [9] reported that fears, anxiety, denial, and depression were frequent emotions cited on treatment in posts by patients with breast cancer. However, among positive emotions, gratitude was included in the top 3 in the study by Cho et al [1], whereas hope was ranked in the top 3 in our study. The question and answer board in Cho et al's study [1] included more posts on gratitude toward health care providers, in contrast to social media posts in this study that often included posts on hope for a cure by patients and their families.

Frequency analysis by the end-node class level indicated that dealing with diet ranked among the top-class concepts for most types of cancer. This finding was consistent with those of previous information-need studies [1,27] in Korean patients with breast cancer. Cho et al [1] reported that patients asked many questions related to diet, and Kim and Hur [27] reported high information-need scores for diet. These findings reflect the importance of diet in disease management that is perceived by the Korean population. Concepts and terms such as food, protein, fruit, vitamin, meal, vegetable, and nutrition were top ranked in the posts collected in this study, demonstrating consumers’ information needs related to a healthy diet.

This study also compared the frequencies of superclass concepts and end-node class concepts by 10 types of cancer.

Regarding the care process of cancer, treatment was the most frequently mentioned for all types of cancer, except cervical cancer. Diagnosis and prevention appeared more frequently in posts on cervical cancer. In Korea, the national cancer screening program for cervical cancer is recommended from 20 years of age. Whereas breast and gastric cancer screenings are recommended from 40 years of age, and colorectal cancer screening is recommended from 50 years of age. As the opportunity to become interested in screening for cervical cancer comes earlier than for other cancers, the active use of social media by young women may have resulted in greater numbers of posts related to diagnosis and prevention.

The frequencies of risk factor–related class concepts, especially diet, chemical, tobacco, obesity, and infectious agent, differed between cancer types. Previous research findings have indicated that differences in post frequency on cancer topics reflect the different information needs for each cancer type [7,8]. Therefore, tailored information on risk factors should be provided according to consumers’ needs by cancer type. In addition, tailored information on cancer can be provided according to the consumers’ specific risk factors.

Posts on leukemia contained a higher rate of feelings of guilt. According to the National Cancer Information Center [22] data, leukemia has a higher incidence in children than those of other types of cancer. In many cases, posts related to leukemia were likely written by the parents of patients rather than by the patients themselves. This finding suggests that emotional management is necessary for not only patients with cancer but also their family members and friends.

Based on these findings, we suggest what information should be provided and how it can be provided. These suggestions would aid information providers, namely clinicians and portals operated by government or professional societies, to improve care for patients with cancer by providing relevant information based on consumers’ information needs.

First, it is necessary to ensure that sufficient information on risk factors is provided to the public. Not only do information needs increase after cancer is diagnosed, but they are also high for risk factor management. Thus, information on risk factors should be provided depending on the concerns of consumers.

Second, the high consumers’ information needs on healthy diets were noteworthy. Qualified information on diet should be provided to patients with cancer. Collaboration with a nutritionist would be effective in providing tailored nutritional information for each cancer type according to the needs of patients.

Third, in general, most information portals provide information through the same organization and the same flow of information for all types of cancer. Tailored information can be provided according to cancer type and the characteristics of the consumer, such as age, gender, and risk exposure. It is possible to make it easier to access the information that consumers want using keyword visualization or navigation. Different navigation routes could be applied according to cancer type. In addition, applying different types of visualization could improve the convenience of consumers before or after the cancer diagnosis.

Finally, providing information and emotional support are not separate but, instead, coexist. It is necessary to provide reliable information and management for emotional care so that people do not rely on only family and caregivers for emotional support. One possible approach is to combine the functions of online support groups with information portals to provide emotional support.

This study had some limitations. Only the top 10 cancer types were analyzed. Future studies should analyze data on other types of cancer. The social media posts were made by not only patients with cancer but also caregivers and the general public. However, these populations could not be distinguished because identifying information of the consumers was not collected. Further research is needed to collect social media data with an identification algorithm to distinguish the status of the authors of the posts and to provide tailored information.

This study was performed to develop a cancer ontology with terminology containing consumer terms to collect and analyze social media data. The ontology consisted of 9 superclasses (cancer type, prevention, diagnosis, treatment, prognosis, risk factor, symptom, dealing with cancer, and emotion), 213 classes, and 4061 synonyms with consumer-generated terms. It used 9 emotional classes (overwhelmed, denial, anger, fear and anxiety, sadness and depression, guilt, loneliness, hope, and gratitude) to investigate emotional status in the social media data on cancer.

This ontology, containing comprehensive cancer-related topics, enabled identification and comparison of consumer interests and concerns about risk factors, dealing with cancer, and emotions as well as the care process in social media data. The results of this study showed that information needs and emotions differ according to cancer type. These observations could be used to provide tailored information to consumers according to the cancer type and care process. Care for patients with cancer can be improved by providing relevant information based on consumers' information needs.