Trending Topics in Research on Rehabilitation Robots during the Last Two Decades: A Bibliometric Analysis

Abstract

Rehabilitation robots, as representative advanced modern rehabilitation devices, are automatically operated machines used for improving the motor functions of patients. Research on rehabilitation robots is typically multidisciplinary research involving technical engineering, clinical medicine, neural science, and other disciplines. Understanding the emerging trends and high-impact publications is important for providing an overview of rehabilitation robot research for interested researchers. Bibliometric analysis is the use of statistical methods to analyze publications over a period of time, which can provide visual insights into the relationships between studies and their publications. In this study, we used “rehabilitation robot*” as a topic term to collect 3527 papers from Web of Science in 127 subject categories published between 2000 and 2019. Rehabilitation robot research has increased rapidly over the past 20 years, 10 key clusters of which were analyzed in this narrative review: improving functional ability after stroke, spinal cord injury, universal haptic drive, robotic-assisted treadmill therapy, treadmill training, increasing productivity, custom-designed haptic training, physical treatment strategies, arm movement therapy, and rehabilitation robotics. Based on this database, we constructed co-citation and co-occurrence networks that were characterized by betweenness centrality values of more than 0.08 and citation bursts with strengths of more than 23, thereby visualizing the emerging trends in the research of rehabilitation robots.

1. Introduction

With improvements in technology and quality-of-life requirements, people who lose their motor functions are increasingly dependent on various medical devices in their daily lives. According to the WHO, one out of eight persons worldwide currently benefits or has benefited from rehabilitation devices [1,2]. Rehabilitation for relieving pain and spasticity, as well as improving walking ability, is always a challenge for clinicians. Rehabilitation robots, as representative advanced modern rehabilitation devices, are automatically operated machines used for improving the motor functions of patients with neurological diseases [3,4,5] or musculoskeletal damage [6,7]. Rehabilitation robots are the result of multidisciplinary research, including engineering, medicine, and other disciplines. Compared to manual therapies, robotic rehabilitation has several advantages: (1) it not only substitutes therapists’ repetitive tasks and boring labor for physical treatment but is also beneficial to reducing human resource consumption for rehabilitation units; (2) it provides personalized training for patients according to their different impairment levels, rehabilitation stages, and recovery processes; and (3) it allows quantitative evaluation and gives real-time feedback on the rehabilitation progress with the help of various types of sensors, such as motion capture, electromyography (EMG), and electroencephalogram (EEG). Research on rehabilitation robots can be traced back two decades or even earlier. More than 3500 academic papers can be found in the Web of Science (WoS) literature database for the years 2000–2019. Some introduced structural and mechanical designs (e.g., serial or parallel mechanism, end-effector or exoskeleton), mainly published in engineering journals [8,9,10,11,12]. Other articles focused on control strategies (e.g., passive, active, or resistance modes) and biomedical signals (e.g., EMG or EEG) applied to human–computer interactions [13]. Additionally, a large number of papers have been published reporting the application of robotic technology to rehabilitation for patients (e.g., stroke, spinal cord injury, traumatic brain injury, or cerebral palsy), mainly in clinical journals [14]. The research on rehabilitation robots involves multidisciplinary theories and technologies, including rehabilitation engineering and medicine, neuroengineering and science, biomechanical and mechanical engineering, control engineering, and computer science, posing a challenge for researchers. Literature reviews, which collect a large volume of relevant information on a certain field or a certain aspect of a research topic, can provide researchers with a comprehensive understanding of particular areas of research. There have been more than 200 review papers related to rehabilitation robots published over nearly two decades and indexed in Web of Science. Most existing review papers are from the reviewer’s perspective, which may be constrained by the subjective evaluation and cognitive approach of the reviewer. In addition, review papers may lack an overview of the whole research area and the relationships between different research fields in statistical analysis, which is extremely important for rehabilitation robot research due to its highly multidisciplinary nature. It is crucial to explore emerging trends [15,16,17,18] and identify breakthrough research related to the technology of rehabilitation robots. Understanding the dynamics of a research field is essential for scientists, analysts, and decision-makers to be able to identify emerging trends and sudden changes in the body of scientific knowledge [19]. The bibliometric method is a quantitative approach for evaluating the state of the field and identifying and tracking evolving trends, providing insights through the visual connections between studies [20].

Conventional bibliometric analysis is often used to evaluate research trends according to the publication outputs of countries, institutes, and journals, as well as the frequency of citations. At present, bibliometric analysis has been used to identify research topics and trends within single fields and between multiple research fields [21,22]. Swanson [23] and his colleagues pioneered a literature-based discovery approach to identify potentially valuable hypotheses and the development trajectory of the scientific field, composed of key turning points [24]. Bibliometrics is a method to observe the current state and future trends in scientific areas by analyzing the academic literature in a given research area. Compared with literature surveys, citation data are used as a bibliometric tool to evaluate and indicate the intellectual impact of the research outputs. Citations indicate the degree to which other scientists are reading and applying published works, or at least that the researchers are building on previous works [25]. In addition, the analysis of keywords has been used to identify hot topics and research trends in recent years [26,27]. However, to the best of our knowledge, no publication has reported research on rehabilitation robots using bibliometric methods. The aim of this work is to present the current state and emerging trends of rehabilitation robots through a bibliometric analysis of the literature in the field. CiteSpace software was employed to construct a document co-citation network and identify keyword co-occurrence. Structural metrics (i.e., betweenness centrality) and temporal metrics (i.e., citation bursts) were used to characterize the networks. By analyzing bibliometric indicators—such as citation frequency, centrality values, and citation bursts—emerging trends and new developments from 2000 to 2019 can be identified, allowing not only academics but also policymakers to recognize future research directions. The study will support researchers’ work by providing potentially relevant information behind a large body of academic publications.

2. Data and Methods

2.1. Data

The data used in this article are limited to scientific literature indexed by the Web of Science Core Collection, Science Citation Index Expanded (SCI-E), and Social Sciences Citation Index (SSCI). The WoS Core Collection is the most commonly used citation database whose data can be processed by a variety of bibliometric analysis software packages. The relevant articles were extracted using topic keywords from the database. There are two criteria commonly used for the analysis of scientometric search terms: recall, and precision. Recall refers to the ability to collect as many relevant documents as possible. Precision refers only to the selection of irrelevant records. In order to improve these two criteria, papers in SCI-E and SSCI containing the key term “rehabilitation robot *" were included. Because a single topic search may not cover thematically relevant literature if the topic does not appear in the titles, abstracts, or keywords, to retrieve highly relevant robot literature, a comprehensive strategy was designed to meet the requirements of data quality and data coverage. In this study, “rehabilitation robot” and related topics—including “rehabilitation robots”, “rehabilitation robotic”, and “rehabilitation robotics”—were searched as “rehabilitation robot *”. The document types were “article” and “review”. Some of the retrieved records included basic attributes such as publication time, author, institution, country, and reference citations. After data retrieval, cleaning, and deduplication, we finally obtained 3527 papers and 73,371 cited references in the field of rehabilitation robots published in the past 20 years. This dataset was used in the subsequent analysis.

2.2. Tools

The principal visual analysis tools are CiteSpace V.5.6.R5, HistCite, VOSviewer, SCI2, and UCINET. In this study, CiteSpace—a Java application program—was used to generate and analyze co-citation networks composed of co-cited references from the literature data or co-occurring keywords, subject categories, or noun phrases. CiteSpace was used to perform a co-occurrence and co-citation analysis of knowledge units in the literature. The data from the WoS Core collection can be used to directly analyze the structure, patterns, and distribution of scientific literature. CiteSpace is a visual analytic system for illustrating emerging trends and critical changes in the scientific literature and has been widely used in scientometric research work. It takes a set of bibliographic records as its input and models the intellectual structure of the underlying domain in terms of a synthesized network based on a time series of networks derived from each year’s publications [28,29]. CiteSpace identifies information bursts by detecting frequency changes in citations and noun terms to determine articles with high burst values. Burst detection is a class of algorithm to identify changes in a variable over a period of time with reference to others in the same population [30]. Furthermore, CiteSpace is a suitable tool to build and visualize knowledge map networks, which can be used in co-citation [31] and co-occurrence analyses.

2.3. Methods

In this paper, we used “rehabilitation robot*” as the topic term to collect data from the WoS database for bibliometric analysis. The retrieved literature records included publication years, countries/territories, institutions, journal sources, subject categories, co-cited references, and keywords. Two sets of scientometric open software—CiteSpace (V5) and VOSviewer (V1.6)—were used for data analysis and visualization. The document co-citation network and co-occurrence network were generated from the bibliographic data. In this paper, we used 6 parameters to construct the co-citation network and the co-occurrence network:

(1)

Time slicing: 2000–2019;

(2)

Years per slice: “1”;

(3)

Term source: title, abstract, author keywords, and keywords plus;

(4)

Node type: category, cited journal, reference, and keywords;

(5)

Selection criteria: the top 50 publications with the highest citation frequency in each time slice (top N = 50);

(6)

Pruning: pathfinder and pruning sliced networks (the functions provided by CiteSpace can reduce the number of connections while retaining the most prominent visualization).

A co-occurrence document analysis was used to explore the relationships between words in the documents. The frequency of word occurrence was assumed to reflect the association of the underlying themes. In this analysis, measurements of the relationships between the intellectual structures identified by the research papers allowed for correlations in multiplex networks, which consist of fixed sets of nodes connected by different types of links. A co-citation document analysis was adapted to complete the comparative analysis by detecting the intellectual structure and emerging rehabilitation robot topics from the selected bibliographic data. This analysis assumes that two documents that appear simultaneously in the reference list of the third related document characterize the structure of intellectual knowledge in terms of networks of co-cited references [32,33].

Co-occurrence analysis is a commonly used text content analysis technique that tends to explore the changes in professional topics in a research field based on the frequency of co-occurrence terms (i.e., words or noun phrases) that appear in the entire body of literature in the selected field [34,35]. Co-word cluster analysis and word frequency analysis are utilized to analyze hot topics and global research trends.

The intellectual structure and emerging trends were examined through overlay visualization of the selected bibliographic data on rehabilitation robots. Co-citation clusters were obtained by cluster algorithms from the synthesized network. The algorithms allow an analysis function of automatic label clustering in CiteSpace [36]. Therefore, all of the clusters were generated automatically, without using manual labeling of specialties or reading the publication titles. In this study, we constructed four networks: the subject category network, the keyword co-occurrence network, the document co-citation network, and the journal co-citation network. The betweenness centrality and the bursts of citations were used to characterize the networks. Betweenness centrality was used to identify pathways between different thematic clusters. It was computed for each node in the merged network based on a fast algorithm introduced by Brandes [37]. The betweenness centrality of a node v is defined as follows:

$$g\left(v\right)={\displaystyle \sum }_{s\ne v\ne t}{\sigma }_{st} \left(v\right)/{\sigma }_{st}$$

where ${\sigma }_{st}$ is the total number of shortest paths from node s to node t, and ${\sigma }_{st}\left(v\right)$ is the number of shortest paths from s to t going through v.

Betweenness centrality measures the extent to which the node is in the middle of a path that connects other nodes in the network [38]. A citation burst is a dramatic increase in the number of citations of an article in a short period of time, and can be taken to represent notable increases in publication frequency. It is an effective indicator of a transformative discovery when it is observed with the betweenness centrality metric. An algorithm reported by Kleinberg et al. [30] was adopted to identify citation bursts.

3. Results

3.1. Trend of Publication Output

Figure 1 shows the number of published papers on rehabilitation robots per year from 2000 to 2019. Since 18 papers were published in 2000, rehabilitation robot publications increased slowly and unsteadily over the first 6 years, reaching 31 papers published in 2005. From 2006 to 2015, an increasing number of scholars paid attention to this area, leading to a steady and rapid increase in rehabilitation robot publications. Starting in 2017, the number of publications rapidly grew, with a rate of increase of almost 38%, from 332 in 2016 to 458 in 2017. The number of total publications showed a significantly increasing trend from 2000 to 2019, especially in the more recent years. The results indicate a continuous increase in research interest in rehabilitation robot studies.

3.2. Subject Category Co-Occurrence Analysis

Each publication indexed in WoS is assigned one or more subject categories. According to Garfield’s judgment, the classification of an author’s discipline subject and research specialty is reflected by the relatively influential literature published by said author [39]. Based on the classification of subject categories in the WoS database, the publications on rehabilitation robots have been distributed in 127 subject categories over the last 20 years. Taking the category as the research object, the subject co-occurrence network was constructed, and the internal connections of the subject mechanism in the field of rehabilitation robots were explored. In this study, the network node was selected as the “category”, and the distribution map of subject categories in the field of rehabilitation robots was obtained using CiteSpace software. As shown in Figure 2, the top four subjects are engineering, rehabilitation, neuroscience and neurology, and biomedical engineering. The most common category is engineering, which has the largest circle. The results show that the major disciplines related to engineering in the field of rehabilitation robots are robotics, computer science, automation and control systems, electrical and electronic systems, mechanical engineering, instrumentation, and artificial intelligence. Related medical disciplines include neuroscience, clinical neurology, sport science, and medical informatics.

All of these disciplines form the foundation of the field of rehabilitation robots. To reveal which disciplines play the most important roles in promoting the development of rehabilitation robots, the number of publications in particular disciplines and their subject centrality values were studied.

Table 1. Subject category co-occurrence analysis.

Frequency	Burst	Centrality	Subject Category
1346	0	0.68	Engineering
1022	11.16	0.11	Rehabilitation
921	0	0.26	Neurosciences and neurology
791	0	0.08	Engineering, biomedical
639	0	0.14	Neurosciences
605	8.64	0.01	Robotics
401	12.51	0.12	Computer science
373	9.35	0.06	Clinical neurology
325	0	0.02	Automation and control systems
293	0	0.04	Engineering, electrical and electronic
234	0	0.03	Engineering, mechanical
214	15.26	0.02	Computer science, artificial intelligence
204	12.52	0.01	Sport sciences
116	0	0.01	Instruments and instrumentation
93	0	0	Engineering, manufacturing
91	0	0.02	Science and technology—other topics
90	0	0.03	Computer science, information systems
83	9.88	0	Medical informatics
81	0	0.06	Chemistry
77	0	0	Multidisciplinary sciences

shows that the subject categories with high values for rehabilitation robots are engineering (centrality = 0.68), neuroscience and neurology (centrality = 0.26), neuroscience (centrality = 0.14), computer science (centrality = 0.12), rehabilitation (centrality = 0.11), and biomedical engineering (centrality = 0.08). These subject categories with high centrality values may lead rehabilitation robot research to new perspectives, play important roles in the network, and can be identified as turning points. Engineering, rehabilitation, neuroscience and neurology, robotics, and computer science are the major turning points linking different subjects and having a significant effect on the evolution of the technology. Through burst detection on subject categories, computer science, artificial intelligence (burst = 15.26), sport sciences (burst = 12.52), computer science (burst = 12.51), rehabilitation (burst = 11.16), medical informatics (burst = 9.88), clinical neurology (burst = 9.35), and robotics (burst = 8.64) showed rapid growth in the number of publications, indicating the most active subject areas in the field. Through burst detection for subject categories, we can identify the most active and rapidly growing topics in this research area. There was an explosive increase in the number of publications in “Computer Science, Artificial Intelligence”, indicating that this subject category was an active area in the study of rehabilitation robots.

3.3. Journal Co-Citation Analysis

CiteSpace was used to generate journal dual-map overlays. The dual-map design can lead to a natural representation of a citation, from its origin to its destination, as if it were the path of a cross-continental flight [40]. Figure 3 shows a journal map established based on the citation relationships among various journals publishing research on rehabilitation robots. The superimposed oval on the right indicates the journal where the cited document is located, representing the basic document; the journal dual-map overlay on the left was constructed based on the citing journals in the field of rehabilitation robots, representing the frontier documents. These overlays reveal the evolution of the knowledge base of the domain by comparing the subject categories of journals in the field of rehabilitation robots. From the subject category perspective, the main areas of specialty cited are mathematics, systems, mathematical medicine, medical clinical, neurology in sports, and ophthalmology in journals on the left, representing research trends such as systems, computing, computer, sports, rehabilitation, and sport. As already mentioned, subject category co-occurrence analysis was conducted by classifying the journals in which articles were published, while journal co-citation analysis was carried out for journals in which the cited articles appeared (cited journals). By identifying frequently cited journals, journal co-citation analysis provides important insights into the journals that collectively form the intellectual basis of a knowledge domain. The top 15 journals in terms of co-citation frequency were J. NeuroEng. Rehabil., IEEE Trans. Neur. Sys. Rehabil. Eng., Arch. Phys. Med. Rehabil., J. Rehabil. Res. Dev., Neurorehab. Neural Repair, Stroke, Phys. Ther., IEEE Int. Conf. Robot., J. Neurophysiol., Int. Conf. Rehabil. Robot., Brain, Clin. Rehabil., Exp. Brain Res., IEEE Trans. Bio-Med. Eng., and IEEE-ASME Trans. Mech.

Table 2. Journal co-citation analysis.

Frequency	Burst	Centrality	Sigma	Subject Category	WOS Subject Categories
1726	0	0.04	1	J. NeuroEng. Rehabil.	Rehabilitation/neurosciences/engineering, biomedical
1722	0	0.09	1	IEEE Trans. Neur. Syst. Rehabil. Eng.	Rehabilitation/engineering, biomedical
1651	0	0.05	1	Arch. Phys. Med. Rehabil.	Rehabilitation/sport sciences
1521	0	0.01	1	J. Rehabil. Res. Dev.	Rehabilitation/rehabilitation
1481	0	0.05	1	Neurorehab. Neural Repair	Rehabilitation/clinical neurology
1331	0	0.03	1	Stroke	Peripheral vascular disease/clinical neurology
1072	0	0.03	1	Phys. Ther.	Rehabilitation/orthopedics
820	0	0.05	1	IEEE Int. Conf. Robot.	/
812	0	0.09	1	J. Neurophysiol.	Neurosciences/physiology
805	0	0.03	1	Int. Conf. Rehabil. Robot.	/
786	3.57	0.10	1.43	Brain	Clinical neurology/neurosciences
776	0	0.04	1	Clin. Rehabil.	Rehabilitation
774	5.73	0.06	1.43	Exp. Brain Res.	Neurosciences
764	0	0.13	1	IEEE Trans. Bio-Med. Eng.	Engineering, biomedical
739	0	0.04	1	IEEE-ASME Trans. Mech.	Automation and control systems/engineering, electrical and electronic/engineering, manufacturing/engineering, mechanical

reveals that these top 15 journals are the primary publishing outlets as well as the dominant sources of cited works for rehabilitation robot researchers. IEEE Trans. Bio-Med. Eng. has the highest betweenness centrality ratio (centrality = 0.13), and its publications have been widely cited by rehabilitation robot scholars. Brain and Exp. Brain Res. have the highest citation burst values (3.57 and 5.73, respectively).

3.4. Document Co-Citation Analysis

Co-citation analysis has been proven to be an effective method to identify the intellectual structure of a field of research. Based on the literature dataset of 3527 articles published between 2000 and 2019, a network of document co-citation networks was constructed and visualized in CiteSpace. The clustering land view of publications generated between 2000 and 2019 is shown in Figure 4. Sixty co-citation clusters of rehabilitation robots were obtained from the literature and automatically labeled with CiteSpace, of which there were 10 larger clusters, each representing a specialty direction. The cluster names were labeled according to the titles of the citing articles with high citation frequency. The modality (Q) and the mean silhouette (S) are two important parameters for measuring the quality of the results of the co-citation cluster network; the higher the parameter indicator value, the more reasonable the clustering result. Large Q values indicate better clusters of nodes, and Q > 0.3 indicates that the community structure of the network is significant. A large S value denotes a higher homogenization of nodes in each cluster, and S > 0.7 generally suggests that the cluster has high credibility [40]. In this research, there were 604 document co-citation network nodes, 1762 links, the modality (Q) was 0.7735, and the mean silhouette value (S) of the largest 10 clusters was 0.8605. Detailed information on these co-citation clusters is shown in

Table 3. Top 10 clusters in the document co-citation analysis.

Cluster ID	Size	Silhouette	Year	From	To	LLR
0	90	0.809	2012	2008	2018	Narrative review (466.3, 1.0 × 10−4); brain–machine interface (416.88, 1.0 × 10−4); walking assistance (388.21, 1.0 × 10−4);
1	87	0.793	2011	2002	2017	Spinal cord injury (306.13, 1.0 × 10−4); gait rehabilitation (299.64, 1.0 × 10−4); stroke rehabilitation (284.17, 1.0 × 10−4);
2	83	0.677	2003	1995	2009	Universal haptic drive (196.71, 1.0 × 10−4); energy recycling (191.57, 1.0 × 10−4); additional design feature (191.57, 1.0 × 10−4);
3	64	0.804	2007	2003	2015	Robot-assisted treadmill therapy (516.58, 1.0 × 10−4); robot-assisted gait training (328.7, 1.0 × 10−4); robotic gait training (250.31, 1.0 × 10−4);
4	44	0.883	2001	1995	2006	Treadmill training (121.96, 1.0 × 10−4); following motor (121.96, 1.0 × 10−4); automated locomotor training (111.32, 1.0 × 10−4);
5	43	0.818	1999	1992	2004	Increasing productivity (74.63, 1.0 × 10−4); robot-aided neurorehabilitation (74.63, 1.0 × 10−4); evidence-based arm rehabilitation (61.47, 1.0 × 10−4);
7	25	0.960	2000	1996	2005	Custom-designed haptic training (62.2, 1.0 × 10−4); restoring reaching ability (62.2, 1.0 × 10−4); post-stroke hemiparesis (62.2, 1.0 × 10−4);
9	17	0.971	1998	1995	2004	Physical treatment strategies (31.23, 1.0 × 10−4); post-stroke motor dysfunction (31.23, 1.0 × 10−4); new therapy (15.36, 1.0 × 10−4);
10	16	0.944	1998	1994	2002	Arm movement therapy (30.68, 1.0 × 10−4); robot assistance (30.68, 1.0 × 10−4); following stroke (28.96, 1.0 × 10−4);
12	9	0.946	1997	1993	2001	Rehabilitation robotics (12.66, 0.001); chronic stroke (0.04, 1.0); stroke rehabilitation (0.02, 1.0);

. Figure 4 shows the co-citation network map of the 50 most cited references in the rehabilitation robot literature between 2000 and 2019. The nodes represent the cited documents and are labeled with the representative author(s) of the cited documents as well as their years of publication. The size of every circle is proportional to the number of citations received by the document. Therefore, a large circle denotes a highly cited document. Generally, cluster labels are ranked by three different algorithms: tf*idf [41], log-likelihood ratio (LLR) tests [42], and mutual information (MI).

In this study, the labels were selected by log-likelihood ratio (LLR) tests, which tend to reflect a unique aspect of a cluster. The activity and impact of the clusters can be judged by the times when the clusters appear and the members of the clusters. Each cluster intuitively reflects the knowledge base formed by the cited reference, and the articles that cite the references in the cluster represent research trends. The knowledge base is a collection of academic literature cited by corresponding research groups, and the research trend is shown by a collection of literature that cites references in the intellectual structure. A common intellectual structure may produce different research trends. The number of members in the cluster indicates the size of the cluster. The more members there are, the larger the cluster. The literature with high centrality in the co-cited network is considered to be representative of the research studies. Table 3 shows the 10 largest clusters. The mean year refers to the average publication year of the corresponding cluster members. For example, the mean year in Cluster #0 is 2012, which means that the literature in the knowledge base of this group was published on average in 2012. Focusing on newly formed clusters can help identify research trends.

To study the evolution process of rehabilitation robots, more information should be revealed through time series. The evolution path of each cluster over time is shown in Figure 5. The publication year is displayed from left to right in the horizontal direction in a timeline view. The vertical direction presents the significant literature forming the knowledge base at each point in time, which can help researchers to quickly understand the research developments in this field. The clustering is arranged in descending vertical order according to the number of cited articles in the cluster, and the largest cluster is displayed at the top of the view. The colored curves represent the co-citation links added to each color’s corresponding year. Large round nodes or nodes with red tree rings deserve special attention because either they are highly cited articles, have high centrality, or both. The nodes with purple tree rings represent citation bursts, meaning that they could be emerging trends. The timeline view presents the clusters’ formation time and evolution process. Each timeline represents a specialty direction, and the evolution of the specialty is determined by the milestone literature in the knowledge base on the timeline. The three papers with the most citations in each year are displayed under the timeline, and the labels of the most cited references are located at the lowest position. It can be seen directly from Figure 5 that Clusters #0, #1, and #3 were formed after 2007.

Cluster #0 “narrative review (2012)”, with 90 members, was mainly dedicated to improving the functional ability after stroke using robot exoskeleton technology and orthoses. Cluster #1 “spinal cord injury (2011)” and Cluster #3 “robotic-assisted treadmill therapy (2007)” are new and active clusters because the cluster members are newly published literature and have steadily increased in number. It can be seen from Figure 5 that Cluster #2 “universal haptic drive (2003)” and Cluster #4 “treadmill training (2001)” were formed before 2003, and the mean publication years of these clusters’ members are 2003 and 2001, respectively. Cluster #2 focuses on practical application research, with research topics on gait training and random crossover studies, such as randomized controlled trials of speed plus treadmill training and the speed of post-stroke movement of the treadmill. Cluster #5 “increasing productivity (1999)”, Cluster #7 “custom-designed haptic training (2000)”, Cluster #9 “physical treatment strategies (1998)”, and Cluster #10 “arm movement therapy (1998)” were formed in approximately 1999 and are inactive clusters. Clusters #10 and #12 do not appear to have many current high-profile publications. It is evident that Clusters #0, #1, and #3 have high concentrations of nodes based on the citation bursts, indicating that these are the most recently formed clusters. Many cluster members with high centrality are found in Clusters #1 and #2.

4. Discussion

The three papers with the most citations in each year are discussed in this section. Analysis of the literature with high burst rates and keyword co-word analysis to detect the research trends in the field of rehabilitation robots in each period are discussed in this section.

4.1. Cluster Discussion

4.1.1. Cluster #0: Narrative Review

Cluster #0 is labeled “narrative review”, including 90 members with a silhouette value of 0.809, and the main research contents are exoskeletons and orthoses in rehabilitation. The literature comprising the cluster was published between 2008 and 2018. This cluster represents a sustained period of over 11 years and remained active until 2018—the most recent year of publication for a cited reference in this study. The key publications constituting the knowledge base and research trends of this cluster are shown in

Table 4. Cited references in cluster #0.

Frequency	Burst	Centrality	Sigma	Author	Year	Source
77	23.25	0.03	1.94	Esquenazi, A. [43]	2012	Am. J. Phys. Med. Rehabil. (American Journal of Physical Medicine and Rehabilitation)
67	20.17	0.03	1.85	Diaz, I. [44]	2011	J. Robot. (Journal of Robotics)
56	15.79	0.01	1.12	Dollar, A.M. [45]	2008	IEEE Trans. Robot. (IEEE Transactions on Robotics)
50	18.09	0.03	1.58	Heo, P. [46]	2012	Int. J. Precis. Eng. Man. (International Journal of Precision Engineering and Manufacturing)
50	18.09	0.04	2.15	Ramos-Murguialday, A.	2013	Ann. Neurol. (Annals of Neurology)
50	17.76	0.02	1.41	Polygerinos, P. [47]	2015	Robot. Auton. Syst. (Robotics and Autonomous Systems)
42	18.28	0	1.04	Meng, W. [48]	2015	Mechatronics
41	14.55	0.03	1.56	Kiguchi, K. [49]	2012	IEEE Trans. Syst. Man Cybern. Part B (IEEE Transactions on Systems, Man, and Cybernetics, Part B)
37	16.09	0.01	1.09	Zeilig, G.	2012	J. Spinal. Cord. Med. (Journal of Spinal Cord Medicine)
36	13.00	0.03	1.52	Belda-Lois, J.M.	2011	J. NeuroEng. Rehabil. (Journal of NeuroEngineering and Rehabilitation)
36	15.66	0.01	1.15	Yan, T.F. [50]	2015	Robot. Auton. Syst. (Robotics and Autonomous Systems)
35	15.09	0	1.05	Jamwal, P.K.	2014	IEEE-ASME Trans. Mech. (IEEE/ASME Transactions on Mechatronics)
35	15.22	0	1.03	Tucker, M.R. [51]	2015	J. NeuroEng. Rehabil. (Journal of NeuroEngineering and Rehabilitation)
33	0	0.05	1	Vitiello, N.	2013	IEEE Trans. Robot. (IEEE Transactions on Robotics)
32	11.55	0.01	1.07	Pennycott, A.	2012	J. NeuroEng. Rehabil. (Journal of NeuroEngineering and Rehabilitation)
32	0	0	1	Mehrholz, J.	2017	Cochrane Database Syst. Rev. (Cochrane Database of Systematic Reviews)
31	9.44	0.01	1.07	Roy, A.	2009	IEEE Trans. Robot. (IEEE Transactions on Robotics)
30	13.04	0.03	1.4	Chen Gong	2013	Crit. Rev. Biomed. Eng. (Critical Reviews in Biomedical Engineering)
30	13.04	0.01	1.11	Kawamoto, H.	2013	BMC Neurol. (BMC Neurology)
28	10.69	0.01	1.06	Dobkin, B.H.	2012	Neurorehab. Neural Repair (Neurorehabilitation and Neural Repair)

.

The top 20 most cited articles among the 90 members show the knowledge base of the cluster. In this cluster, the most cited article with the highest centrality value is the paper “The ReWalk Powered Exoskeleton to Restore Ambulatory Function to Individuals with Thoracic-Level Motor-Complete Spinal Cord Injury”, which evaluated the safety and usability of the ReWalk exoskeleton robot in restoring the motor functions of patients with spinal cord injuries [43]. The cluster is guided by reviews on subjects including lower-limb robotics [44]; actuation, sensory, and control systems of lower-extremity exoskeletons and rehabilitation orthoses [45]; BMI training in chronic stroke patients [46]; and soft robotic gloves [47]. Among the top 15 most frequently cited articles, 12 focused on robotic exoskeletons and orthoses, and Yan [50] reviewed the assistive strategies utilized by active locomotion augmentation orthoses and exoskeletons; the other 3 studied mechanisms and control strategies for lower-limb rehabilitation [11,48,49].

The citing articles in Cluster #0 form the extension of the knowledge base of the cluster and provide more valuable and novel information for the development of the cluster’s research trend.

Table 5. Citing papers in Cluster #0.

Coverage	GCS	LCS	Bibliography
11	150	1	Ang, K.K.; Chua, K.S.G.; Phua, K.S.; Wang, C.; Chin, Z.Y.; Kuah, C.W.K.; Low, W.; Guan, C. A randomized controlled trial of EEG-based motor imagery brain-computer interface robotic rehabilitation for stroke. Clin. EEG Neurosci. 2015, 46, 310–320. https://doi.org/10.1177/1550059414522229.
9	39	1	Naros, G.; Gharabaghi, A. Reinforcement learning of self-regulated beta-oscillations for motor restoration in chronic stroke. Front. Hum. Neurosci. 2015, 9, 391. https://doi.org/10.3389/fnhum.2015.00391.
9	43	1	Ang, K.K.; Guan, C. Brain–computer interface for neurorehabilitation of upper limb after stroke. Proc. IEEE 2015, 103, 944–953. https://doi.org/10.1109/JPROC.2015.2415800.
9	102	1	Soekadar, S.R.; Birbaumer, N.; Slutzky, M.W.; Cohen, L.G. Brain–machine interfaces in neurorehabilitation of stroke. Neurobiol. Dis. 2015, 83, 172–179. https://doi.org/10.1016/j.nbd.2014.11.025.
8	40	1	Brauchle, D.; Vukelić, M.; Bauer, R.; Gharabaghi, A. Brain-state-dependent robotic reaching movement with a multi-joint arm exoskeleton: Combining brain–machine interfacing and robotic rehabilitation. Front. Hum. Neurosci. 2015, 9, 564. https://doi.org/10.3389/fnhum.2015.00564.
8	21	1	García-Cossio, E.; Severens, M.; Nienhuis, B.; Duysens, J.; Desain, P.; Keijsers, N.; Farquhar, J. Decoding sensorimotor rhythms during robotic-assisted treadmill walking for brain–computer interface (BCI) applications. PLoS ONE 2015, 10, e0137910. https://doi.org/10.1371/journal.pone.0137910.
8	29	1	Yong, X.; Menon, C. EEG classification of different imaginary movements within the same limb. PLoS ONE 2015, 10, e0121896. https://doi.org/10.1371/journal.pone.0121896.
8	20	1	Ramos-Murguialday, A.; García-Cossio, E.; Walter, A.; Cho, W.; Broetz, D.; Bogdan, M.; Cohen, L.G.; Birbaumer, N. Decoding upper-limb residual muscle activity in severe chronic stroke. Ann. Clin. Transl. Neurol. 2015, 2, 1–11. https://doi.org/10.1002/acn3.122.
6	40	1	Hussain, S.; Xie, S.Q.; Jamwal, P.K. Robust nonlinear control of an intrinsically compliant robotic gait training orthosis. IEEE Trans. Syst. Man Cybern. -Syst. 2012, 43, 655–665. https://doi.org/10.1109/TSMCA.2012.2207111.
6	50	1	Ang, K.K.; Guan, C.; Phua, K.S.; Wang, C.; Zhao, L.; Teo, W.P.; Chen, C.; Ng, Y.S.; Chew, E. Facilitating effects of transcranial direct current stimulation on motor imagery brain–computer interface with robotic feedback for stroke rehabilitation. Arch. Phys. Med. Rehabil. 2015, 96, S79–S87 https://doi.org/10.1016/j.apmr.2014.08.008.

lists the top 10 of 151 citing articles, indicating the direction of the research frontier. The most representative and influential terms from the citing articles are used to determine the cluster characteristics and label the cluster names. The coverage value is the number of documents cited in the knowledge base in this article; the larger the value, the stronger the representativeness of the document. The global citation score (GCS) is the number of citations of the document in the entire database, while the local citation score (LCS) is the citation count of the literature in the data. These two indicators measure the impact of the paper. Coverage and citation values from the cited literature were comprehensively considered, and documents with coverage values greater than 5 were selected to track and monitor emerging trends and patterns to reveal the contribution of the research progress (see Table 5). In Cluster #0, the most active citer in the cluster was the research paper by Ang, K.K. (2015) [13], with a GCS of 150 and an LCS of 1. Ang reported “A Randomized Controlled Trial of EEG-Based Motor Imagery Brain-computer Interface Robotic Rehabilitation for Stroke”. In his study, he illustrated the significant results of the correlation of rBSI with motor improvements, suggesting that the rBSI can be used as a prognostic measure for brain–computer interface (BCI)-based stroke rehabilitation. The top 10 citing articles of the cluster reveal that brain–machine interface research is active and is an emerging trend in rehabilitation robots. As a neurorehabilitation intervention, BCI is an emerging trend in the field. Its research includes clinical evidence of its effectiveness, high-intensity learning, kinematic and dynamic modeling, self-administered training, and rehabilitation motor sensory rhythm. Electroencephalography (EEG)-based motor imagery (MI) brain–computer interface (BCI) technology has the potential to restore motor function by inducing activity-dependent brain plasticity. The use of BCI/BMI technology increases adherence to MI training more efficiently than interventions with sham or no feedback [52].

In brief, Cluster #0 was mainly dedicated to improving the functional ability after stroke using robot exoskeleton technology and orthoses. The research trends focus on the development of clinical applications of exoskeletons and strict quantitative assessment of their performance, energy consumption, and user experience. In other words, robot exoskeleton technologies can be combined with the accuracy of the exoskeleton model, adaptive control strategy, brain–computer interface, and biological design to achieve personalized motion assistance and to enhance power efficiency.

4.1.2. Cluster #1: Spinal Cord Injury

Cluster #1 has 87 members with a silhouette value of 0.793. It is labeled “spinal cord injury”, and its duration ranges from 2002 to 2017. This cluster focuses on the study of post-stroke exercise recovery, with an emphasis on the treatment of specific exercises with high intensity and repetitive tasks, including exercise evaluation, control strategies, system evaluation, and randomized controlled trials. The most prominent contributions in this period include the article by Lo, A.C. (2010) [53] that provides experimental evidence that high-strength robot-assisted rehabilitation treatment of at least 6 months is effective for stroke with moderate-to-severe upper-limb injuries; a survey by Maciejasz (2014) [3] on robotic devices for upper-limb rehabilitation, discussing these devices’ application field, target group, type of assistance, mechanical design, control strategies, and clinical evaluation; a review by Langhorne (2011) [54] focusing on the evidence underlying stroke rehabilitation, including the principles of rehabilitation practice, systems of care, and specific interventions; and an article by Lo, H.S. (2012) [55] on the prospects of exoskeleton robots for upper-limb rehabilitation. The top 20 most cited articles that reveal the evolution of research in this field are listed in

Table 6. Top 20 most cited references in Cluster #1.

Frequency	Burst	Centrality	Sigma	Author	Year	Source
270	23.70	0.08	6.36	Lo, A.C. [53]	2010	N. Engl. J. Med.
262	43.85	0.03	3.42	Kwakkel, G. [56]	2008	Neurorehab. Neural. Repair
135	40.78	0.03	3.71	Maciejasz, P. [3]	2014	J. NeuroEng. Rehabil.
105	18.19	0.04	1.89	Langhorne, P. [54]	2011	Lancet
96	14.82	0.04	1.72	Takahashi, C.D.	2008	Brain
93	26.67	0.01	1.4	Klamroth-Marganska, V.	2014	Lancet Neurol.
87	18.79	0.02	1.36	Norouzi-Gheidari, N.	2012	J. Rehabil. Res. Dev.
80	12.56	0.04	1.67	Huang, V.S.	2009	J. NeuroEng. Rehabil.
75	21.18	0.04	2.45	Wolf, S.L.	2006	JAMA J. Am. Med. Assoc.
73	15.23	0.02	1.44	Langhorne, P.	2009	Lancet Neurol.
73	4.83	0.04	1.21	Lo, H.S. [55]	2012	Med. Eng. Phys.
61	18.67	0.01	1.26	Mehrholz, J.	2012	Cochrane Database Syst. Rev.
59	16.84	0.02	1.46	Basteris, A.	2014	J. NeuroEng. Rehabil.
57	9.27	0.04	1.5	Housman, S.J.	2009	Neurorehab. Neural Repair
55	14.12	0.11	4.12	Volpe, B.T.	2008	Neurorehab. Neural Repair
52	18.48	0.02	1.36	Mehrholz, J.	2015	Cochrane Database Syst. Rev.
51	0	0.02	1	Veerbeek, J.M.	2017	Neurorehab. Neural. Repair
50	16.90	0.02	1.39	Chang, W.H.	2013	J. Stroke
43	10.90	0	1.04	Masiero, S.	2011	J. Rehabil. Res. Dev.
42	7.86	0.02	1.2	Bosecker, C.	2010	Neurorehab. Neural. Repair

.

4.1.3. Other Clusters

Cluster #2 is labeled “universal haptic drive”—a research topic concerning gait training and random crossover studies, such as randomized controlled trials of speed plus treadmill training and the speed of post-stroke movement on the treadmill. This cluster focuses on practical application research. The cluster was active from 1995 to 2009. In this cluster, Ferraro, M. (2003) [57], Marchal-Crespo, L. (2009) [58], and Hesse, S. (2003) [52] produced key documents and published influential results. Cluster #3 is labeled “robotic-assisted treadmill therapy”. The research period was from 2003 to 2015. The primary content of the cluster is the basic research of rehabilitation-robot-based assistive technology. Currently, this cluster is nearly inactive. Cluster #4 is labeled “treadmill training”, focusing on mechanistic research such as walking function recovery, repeated training on hand function, nerve matrix effects on rehabilitation training, and mechanical gait training. This cluster is inactive, and its intellectual structure presents a steady state. The remaining clusters are relatively small or are inactive; therefore, the discussion of these clusters is omitted.

4.2. Publications with High Betweenness Centrality Ratios

The betweenness centrality of a node is a measurement of the importance of the position of the node in the co-citation network. A node with a high betweenness centrality ratio means that the node is highly connected to other nodes or connects different co-citation clusters.

Table 7. Top 10 betweenness centrality values.

Frequency	Centrality	Sigma	Author	Year	Source	Cluster
55	0.14	12.37	Ferraro, M. [57]	2003	Neurology	2
140	0.13	40.13	Marchal-Crespo, L. [58]	2009	J. NeuroEng. Rehabil.	2
55	0.11	4.12	Volpe, B.T. [60]	2008	Neurorehab. Neural Repair	1
61	0.10	6.60	Hesse, S. [59]	2003	Arch. Phys. Med. Rehabil.	2
52	0.09	5.56	Fasoli, S.E.	2004	Arch. Phys. Med. Rehabil.	2
270	0.08	6.36	Lo, A.C. [53]	2010	N. Engl. J. Med.	1
101	0.08	17.16	Lum, P.S. [61]	2002	Arch. Phys. Med. Rehabil.	2
119	0.08	2.67	Hidler, J.	2009	Neurorehab. Neural Repair	3
16	0.08	1.78	Colombo, G. [62]	2001	Spinal Cord.	4
16	0.08	1.76	Whitall, J.	2000	Stroke	9

lists the references with the top 10 betweenness centrality ratios of all nodes in the network, and literature with high betweenness centrality is usually the key hub connecting two different areas—also called the turning point. The nodes with betweenness centrality ratios > 0.1 were turning points. These articles were considered landmark studies in the research field of rehabilitation robots. The work of Ferraro, M. (2003) [57] has the highest betweenness centrality (0.14). In this pilot study, task-specific motor training attenuated chronic neurological deficits well beyond the expected period for improvement after stroke. Hesse, S. (2003) [59] studied arm trainers and enabled intensive bilateral elbow and wrist training of severely affected stroke patients. Volpe, B.T. [60] studied new protocols rendered by either therapists or robots that could be standardized, tested, and replicated and potentially contribute to rational activity-based programs.

The burst rate of a document can reflect its citation burst in a certain area of specialization in a certain period. A stronger burst shows higher attention to this research topic, which represents the research trend in certain periods. The research trend of rehabilitation robots in each period can be detected by analyzing the literature with high burst rates. In

Table 8. Top 15 citation bursts until 2019.

References	Year	Strength	Begin	End
Yan, T.F., 2015, Robot. Auton. Syst.	2015	37.9735	2017	2019
Wang, S.Q., 2015, IEEE Trans. Neurol Sys. Rehabil. Eng.	2015	26.9781	2017	2019
Tucker, M.R., 2015, J. NeuroEng. Rehabil.	2015	25.6632	2017	2019
Polygerinos, P., 2015, Robot. Auton. Syst.	2015	23.9115	2017	2019
Meng, W., 2015, Mechatronics	2015	23.0363	2017	2019
Maciejasz, P., 2014, J. NeuroEng. Rehabil.	2014	47.8446	2016	2019
Collins, S.H., 2015, Nature	2015	26.5816	2016	2019
Zeilig, G., 2012, J. Spinal Cord Med.	2012	23.5634	2016	2019
Kiguchi, K., 2012, IEEE Trans. Syst. Man Cybern. Part B	2012	23.5634	2016	2019
Esquenazi, A., 2012, Am. J. Phys. Med. Rehabil.	2012	41.795	2015	2019
Klamroth-Marganska, V., 2014, Lancet Neurol.	2014	29.9453	2015	2019
Ramos-Murguialday, A., 2013, Ann. Neurol.	2013	24.2579	2015	2019
Langhorne, P., 2011, Lancet	2011	23.9969	2015	2019
Diaz, I., 2011, J. Robot.	2011	23.9014	2015	2019
Lo, A.C., 2010, N. Engl. J. Med.	2010	38.0828	2011	2019

, each colored segment represents a time slice of 1 year. The period in which a study was found to exhibit a burst is shown as a red line segment, indicating the beginning and the end of the duration of the burst. Burst detection to analyze co-citation analysis can reveal emerging research trends in the knowledge field. The burst rate indicates that the literature experiences a sharp increase in literature citations over a certain period, which is shown as citation emergence on the indicator. The major milestones in the development of rehabilitation robots can be identified from the highest burst values, as shown in Table 8. Documents with higher burst rates can be regarded as milestone documents that represent scientific research in certain periods. Current research fronts in the rehabilitation robot domain include exoskeletons, mechanisms and control strategies, EMG-based control, soft exosuit stroke rehabilitation, wearable robots, prosthetics and orthotics, and hybrid assistive limbs. Exoskeleton research focusing on methods and theories includes unpowered exoskeletons, powered exoskeletons, design and control of exoskeletons, safety and tolerance of exoskeleton suits, and production cost reduction, having used exoskeleton theories and methods for the current application.

4.3. Keyword Co-Word Analysis

Keywords in the literature on rehabilitation robots were analyzed by a co-occurrence network analysis performed using both VOSviewer software and CiteSpace software. Figure 6 shows the density visualization network of keyword co-occurrence with frequencies >10. There are 515 keywords with occurrence frequencies over 10. The keywords with the highest co-occurrence frequencies indicate the hot research topics in rehabilitation robots, including rehabilitation (1497), stroke (1071), recovery (559), therapy (468), robotics (416), walking (370), arms (339), gait (348), design (451), exoskeleton (352), robot (368), rehabilitation robotics (327), reliability (214), performance (221), upper limbs (200), neurorehabilitation (192), upper extremity (185), stroke rehabilitation (173), and spinal cord injury (170). Co-occurrence clustering was used to analyze new trends in the keywords from 2000 to 2019. The strength of the links among co-occurring keywords indicates shared research topics. Through burst detection of keywords, we can identify the rapidly growing topics in this research area.

Table 9. Top 20 keyword bursts.

Keywords	Year	Strength	Begin	End
Manipulator	2000	12.8896	2016	2019
Brain–computer interface	2000	10.8979	2017	2019
Balance	2000	9.058	2017	2019
Gait rehabilitation	2000	8.3232	2015	2019
Randomized controlled trial	2000	6.4838	2014	2019
Orthosis	2000	5.4585	2015	2019
Robotic rehabilitation	2000	11.786	2016	2017
Human–robot interaction	2000	10.396	2016	2017
Chronic stroke	2000	9.3485	2014	2017
Impedance control	2000	8.6506	2014	2017
Device	2000	5.4634	2016	2017
Validity	2000	10.7783	2015	2016
Modulation	2000	8.9653	2015	2016
Proprioception	2000	8.437	2015	2016
Functional electrical stimulation	2000	7.8711	2014	2016
Lokomat	2000	7.0776	2015	2016
Quality of life	2000	6.7234	2015	2016
Motor function	2000	6.05	2009	2016
Robot-assisted therapy	2000	4.7467	2015	2016
Coordination	2000	11.1387	2010	2015

shows the keywords with the strongest citation bursts until 2015. Some research topics were identified as new trends in rehabilitation research that continued until 2019, including randomized controlled trials, gait rehabilitation, orthoses, manipulators, brain–computer interfaces, and balance. Additionally, some of the topics appear for the first time in citation bursts in 2014 and 2015, including randomized controlled trial, chronic stroke, impedance control, functional electrical stimulation, gait rehabilitation, orthoses, validity, modulation, proprioception, Lokomat, quality of life, and robot-assisted therapy.

5. Conclusions

In the past two decades, the literature on rehabilitation robots has increased significantly, mainly due to the increasing demand for disease rehabilitation and broader applications of the theory and technology in sports rehabilitation and robotic clinics. Interest in rehabilitation robots is rapidly growing and will continue to develop in the future. According to the distribution of output in subject categories, rehabilitation robots represent a typical interdisciplinary research field that involves numerous subject categories. The literature was obtained from various fields—primarily from engineering, followed by rehabilitation, neuroscience and neurology, and biomedical engineering. The results show that engineering is the foundation of rehabilitation robot research, and the core disciplines are clinical neurology, rehabilitation, and biomedical engineering. Orthopedics, sport sciences, computer science, artificial intelligence, clinical neurology, and medical informatics are also the six most active subject categories. The subject categories of these journals reveal the close integration of engineering and clinical applications in rehabilitation. Furthermore, the distribution of journals was analyzed by journal dual-map overlays, revealing the evolution of the knowledge base in the field of rehabilitation robots. Brain and Exp. Brain Res. had the highest burst values (3.57 and 5.73, respectively).

Based on the DCA results, rehabilitation robots can be divided into 10 mainstream research fields covering 478 members in 604 nodes, accounting for 79% of the members. The emerging trends were identified from the latest and largest clusters, including brain–machine interfaces and spinal cord injuries, which have become the most promising research fronts at present. Traditional rehabilitation robot research topics—including lower-limb rehabilitation technology, upper-limb rehabilitation technology, and treadmill therapy—remain current research fronts. We also observed that the rehabilitation robot research field has shifted from traditional physical therapy strategies to new therapies, such as exoskeleton robots, soft robot gloves, and BCIs. Furthermore, traditional assisted-therapy-based evaluation methods such as “increasing productivity”, “custom-designed haptic training”, “physical treatment strategies”, and “arm movement therapy” do not currently capture researchers’ attention and are no longer active research areas. Figure 7 shows the overall picture of this paper.

We found that the evolution of research topics between 2000 and 2019 is revealed by the timeline knowledge map through year-by-year DCA clustering. First, through further analysis of the knowledge structures of the 10 clusters, the development track of core fields related to rehabilitation robots was determined, thereby establishing the intellectual knowledge base for probing the emerging development directions. Emerging research topics were identified in Cluster #0 and Cluster #1 by structure and time indicators, such as citation frequency, centrality, and burst indicators. These two clusters have attracted the attention of many scholars in the period since 2011. Lower-limb robotics has become the core topic in the rehabilitation robot domain and has driven the shift in research topics for the entire rehabilitation field. By analyzing highly cited papers and the cluster members’ contents, some future research directions that have continued into 2019 can be identified, such as orthoses. In particular, “BCI” is a new area of interdisciplinary science that has developed in the last 20 years, combining BCI theory with information theory. In recent years, this subject has steadily moved from theory to viable advancement, i.e., more clinical tests. The high proficiency and security of inducing movement data transmission is receiving increasing attention. Considering the fundamental standards of quantum mechanics, this could become an exploration hotspot in quantum material science and data science. Under the four stages of the theoretical framework of discipline development, studies on BCIs in rehabilitation form the primary blueprint, following the logical steps of concept definition, tool construction, and application phases. Our results show that stroke (1071), recovery (559), therapy (468), walking (370), arms (339), gait (348), design (451), exoskeleton (352), reliability (214), performance (221), upper limbs (200), neurorehabilitation (192), upper extremities (185), stroke rehabilitation (173), and spinal cord injury (170) are the most frequent keywords, in addition to the search terms rehabilitation (1497), robotics (416), robot (368), and rehabilitation robotics (327). On the grounds of citation burst intensity and duration, randomized controlled trials (burst = 6.4838), gait rehabilitation (burst = 8.3232), orthoses (burst = 5.4585), manipulators (burst = 12.8896), BCI (burst = 10.8979), and balance (burst = 9.058) are becoming emerging trends in the field of rehabilitation robots from the perspective of the number of keyword burst citations. In addition, brain–computer interface technology—which has greatly strengthened the fusion of clinical neurology, robot control, and motor functions with reinforcement learning and control feedback—will profoundly impact the improvement of rehabilitation robots. Figure 8 shows the technical concept tree diagram drawn from the knowledge base of DCA clusters.

In this article, bibliometric analysis and concepts are applied to the rehabilitation robot literature. Assessing the impacts of scientific activity and scientific discoveries has practical implications for the scientific revolution, strategic research, and development planning. This article provides researchers with an intuitive understanding of the current research status of rehabilitation robots and a higher capability of identifying relevant publications. The findings and discussion of the subject composition and subject trends of rehabilitation robots can provide reference and suggestions for scholars, policymakers, and enterprise decision-makers involved in this field to establish further research directions. In addition, this study provides scientists with a way to obtain and combine various disruptive information links to locate scientific discoveries more easily. It is foreseeable that research on rehabilitation robots will continue to grow in the next 10 years to improve motor functions and quality of life for humans.

In this article, the bibliometric analysis method and its associated concepts are applied to the rehabilitation robot literature. Due to the time delays of research papers and reference citations, publication- and citation-based indicators are insufficient to reveal the latest research trends in a single research field; thus, the burst indicator was chosen to identify emerging trends. Limited by database incompatibility, this research is also extensible to the use of different structural data analysis methods as well as other database and patent data. Patent and document index strategy research is a promising route to identify trends. In the future, interdisciplinary and multilevel research methods should be developed to promote innovation and collaboration between engineering R&D and clinical trials, applications should be enhanced, and complementary empirical systems should be devised. Furthermore, the extent of the literature data was enormous, accounting for the entire field of rehabilitation robots. However, it may be difficult to achieve an in-depth analysis of all of the variables and parameters in each cluster.