A comparative quantitative study of utilizing artificial intelligence on electronic health records in the USA and China during 2008–2017

Background The application of artificial intelligence techniques for processing electronic health records data plays increasingly significant role in advancing clinical decision support. This study conducts a quantitative comparison on the research of utilizing artificial intelligence on electronic health records between the USA and China to discovery their research similarities and differences. Methods Publications from both Web of Science and PubMed are retrieved to explore the research status and academic performances of the two countries quantitatively. Bibliometrics, geographic visualization, collaboration degree calculation, social network analysis, latent dirichlet allocation, and affinity propagation clustering are applied to analyze research quantity, collaboration relations, and hot research topics. Results There are 1031 publications from the USA and 173 publications from China during 2008–2017 period. The annual numbers of publications from the USA and China increase polynomially. JAMIA with 135 publications and JBI with 13 publications are the top prolific journals for the USA and China, respectively. Harvard University with 101 publications and Zhejiang University with 12 publications are the top prolific affiliations for the USA and China, respectively. Massachusetts is the most prolific region with 211 publications for the USA, while for China, Taiwan is the top 1 with 47 publications. China has relatively higher institutional and international collaborations. Nine main research areas for the USA are identified, differentiating 7 for China. Conclusions There is a steadily growing presence and increasing visibility of utilizing artificial intelligence on electronic health records for the USA and China over the years. The results of the study demonstrate the research similarities and differences, as well as strengths and weaknesses of the two countries.


Background
With the expanding use and increasing possibility of including information relating to patient outcomes and functionality such as clinical decision support, Electronic Health Records (EHRs) becomes increasingly valuable information about patient health conditions and responses to treatment over time [1]. The field of utilizing artificial intelligence techniques on EHRs data processing has attracted increasing interests from scientific community, reflected by the increasing of publications from major scientific literature databases such as Web of Science (WoS) and PubMed. The USA and China are top 2 largest economies in the world. According to literature retrieval in WoS, the two countries have the most publications the field in the last decade. Therefore, it is meaningful to conduct a quantitative analysis of the research publications from the two countries to compare their research similarities and differences, as well as strengths and weaknesses.
Research publication plays an important role in providing key linkage between knowledge generation, uptake and use in the scientific process [2]. Bibliometrics involves statistical analysis of written publications. It has been the method of choice for quantitative assessments of academic research to comprehensively explore the research advances in the past and identify future research trends in a specific field [3]. Bibliographic data from citation indexes, e.g., titles, journal, abstracts, author addresses, and etc., are analyzed statistically to recognize the popularity and impact of specific publications, authors, affiliations, or an entire field. Bibliometrics has been widely performed in the evaluation of various research areas [4,5]. Especially, it has also been adopted to the evolution of interdisciplinary research field, e.g., natural language processing in medical research [6], natural language processing empowered mobile computing research [7], technology enhanced language learning research [8], and text mining in medical research [9].
To that end, relevant publications in the field were retrieved from both WoS and PubMed to quantitatively explore the academic performances of the two countries in terms of current research status, research intellectual structures, and research focuses. Analyzing techniques include bibliometrics, geographic visualization, collaboration degree calculation, social network analysis, latent dirichlet allocation, and affinity propagation clustering.
Specifically, the following comparisons are conducted: 1) studying the quantitative distributions and growth characteristics of the publications, 2) identifying prolific publication sources, authors, and affiliations, 3) exploring publication geographical distributions, 4) investigating collaboration degrees and collaboration patterns, 5) visualizing scientific collaboration relations, and 6) discovering hot research topics and topic evolutions.

Data sources
The publications in the research field during 2008-2017 from WoS and PubMed databases were preferred. With a list of search keywords determined by domain experts, as shown in Table 1, publications with "Article" type were retrieved and downloaded as plain texts. After manual review, 1031 records from the USA and 173 records from China were obtained for comparison analysis. Key elements including title, publication year, keywords, abstract, author address were extracted. In addition, corresponding affiliations and regions were automatically extracted from author address information. Key words from author keywords, Keywords Plus/ PubMed MeSH, title, and abstract, were extracted by our developed natural language processing module.
In addition to basic bibliometric analysis, the techniques used in this paper include: geographic visualization, co-authorship index and collaboration degree calculation, social network analysis, and topic modelling analysis.

Geographic visualization analysis
Geographic visualization [10] refers to a set of visualization technologies for supporting geospatial data analysis. It provides ways to explore both the information display and the data behind the information itself to more readily view complex relations in images Table 1 Search keywords related to "artificial intelligence" and "EHR" Keywords related to "artificial intelligence" "artificial intelligence" OR "intelligent information processing" OR "machine learning" OR "pattern recognition" OR "information retrieval" OR "information extraction" OR "data mining" OR "text mining" OR "deep learning" OR "neural network" OR "natural language processing" OR "NLP" OR "semantic analysis" OR "question answering" OR "word sense disambiguation" OR "named entity recognition" OR "language modeling" OR "intelligent computing" OR "intelligent computation" OR "speech recognition" OR "smart learning" OR "knowledge graph" OR "automated reasoning" OR "automated inference" OR "knowledge representation" OR "fuzzy logic" OR "bayesian network" OR "machine intelligence" OR "natural language generation" OR "natural language understanding" OR "bayesian networks" OR "neural networks" OR "classification algorithm" OR "clustering algorithm" OR "association rule mining" Keywords related to "EHRs" " electronic medical record" OR "clinical notes" OR "clinical summary" OR "discharge summary" OR "EMR" OR "medical data" OR "electronic patient record" OR "medical record" OR "medical records" OR "electronic medical records" OR "electronic health record" OR "EHR" OR "electronic health records" OR "EHRs" OR "EMRs" OR "clinical note" OR "electronic patient records" OR "personal health record" [11,12]. Geographic visualization works essentially by helping people see the unseen more effectively in a visual environment than when using textual or numerical description. In this study, we apply geographic visualization analysis to explore publication geographical distributions in the USA and China, respectively.

Co-authorship index and collaboration degree
Co-authorship index shown as Eq. (1), was firstly elaborated by Schubert and Braun [13]. It is obtained by calculating proportionally the publications co-authored by single, two, multi-and mega-authors for different countries. Here, the publications have been firstly divided into four categories according to author count, i.e., single-author, two-author, multiple-author publications with three to four authors, and mega-author publications with five or more authors.
In the equation, N ij is the publication count co-authored by j authors in the i th country, N io is the publication count in the i th country, N oj is the publication count co-authored by j authors in all countries, N oo is the publication count in all countries. CAI = 100 represents the average level. CAI > 100 indicates higher than the average, while CAI < 100 reflects lower than the average. As a measure of scientific research's connective relation to the level of author, affiliation, or country, the collaboration degree can be calculated as Eq. (2) [14,15].
In the equation, C Ai indicates the collaboration degree of the i year in the author, affiliation or country level. α j donates the count of author, affiliation or country for each publication. N is the annual publication count.
In this study, co-authorship index is used to study collaboration patterns of authors, and collaboration degree is applied to measure the scientific research's connective relation to the three levels.

Social network analysis
Social network analysis (SNA) focuses on the structure of ties within, e.g., persons, organizations, or the products of human activity or cognition such as web sites [16]. SNA works based mainly on networks and graph theory [17], and it provides both a visual and a mathematical analysis of human relations. In this study, the collaboration relations for authors, affiliations and countries are explored using social network analysis. In the network, the nodes are specific authors, affiliations or countries, and the lines are the collaboration relations.
The size of node indicates the publication count of a specific author, affiliation or country. The width of link indicates the collaboration frequency between the two authors, affiliations or countries.

Topic modelling analysis
Topic modeling extracts semantic information from a collection of texts using statistical algorithms. Latent Dirichlet Allocation (LDA) is an improved three-layer Bayesian model developed by Blei et al. [18]. In LDA, each document in the text corpus is modeled as a set of draws from a mixture distribution over a set of hidden topics, where topics are assumed to be uncorrelated and each is characterized by a distribution over words. In LDA, a word is defined as an item from a vocabulary indexed by {1, …,V}, a document is a sequence of N words denoted by d = (w 1 , …, w N ), and a corpus is a collection of M documents denoted by D = {d 1 , …, d M }. The generation process is as follows: 1) The term distribution β indicating the probability of a word occurring in a given topic is as β~Dirichlet(δ); 2) The proportions θ of the topic distribution for a document d are determined by θ~Dirichlet(α); 3) A topic is chosen by the distribution z i~M ultinomial(θ) for each word w i in the document d, and a word is chosen from a multinomial probability distribution conditioned on the topic z i : p(w i | z i , β). As for variational expectation-maximization, the log-likelihood for one document d ∈ D is given by Eq. (3), and the likelihood for Gibbs sampling estimation with k topics is as Eq. (4).
Further, Affinity Propagation (AP) clustering is used for the cluster analysis of the topics identified by LDA. AP was proposed by Frey and Dueck [19] with a basis of message passing. It does not require users to set cluster count in advance, but considers all data points to be potential exemplars and transmits real-valued messages recursively until a set exemplars of high-quality emerges [20]. AP was found to identify clusters with lower error rate and less time [21]. AP calculates the "responsibility" r(i, k) and the "availability" a(i, k), shown as Eqs. (5) and (6) for each node i and each candidate exemplar k. r(i, k) is the suitableness of k as an exemplar for i, while a(i, k) is the evidence that i should choose k as an exemplar.
In the equations, s(i, k) is the similarity between two nodes i and k. When a good set of exemplars emerges, Eqs. (5) and (6) will stop iterating. Each node i can then be assigned to the exemplar k that maximizes a(i, k) + r(i, k). If i = k, then i is an exemplar. Numerical oscillations is controlled using a damping factor between 0 and 1.
In this study, words from author keywords and Keywords Plus/PubMed MeSH, publication title, as well as abstract with weights 0.4, 0.4, and 0.2 determined by our former study [6] are used as analysis units in topic modelling analysis. Term Frequency-Inverse Document Frequencies (TF-IDF) is used to filter out unimportant terms.

Growth of publications
The distributions of total publications by year for the USA and China are shown in Fig. 1

Prolific publication sources
The 1031 records from the USA are published in 347 unique journal or conference proceeding sources, and 92 publication sources contribute to China's 173 publications. The top 16 publication sources for the USA in Table 2

Prolific authors and affiliations
Three thousand three hundred fifty authors and 542 affiliations from the USA contribute to the 1031 publications, and 635 authors and 208 affiliations from China for the 173 publications. Table 3

Geographical distribution of publications
We study the concentration of researches in the USA and China at regional levels. The spatial characteristics of the publications from the two countries are explored. 46 states in the USA involve in the 1031 publications and 25 regions in China contribute to the 173 publications.

Authorship pattern and collaboration
The profiles of CAI for the USA and China have been illustrated in Fig. 4. It is clearly indicated that CAIs of multi-and mega-author publications in the research filed in China are slightly higher than the average. However, the CAIs of multi-and mega-author publications in the USA are lower than the average. Figure 5 shows the collaboration degrees at the country, affiliation and author levels in the two countries. On the whole, the international collaboration degree is growing relatively slowly than the author and affiliation collaboration degrees. On average, 5.83 authors, 2.63 affiliations and 1.18 countries participate in each publication from the USA. As for China, on average each publication has 5.79 authors, 2.84 affiliations and 1.39 countries. The average degrees of affiliation and country for China's publications are higher than that for the USA's publications, while the average degrees of author is on the contrary. The collaboration among countries/regions for the USA's publications is then visualized as Fig. 6 (access via the link [22]). From the figure, the USA (the largest node in blue color) in the center of the network has the most collaborations with other countries/regions. The USA-China collaboration (the thickest line) is ranked at 1st. The collaboration networks among affiliations with publications > = 15 (access via the link [23]) and among authors with publications > = 12 (access via the link [24]) are also visualized. Furthermore, we also visualize the collaborations for China's publications including country/region collaboration (access via the link [25]), collaboration among affiliations with publications > = 3 (access via the link [26]), and collaboration among authors with publications > = 3 (access via the link [27]). By accessing to the  the perplexities of models fitted using Gibbs sampling with different topic counts. The results suggest that the optimal topic count can be set to 35 for both the USA and China. The α is then set to 0.01339416 for the USA and 0.008163102 for China. We estimate the LDA models using Gibbs sampling with the parameters. Potential themes are assigned to each topic through semantics analysis of representative terms and text intention reviewing. Table 6 displays the top 5 best matching topics for the USA including Drug adverse event, Vaccine, Diabetes mellitus, Health data confidentiality, and Health data analysis technique, while the top 5 for China are Named entity recognition, Drug adverse   groups. For identifying emerging research topics, we firstly assign each publication to the topic with the highest posterior probability. We then explore the trends of research topics shown in Figs. 11 and 12. We also conduct Mann-Kendall test [28] to examine whether topics present increasing or decreasing trends.

Discussion
In this study, a comparative quantitative analysis of literature of utilizing artificial intelligence on electronic health records in the USA and China are conducted. This study identifies 1031 publications from the USA and 173 publications from China for the comparative analysis. Significant and polynomial increases in publication counts for both two countries can be found. This reflects a growing interest in the research field. However, the publication count of China is not at par with that of the USA, this can also be reflected by Tables 3 and 4, where the top prolific authors and affiliations of the USA own relatively more publications than that of China. Most prolific publication sources are journals, while only some are conferences such as AMIA Annual Symposium Proceedings, indicating a wide influence of journal in the research field. From the publication distributions in region levels, it is obvious that for both the USA and China, most top prolific regions are also of economic prosperity. From the authorship pattern analysis, it is found that publications published by scientists in the research field in China prefer to work in larger collaboration groups. This is consistent with the finding of Guan and Ma [29]    etc., present statistically significant increasing trends at the two-sided p = 0.05 level. The same is for 6 topics for China, including Named entity recognition, Risk event, Chinese medicine, Brain imaging, Drug adverse event, and Cancer. As an emerging focus in drug and cancer research topics, drug resistance has currently been one of the biggest obstacles in the treatment of cancers in clinical practice [30]. Some existing examples of cancer drug resistance research are as follows. Sun et al. [31] proposed a novel stochastic model connecting cellular mechanisms underlying cancer drug resistance to population-level patient survival for the examination of therapy-induced drug resistance and cancer metastasis. Sun and Hu [30] conducted a systematic review on the literature of mathematical modeling approaches and computational prediction methods for cancer drug resistance.
In this study, there are some limitations that are inherent to the database used and to search query developed by the authors. Such limitations were also encountered in the existing bibliometric studies, e.g., [32,33]. Firstly, despite the fact that WoS is a widely applied repository for bibliometric analysis and PubMed is an important data source on life sciences and biomedical topics, there are still unindexed conference proceedings and journal articles. Secondly, we treat publications of journal and conference types equally important in the analysis rather than bestowing weights for publications of different types. Furthermore, since no search query is 100% perfect, thus false positive and false negative results are always a possibility. In addition, the ranking of authors and affiliations in the study is based on data presented by WoS and PubMed. However, it is possible that some authors or affiliations might have different name spelling or more than one names, which might lead to an inaccuracy in the productivity of these authors or affiliations. Despite all these limitations, our study is the first to conduct a quantitative analysis of the research publications of utilizing artificial intelligence on electronic health records from the USA and China to compare Fig. 12 The trends of research topics for China's publications their research similarities and differences, as well as strengths and weaknesses. The findings of our study can potentially help relevant researchers, especially newcomers, understand and compare the research performance and recent development in the USA and China, especially, as well as optimize research topic decision to keep abreast of current research hotspots.

Conclusions
Utilizing artificial intelligence techniques on EHRs research is an emerging and promising field. This research provides a most up-to-date quantitative analysis for exploring and comparing the research performance and development trends of the research field from the USA and China during the period 2008-2017. Results of this exploration present a comprehensive overview and an intellectual structure of the research, especially, research topics, for the two countries in the last decade.