Content Characteristics of Knowledge Integration in the eHealth Field: An Analysis Based on Citation Contexts

We selected two leading journals in the eHealth field, Journal of Medical Internet Research (JMIR) and JMIR mHealth and uHealth (JMU), as our data sources. Our reasons are threefold. First, according to an expert survey of 398 active e-health researchers, JMIR and JMU were ranked as top A+ and top A journals out of 63 peer-reviewed eHealth related journals, respectively (Serenko, Dohan, & Tan, 2017). Second, JMIR was established in 1999, when the eHealth field was just emerging (Della Mea, 2001). This could provide us with a comprehensive understanding about the formation and evolution of the eHealth field. JMU is a newer spin-off journal of JMIR, focusing on more technical and developmental papers than JMIR. It covers more frontier scientific and technological contents in the eHealth field. Third, both JMIR and JMU provide open access articles in XML format. Since we aim at investigating the content characteristics of knowledge integration through citation context analysis, the availability of full text articles is helpful for us to obtain citation contexts. Other journals in the eHealth field often provide PDF-format articles, which require heavy and error-prone text processing to obtain the text content of articles (Bertin et al., 2016).

We collected all papers published by the two journals from 1999 to 2018, and selected 3,221 articles with the type of “original papers”, “reviews”, and “viewpoints”. Other types of articles, such as “Corrigenda and Addenda”, “Editorial”, and “Letter to the Editor”, which list fewer references, were excluded.

For each article, we parsed the metadata (DOI, publish year, etc.), bibliography information (title, PMID, journal, publish year, etc.), and citation contexts. The context of a citation in this study is defined as the sentence where the citation occurs rather than a longer text span so that the association between the citation context and its corresponding reference will be closer (Small, Tseng, & Patekc, 2017).

We augmented the metadata information (abstract, keyword, Keyword Plus, MeSH term) of the references by linking them to Web of Science (WoS) and PubMed. The disciplines of the references were determined as the WoS subject categories of the journal where it was published. The references without WoS subject categories were not analyzed.

In total, 119,598 citation sentences were obtained, as well as 101,751 reference records (i.e. bibliographic items) with metadata information, which account for 93.00% of all journal references and 72.38% of all references.

Most previous studies used expert knowledge to identify cited objects in citation sentences by human annotation, which were then applied to investigate the domain knowledge used in interdisciplinary research (Wang & Zhang, 2018). In this study, we used an automatic approach proposed in our previous study (Mao, Wang, & Shang, 2020) to identify associated knowledge phrases (AKPs), which can be regarded as explicit integrated knowledge content spread from references to citing papers.

The approach extracts noun phrases from citation sentences as well as titles and abstracts of references by using spaCy, an open-source natural language processing package. Several pre-processing operations were performed before the noun phrases from the two sources were matched. Single characters and the phrases starting or ending with numbers were removed. Author keywords, Keyword Plus terms, and MeSH (Medical Subject Headings) terms in the references are also treated as noun phrases of references. All phrases from the two sources were lemmatized using the NLTK Python package. Next, the noun phrases appearing in each pair of citation sentence and the corresponding reference were compared by our stem-matching approach. The noun phrases between the pair were matched if their stemmed forms were the same. We also matched the stemmed noun phrases extracted from the citation sentence with the stemmed sentences in the corresponding reference (including its title and abstract). Then, we denote the matched noun phrases of the citation sentence as the AKPs. This method recalled 78.57% phrases (209 of all 266 phrases) according to the evaluation on a randomly sampled 100 citation sentences. A total of 246,167 AKPs were extracted from our dataset, with 25,764 distinct ones.

To characterize the knowledge integrated by the interdisciplinary field, we designed a knowledge classification schema to categorize the identified AKPs. Recently, a few studies have attempted to discern the functions of knowledge played in a domain. Ding et al. (2013) pointed out that scientific papers embed many types of micro-level entities, including datasets, methods, and domain-specific entities. Heffernan and Teufel (2018) focusd on the identification of problems and solutions in scientific text. Lu et al. (2019) proposed a classification schema for author selected keywords, reflecting how they function semantically in scientific manuscripts. To favor the investigation of micro-level knowledge integration relationships, we also designed a knowledge classification schema based on the functions of knowledge in scientific articles.

We recruited two graduate students to annotate the types of all distinct AKPs based on the knowledge classification schema in Table 1. Each distinct AKP and one of its citation sentences that was randomly selected were given for the coders. Some examples are given in Table 2. First, two coders independently annotated 500 identical randomly selected knowledge phrases for pre-annotation. However, the kappa coefficient between the annotation of two coders was only 0.65. Therefore, an expert in the eHealth field was invited to guide the annotation work and helped the coders to distinguish the ambiguous cases. We found that some phrases could be labeled into different categories in different contexts. To avoid ambiguity, we only considered the frequently used meaning of the term in our annotation process. After discussion, two coders reached a consensus. Then, they independently annotated all 24,132 unique phrases that are associated with the disciplines of our interests. During the annotation process, two coders kept in communication with each other to reach an agreement. Among all 24,132 distinct phrases annotated in our previous study (Mao, Wang, & Shang, 2020), 24,063 distinct phrases were related to the WoS subject categories of this study’s interest, and another 1,701 distinct AKPs from the remaining references were annotated by the two coders in the same way for this study.

We introduce several indicators to measure the integration characteristics of different types of knowledge based on the identified AKPs. The indicators are defined as follows:

Knowledge amount: the number of AKPs.

To further understand the relationship of different knowledge in the integration process, we also analyzed the co-occurrence of different types of knowledge in the same citation contexts.

The descriptive information of our dataset is shown in Table 3. From the dataset, 119,598 citation sentences and 101,751 references with metadata information were extracted. Since a citation sentence may contain more than one in-text citation (Small, Tseng, & Patekc, 2017), the number of in-text citations (199,461) exceeds the number of citation sentences. In total, we obtained 246,167 AKPs with 25,764 distinct ones.

The annotation results of AKPs classification are shown in Table 4. The number of references and source disciplines, as well as knowledge integration density and average citation interval, are presented for each knowledge type. It is observed that the knowledge amount for different knowledge types is uneven. The phrases in the category of Research Subject are the most, followed by Others. The category of Theory contains the fewest AKPs, however, the knowledge integration density of Theory exceeds that of most other knowledge types, ranking the second place among all knowledge types. This indicates that Theory related references may carry more phrases of theories in each citation.

The average citation interval shows that different knowledge types have significantly different time lags. As Table 4 presents, Theory related phrases have the longest time lag in the knowledge integration, followed by Research Methodology, while Technology has the shortest time lag. This result could be explained by that theory and methodology need more time to be verified by the scientific community, while technology is updated rapidly.

We next turn our attention to the source disciplines of each type of AKPs. In this paper, we defined the source disciplines of AKPs as the WoS subject categories of the references carrying the AKPs.

Table 5 illustrates the top 10 highly contributed disciplines with the largest number of AKPs for each knowledge type. Overall, except Theory, Health Care Sciences & Services is the largest knowledge provider, followed by Medical Informatics. Nonetheless, the top 10 highly contributed disciplines rank significantly different among the knowledge types. Medical, healthcare, and psychology related disciplines provided the eHealth field with more knowledge about Research Subject, Entity, and Research Methodology, while for Technology and Data, information and computer science related disciplines contributed more. Psychology and management related disciplines supplied the eHealth field with more AKPs of Theory. This demonstrates that different disciplines may play different roles in the formation of the interdisciplinary field of eHealth according to their contributions in different knowledge types.

In this section, we present the integration characteristics in terms of the proposed indicators.

3.4.1

Knowledge amount

Fig. 1 displays the knowledge amount of each knowledge type over time. For every type, the number of AKPs remained stable before 2010 and has been rising since then. This trend is along with the increasing publication tendency of the eHealth papers (Fig. 1a), which reveals the emergence of the eHealth field in recent years. It appears that the category of Research Subject has grown the fastest, followed by Entity and Technology, while Theory has grown the slowest. It shows the abundance of research subjects in the interdisciplinary field of eHealth. The highly cited research subjects include “information”, “intervention”, “depression”, “physical activity”, “health”, “diabetes”, etc. These research subjects reflect the research hotspots in the eHealth field from the citation content perspective.

Figure 1

To deeply understand the patterns of different knowledge categories, we further analyzed the proportion of each knowledge type in each year, as shown in Fig. 1b. It is observed that the proportion of every knowledge type has gradually remained stable after the fluctuations in the early years. As the knowledge structure of the eHealth field has been formed over time, the integration pattern of different knowledge types has become relatively fixed. Besides, Technology was gradually surpassed by Entity, which shows that human beings and related organizations are highly involved in the field.

3.4.2

Number of references

As Fig. 2 presents, similar to the growing trend of knowledge amount, the number of references remained stable before 2010 and has been increasing afterward. For the proportion of references (Fig. 2b), it also shows a similar pattern to the knowledge amount, which remained stable in later years after the fluctuations in early years. This further proves the integration patterns of different types of knowledge have gradually remained stable in recent years.

Figure 2

3.4.3

Number of source disciplines

The number of source disciplines involved by each type of AKPs has continued to grow dramatically since 1999, as shown in Fig. 3a, which demonstrates the increase of interdisciplinarity in the eHealth field. The proportion of distinct source disciplines for each knowledge type also shows an upward trend, and the growth rate has slowed down recently.

Figure 3

3.4.4

Citation interval

Fig. 4 presents the average citation interval of AKPs, which represents the time lag that eHealth integrates these types of knowledge. Overall, the citation interval of every knowledge type increased steadily with the development of the field. This may be due to that some classic publications of pioneering research work in the field would increase the citations in the following years (Sun & Latora, 2020). As a result, the average citation age would increase over time. On the other hand, as shown before (Fig. 3a), the interdisciplinary character of the eHealth field has been rising over time. Since the cross-disciplinary knowledge flow often has a longer time lag (Rinia et al., 2001), the citation intervals between cited papers from other disciplines and citing papers in the eHealth field would also increase with the rise of interdisciplinarity.

Figure 4

We notice that there were no Theory related AKPs in some early years, therefore, the curve of Theory is not continuous. It may be driven by several reasons. First, the early studies in the eHealth field were more focused on the application of information technology to assist the information acquisition process of medical workers but were concerned less about the theory of interaction between humans and technology. Second, the definition of Theory in the present study is very narrow as we only included the phrases with specific theory names due to the operability of annotation. Finally, we only used the metadata of references to do the matching process. However, some references from the early years may not have recorded abstract or the theory related information was not covered in the metadata, which prevents us from annotating the AKPs of theory.

Moreover, we observe that the curve of Theory in Fig. 4 has fluctuated during the period. The rapid increase from 2008 to 2010 may be attributed to the rapid growth of the publications in the period, and they cited a few classical theory models (e.g. “social cognitive theory”) which were proposed in the early years. On the other hand, the theories cited by the eHealth field covered both relatively new information technology theories (e.g. “sensor acceptance model”) and classic cognitive theories (e.g. “social cognitive theory”). Therefore, the curve of the Theory has fluctuated during the later years. For Research Methodology, it shows a relatively long rise before 2007. At the moment, eHealth research absorbed some traditional psychology questionnaires (e.g. “SCL90R”, “CES D”). Then, it experiences a falling interval between 2007 and 2010. In this period, some novel data analysis approaches (e.g. “text mining”, “natural language processing”, “thematic analysis”) were introduced into the eHealth field. As the development of the eHealth field, more and more psychology questionnaires were used to assist the eHealth research, thus, the citation interval was increased again and gradually remained stable.

We further analyze the co-occurrence pattern of knowledge types within citation contexts to disclose their interactions in the knowledge integration process, as shown in Fig. 5. The ratio value in the figure is calculated as twice the co-occurrence frequency divided by the total frequency of the two knowledge types. It is clear that the most frequent pair of knowledge types is Research Subject and Research Subject, followed by Research Subject and Entity, then Research Subject and Technology. It is reasonable because authors often need to describe research subjects related information when citing the references, and it demonstrates Entity and Technology are two types of knowledge that are often integrated across different research topics. However, the co-occurrence of Theory and Data is the fewest. This may be due to the fewest total number of theory related knowledge. We also observe an interesting finding that the cells along with the diagonal line exhibit a relatively high ratio value. This phenomenon may be driven by that when we cite a knowledge entity (e.g. a methodology or a theory), we usually compare it with other similar types of entities. For example, in our dataset, “TAM” theory is frequently occurred with “TPB” theory.

Figure 5