Global research of artificial intelligence in strabismus: a bibliometric analysis

Purpose To analyze the global publications on artificial intelligence (AI) in strabismus using a bibliometric approach. Methods The Web of Science Core Collection (WoSCC) database was used to retrieve all of the publications on AI in strabismus from 2002 to 2023. We analyzed the publication and citation trend and identified highly-cited articles, prolific countries, institutions, authors and journals, relevant research domains and keywords. VOSviewer (software) and Bibliometrix (package) were used for data analysis and visualization. Results By analyzing a total of 146 relevant publications, this study found an overall increasing trend in the number of annual publications and citations in the last decade. USA was the most productive country with the closest international cooperation. The top 3 research domains were Ophthalmology, Engineering Biomedical and Optics. Journal of AAPOS was the most productive journal in this field. The keywords analysis showed that “deep learning” and “machine learning” may be the hotspots in the future. Conclusion In recent years, research on the application of AI in strabismus has made remarkable progress. The future trends will be toward optimized technology and algorithms. Our findings help researchers better understand the development of this field and provide valuable clues for future research directions.


Introduction
Strabismus, featured by misalignment of the eyes, is one of the most common ocular diseases (1).It can occur at all ages, while children are diagnosed more commonly, with a prevalence ranging from 2.40 to 5.65% (2)(3)(4)(5)(6).For young children, it is typical that strabismus occurs spontaneously, probably stemming from inheritance, low birth rate, prematurity and etc. (7).Individuals suffering from strabismus later in life may have gone through head injury, stroke, eye muscle damage during surgery or developed diseases, such as Graves' disease, diabetes and myasthenia gravis (1,8,9).Traditionally, the diagnosis of strabismus requires for specialized examinations like the cover and uncover test, the Hirschberg test and the alternate prism cover

Data collection
Based on the recommendation of collecting bibliometric data from single database (20), we used Web of Science Core Collection (WoSCC) as the research database, which is the most commonly used and approved database for bibliometric analysis.To search for relevant data, at least one keyword related to strabismus and at least one keyword related to AI were combined to form query formulation.The detailed query formulation was described as follows: TS = (AI OR "artificial intelligence" OR intelligent OR "data learning" OR "robotic*" OR "computer vision" OR "machine learning" OR "deep learning" OR "deep network*" OR "neural learning" OR automat* OR algorithm OR "neural network*" OR "expert* system*") AND TS = (strabismus OR esotropia OR exotropia OR hypertropia OR hypotropia OR heterotropia OR esophoria OR exophoria OR hyperphoria OR hypophoria OR heterophoria OR "dissociated vertical deviation" OR "dissociated horizontal deviation" OR "dissociated torsional deviation" OR "[(eye OR ocular) AND (motility OR movement*) AND disorder*]" OR "[(eye OR ocular) AND (alignment OR deviation)]" OR "media rectus" OR "lateral rectus" OR "superior rectus" OR "inferior rectus" OR "superior oblique" OR "inferior oblique" OR "third nerve" OR "fourth nerve" OR "sixth nerve" OR "abducens nerve" OR "oculomotor nerve" OR "trochlear nerve").The time span was from January 1, 2002 to March 31, 2023, and the document types were limited to articles, reviews and proceedings papers.The last search was conducted on April 29, 2023.A total of 280 retrieved documents were prepared for the following screening.

Data screening
To exclude irrelevant documents in retrieved documents, we set the practical inclusion criteria as follows: (i) involvement of AI technology, including deep learning, machine learning and automated devices; (ii) involvement of strabismus, including: (1) researches focused on strabismus; (2) researches focused on the characterized clinical features of strabismus; (3) researches focused on multiple diseases and strabismus.After reading the titles and abstracts of each document carefully, 148 documents were included for the bibliometric analysis.

Data analysis and visualization
Publications and citations of each year, countries, institutions, authors, journals, research domain and H-index were acquired from WoSCC.H-index, as a reference index, indicates the impact of a researcher, country, institution, journal on the development of the certain scientific field (29,30).The WoSCC intrinsic toolkits help to analyze general features mentioned above.Microsoft Excel 2019 was used to conduct polynomial regression analysis and to export charts and tables of publications, top-cited documents, productive countries, institutions, authors and journals, and hot research domain.The average growth rate of publications was calculated as follows: 1 100 Where t 1 : first year; t 2 : last year; p 1 : publication count of the first year; p 2 : publication count of the last year.
We used VOSviewer (software, version 1.6.19)(31) to visualize the collaboration of countries, institutions, authors and co-occurrence of keywords, and calculate the total link strength.Bibliometrix (package, version 4.1.2) of R (programming language, version 4.3.0)(32) was applied to create WordCloud of keywords and calculate fractionalized frequency of an author which reflect one's contribution to the publications.Fractionalized frequency was calculated as follows: Where AU j : the set of documents co-authored by the author j; h: a document included in AU j .

Analysis of publications and citations
On the basis of search strategy and inclusion criteria, we included 146 documents (Figure 1), including 110 articles, 26 proceedings and 10 reviews.The detailed publication numbers of different document , respectively.To model the publication and citation trends, polynomial regression analysis was conducted (2023 was excluded due to incomplete indexing).The fitted curves of y1 = −2E-05x 6 + 0.0013x 5 -0.0332x 4 + 0.3959x 3 -2.2868x 2 + 5.8559x − 2.627(R 2 = 0.8423) and y2 = −0.0003x 6+ 0.0212x 5 -0.498x 4 + 5.5629x 3 -29.853x 2 + 70.569x − 50.265(R 2 = 0.9642) indicated changes in the quantities of publications and citations with time, respectively.The trend of citations showed a similar tendency as publications, indicating AI in strabismus drew researchers' attention in the last decade and is in the developing phase.Additionally, we listed the top 10 documents ranked by annual citations count in Table 1.There were 7 articles, 2 reviews and a proceedings paper.Eight documents were published in the last 5 years.The earliest document, also the most impactful one, was by Donahue et al. (33).It drew up guidelines for automated preschool vision screening, including the identification of children with strabismus.The Pediatric Vision Scanner, detecting the absence of foveal fixation as a harbinger of strabismus and amblyopia, was regarded as a potentially recommended instrument.The top 10 documents ranked by total citations count were also listed in Table 1.Four articles and 1 proceedings paper published before 2014 were newly on the list.

Analysis of top productive countries and collaboration networks of countries
A total of 28 countries had published related research on this topic.Countries ranked by publication count were listed in the Table 2. Accounting for 35.62% (52/146) of included documents, USA was the most productive country with the most publications and the highest citations (687 times) and H-index (15).According to publication count, China (29/146, 19.86%) was the second leading country, followed by South Korea (11/146, 7.53%).When it came to collaboration networks of countries, we visualized it by means of VOSviewer (Supplementary Figure S2).The number of publications determined the size of the circle and the lines between circles stood for the co-authorship between countries.The thickness of lines indicated the strength of cooperation (termed total link strength).USA had the highest total link strength, followed by England and China.Seventeen countries with cooperative relationships were presented, while 11 countries lacking international cooperation were not shown in Supplementary Figure S2, like South Korea, Brazil etc.

Analysis of top productive institutions and collaboration networks of institutions
There were 248 institutions participating in related research.Table 2 summarized institutions ranked by publication count.Johns Hopkins University had the greatest contribution with 11 documents in total, followed by Vanderbilt University, Universidade Federal do Maranhao and Seoul National University, each of which published 6 documents.Although the number of documents was inferior to Johns Hopkins University, Vanderbilt University had the highest citations.Figure 2 visualized the co-authorship network of institutions.While  Nevertheless, the majority of institutions were dispersive and lack of cooperation (Supplementary Figure S3).

Analysis of top productive authors and collaboration networks of authors
Top productive authors were listed in Table 3 and the collaboration relationship among authors were displayed in Figure 3.Because some authors were not connected to each other, the largest set of connected items consisting of 30 authors were visualized.According to Figure 3, we could see that authors from the same country were more likely to have close cooperation.

Analysis of research domains and top productive journals
The 10 most common research domains were listed in Top 10 productive journals were listed in Table 5.To evaluate the academic influence of journals, we consulted impact factor (IF) and the Journal Citation Reports (JCR) (2021) assessment system.Among these journals, JAMA Ophthalmology had the highest IF with 8.253.Journal of AAPOS (IF: 1.325, Q4) had the most publications and highest citations.Four journals ranked Q1 in JCR (2021), three journals ranked Q2, two journals ranked Q3 and one journal ranked Q4.
To understand when these hotspots emerged and how they evolved, we divided the documents into 3 group based on publication  S4) were conducted by Bibiometrix Package in R. The more frequently-occurred, the bigger scale."Strabismus" was the most dominant keyword for the entire period, followed by "amblyopia".During 2002 and 2012, "polarization optics", "vision screening" was ascendent.During 2013 and 2017, "retina birefringence" and "fixation detection" were more prevailing compared with "artificial neural networks".During 2018 and 2023, "deep learning" and "machine learning" were dominant.

Discussion
In this study, we performed a bibliometric analysis to explore the features of the publications concerning the application of AI in strabismus from 2002 to 2023 and gave a comprehensive look at this research trends for the first time.
The production and growth rates of the scientific papers reflect how a particular subject of study is progressing.There was a relatively slow stage of development from 2002 to 2012.The number of publications during this period accounted for 18.49% (27/146) of the total.Then, there has been an overall rising tendency since 2013.*The total link strength was analyzed from all countries using VOSviewer software.

#
The total link strength was analyzed from all institutions using VOSviewer software.The advancements in AI technology and its use in ophthalmology were primarily the reason for the developing trend.The most influential document was guideline for automated preschool vision screening by Donahue et al. (33).Its high citations were closely related to its guiding value.And other impactful documents were focused on aspects including strabismus recognition using eye-tracking data and convolutional neural networks, instrument-based vision screen and eye alignment measurement using smartphone app (34-37, 39, 40, 44, 47).USA, China and South Korea were the main forces in researches concerning the application of AI in strabismus.USA was the most productive country with the most publications and the highest citations and H-index.A significant gap between the number of publications of USA and China was noticed, while China ranked second in the number of publications.This was intimately tied to the fact that USA had the most advanced research equipment and a large number of scientific researchers in the whole world.Meanwhile, USA had the closest international cooperation, while other countries such as South Korea and Brazil sustained high scientific productions with poor international collaboration.Publications from USA in recent years inclined to study new screening devices which were able to detect amblyopia risk factors or directly detect strabismus (48,49).Publications from China brought more focus on deep learning-based image analysis.Zhang et al. (50) proposed multi-feature fusion model which achieved an accuracy of 97.17%, sensitivity of 96.06%, specificity of 97.79%, and AUC of 0.969 in detecting strabismus.Shi and Tang (51) proposed a multitask deep learning model based on deep snake to improve strabismus iris recognition in complicated scenes.Publications from South Korea also paid more attention to automated detection of strabismus (52,53).
Johns Hopkins University had excellent academic reputation and played a dominant role in this field.Research groups from here developed and reported a pediatric vision screener which was an efficient tool for detecting central fixation based on retinal birefringence scanning, and Gramatikov and Guyton did dedicate   The co-occurrence map of keywords.The top 3 research domains were Ophthalmology, Engineering Biomedical and Optics.Journal of AAPOS was ranked the first in the most productive journal.By analyzing keywords, "deep learning" and "machine learning" may be the hotspots in the near future.Meanwhile, "screening" and "smartphone app" had been the common interests of researchers.This may be due to the fact that the diagnosis of strabismus was important in clinical procedure and it was image-dependent which was suitable for AI technology to handle.AI-based algorithms can extract characteristics from ocular images to construct models to facilitate diagnosis and quantitively assessment.
By analyzing recent publications, the latest advances were manifested in different ways.When it came to screening devices, Monahan et al. (49) held the view that the blinq™ Vision Screener was suitable for detecting strabismus in children with sensitivity of 87.5% and specificity of 51.3%.As for new algorithms, Huang et al. (61) proposed a method that could automatically identify the oculomotor nerve from dMRI tractography and Xie et al. (62) put forward a novel multimodal deep-learning-based multi-class network for automated cranial nerves tract segmentation called CNTSeg, which would be beneficial to the diagnosis of incomitant strabismus.Lou et al. (63) proposed a novel deep learning-based approach to automatically evaluate the amount of inferior oblique overaction from frontal facial images via GAR2U-Net.
With the prosperity of AI technology, the development direction is changing from digital devices [like photoscreener, eye trackers, virtual reality headsets (39,44,64)] to automatic screening methods, which is more economical and practical in undeveloped areas.In the future, optimizing algorithms and proposing new techniques will become the focus of this field.For one thing, with the development of instrument-based vision screening technology, promoting the application of efficient and economical devices in strabismus screening is an important topic.For another, as artificial intelligence is racing ahead, it is vital to develop segmentation algorithms with higher accuracy and models with better performance.Also, an automated integrated platform with functions such as diagnosis, evaluation and postoperative prediction is highly expected.
Compared with AI applications in diabetic retinopathy, glaucoma and other eye diseases, AI applications in strabismus was relatively fewer.It may be because of the following reasons.On the one hand, the dynamic tests are required during the diagnosis of strabismus while a static fundus photograph is of great value for the diagnosis of eye diseases like diabetic retinopathy.On the other hand, public large-scale datasets of strabismus are rare, which is inconvenient for academic groups worldwide to test their models and algorithms.As a result, it is essential to grasp the specific characteristics of strabismus and promote the establishment of public large-scale datasets so as to provide opportunity for researchers worldwide to train their models.Research on the application of AI in strabismus has made significant progress in recent years, but there is still a distance between the current AI in strabismus and the clinical practice.We anticipate a more effective use of artificial intelligence in strabismus.
There are several limitations of the study.First, we only retrieve data from WoSCC.Other databases like Pubmed, Embase, and Scopus were not obtained.Secondly, our query formulation may not be perfect enough to retrieve all publications in the research field.Third, when analyzing keywords, similar keywords like "image processing" and "image processing techniques" were not merged.

Conclusion
This study presented an overview regarding the global researches of AI in strabismus.It was the first bibliometric analysis of this field.Our findings help researchers better understand the development of this field and provide valuable clues for future research directions.Research on the application of AI in strabismus has made remarkable progress in recent years.The future trends will be toward optimized technology and algorithms.

FIGURE 1
FIGURE 1Detailed flowchart of this study.

FIGURE 2 Co
FIGURE 2 Co-authorship network visualization map of institutions.Each node represents one institution.The number of publications determines the size of the circle.Connecting lines represent collaboration between institutions.

FIGURE 3 Co
FIGURE 3 Co-authorship network visualization map of authors.Each node represents one author.The number of publications determines the size of the circle.Connecting lines represent collaboration between authors.
It reveals 5 clusters (in 5 colors): the diagnosis of strabismus (in red), diseases related to strabismus (in yellow), terms related to AI (in blue), terms related to epidemiology (in green) and technical key point (in purple).Each node represents one keyword.The number of occurrences determines the size of the circle.Connecting lines represent co-occurrence between different keywords.

TABLE 1
Top documents ranked by annual citations/total citations.
some institutions were not connected to each other, Figure2only demonstrated the largest set of connected items consisting of 48 items.University College London had the highest total link strength, followed by Moorfields Eye Hospital and Harvard Medical School.

TABLE 2
Top countries/institutions ranked by publication count.

TABLE 3
Top authors ranked by publication count.

TABLE 4
The most related research domains ranked by publication count.

TABLE 5
Most productive journals ranked by publication count.