Mapping interconnectivity of digital twin healthcare research themes through structural topic modeling

Abstract

Digital twin (DT) technology is revolutionizing healthcare systems by leveraging real-time data integration and advanced analytics to enhance patient care, optimize clinical operations, and facilitate simulation. This study aimed to identify key research trends related to the application of DTs to healthcare using structural topic modeling (STM). Five electronic databases were searched for articles related to healthcare and DT. Using the held-out likelihood, residual, semantic coherence, and lower bound as metrics revealed that the optimal number of topics was eight. The “security solutions to improve data processes and communication in healthcare” topic was positioned at the center of the network and connected to multiple nodes. The “cloud computing and data network architecture” and “machine-learning algorithms for accurate detection and prediction” topics served as a bridge between technical and healthcare topics, suggesting their high potential for use in various fields. The widespread adoption of DTs in healthcare requires robust governance structures to protect individual rights, ensure data security and privacy, and promote transparency and fairness. Compliance with regulatory frameworks, ethical guidelines, and a commitment to accountability are also crucial.

Introduction

A Digital Twin (DT) is a digitized representation of a possible or actual process, person, place, system, or device. The original definition of DTs included physical products, virtual products, and their connections¹. However, rapid developments in communication technologies, sensor technologies, cloud computing, big-data analytics, the Internet of Things (IoT), and simulation technologies have resulted in DT research growing exponentially^2,3, with more recently the application of DTs also receiving increasing attention from industry.

DTs consist of a computational model that is calibrated to the patient and is dynamically updated from data repeatedly collected from the patient. A DT integrates and analyzes data collected from a patient’s electronic health record, wearable devices, medical equipment, and other sources in real time. This gives healthcare providers a clearer picture of a patient’s overall health and allows them to formulate personalized treatment plans that take into account individual patient characteristics, medical history, genetic factors, and real-time physiological data, which can both improve patient outcomes and increase patient satisfaction⁴. DTs can be used to leverage patient data and machine learning (ML) algorithms to detect health risks early, predict disease progression, and implement preventative measures. Such a proactive approach helps to reduce healthcare costs and improve patient outcomes by enabling healthcare providers to intervene before medical conditions worsen^5,6. DTs can play an important role in optimizing the clinical operations in healthcare systems. By integrating and analyzing data from various sources, DTs can improve workflows, optimize resource allocation, and increase operational efficiency⁷. This enables hospitals to anticipate future demand, allocate resources appropriately and when they are needed, and increase operational efficiency⁸. DTs can be an important tool for training and simulation in healthcare. DTs can be utilized in simulated critical scenarios such as emergencies, allowing medical staff to practice their response skills in stressful situations and prepare them to be more effective in real-life emergencies⁹.

The above discussion indicates the great potential that DTs have in transforming healthcare systems in many ways, including patient care, predictive analytics, and in the optimization, training, and simulation of clinical operations. Reviewing the literature on how DTs have been applied in healthcare would provide insights into how they could be applied in the future. However, research trends in digital twins have mainly focused on systematic literature reviews^3,10–12. These systematic reviews only selected and analyzed targeted studies, and so there has been no overall review of DT articles in the healthcare field. In addition, digital twins are a field of research based on real-time big data. Therefore, with the advancement of DT, various challenges are arising in security and privacy issues, unethical use of data, and data governance^13,14. However, systematic review studies focusing on specific areas have limitations in providing insight into these aspects.

This study employed structural topic modeling (STM) to analyze a large corpus of healthcare-related DT literature retrieved from online databases. STM involves the use of topic models designed to uncover latent thematic clusters in text collections. It extends the capabilities of traditional models, such as latent Dirichlet allocation (LDA)¹⁵ and correlated topic modeling¹⁶. Conventional topic modeling techniques such as LDA assume that topics are discrete and independent, but in many applications the topics can overlap and documents may exhibit complex relationships between them. STM addresses these limitations by incorporating structural information into the modeling process. In addition to identifying topics, the approach estimates the prevalence of these topics by considering document-level covariates. Given the above-described situation, this study aimed to identify the research keywords related to DTs through STM to identify research trends and provide basic data for the application of DT technology in the healthcare field.

Methods

Article collection

Literature related to the focus of this study was acquired by searching five electronic databases to identify scholarly publications related to healthcare DTs. Google Scholar, Scopus, and PubMed were selected to ensure comprehensive coverage and the inclusion of rigorous peer-reviewed academic and medical literature. Concurrently, the Korean Research Information Sharing Service (RISS) and Korean Studies Information Service System (KISS) databases were searched to facilitate the inclusion of relevant South Korean resources. We conducted a systematic literature search using the keyword “digital twin with healthcare” across five major academic databases, covering the period from 2018 to 2024. The year 2018 was chosen as the starting point, as it marks the emergence of scholarly interest in DT technologies. Given the high volume of tangentially related publications retrieved during the initial search, we employed a screening process to ensure the inclusion of those studies with high relevance and methodological quality. Inclusion criteria required that articles be written in English and explicitly reference both “digital twin” and “healthcare” in either the title or abstract. Duplicate elimination procedures were applied to results retrieved from Google Scholar, which aggregates a broad spectrum of publications from sources including Scopus and PubMed. Duplicate entries were identified and removed through title-matching comparisons with records obtained directly from Scopus and PubMed. Following this process, the search yielded a total 758 unique scholarly articles. The distribution of these articles across the five databases is presented in Fig. 1a.

Fig. 1.

Statistical information of collected articles.

The temporal progression of scholarly publications investigating DTs in healthcare field was examined systematically. Figure 1b illustrates the annual publication volume from 2018 to 2024. The publication volume was initially negligible (in 2018), but this subsequently showed a consistent upward trajectory, culminating in approximately 250 scholarly contributions by 2024. The recent marked increase in publication volume suggests an increasing recognition of the potential importance of applying DTs in healthcare contexts.

Preprocessing

This study examined trends in healthcare DT technology applications by analyzing titles and abstracts, which have been established as appropriate sources for conceptual reviews due to them encapsulating scholarly information¹⁷. The data extraction methodology varied with the database: an API was used to facilitate title and abstract retrieval from Scopus and PubMed, whereas manual extraction was employed for RISS and KISS. For Google Scholar, which offers an extensive corpus but lacks an extraction API, PDF documents were processed using ChatGPT, which was selected due to its demonstrated proficiency in document parsing¹⁸. To avoid the inclusion of hallucination-related inaccuracies, a rigorous screening protocol was implemented wherein two authors independently verified the extracted content. Publication dates were additionally collected as metadata to facilitate a temporal analysis of how research priorities have evolved.

Text preprocessing was conducted to optimize data quality using the following sequential protocol: consolidation of titles and abstracts into a unified analytical corpus; elimination of punctuation, control characters, URLs, numerical values, superfluous spaces, and stopwords to enhance the data consistency and computational efficiency; application of lemmatization and conversion to lowercase to standardize lexical formats; and exclusion of terms comprising five or fewer characters to eliminate nonsubstantive vocabulary items.

STM analysis

In STM¹⁹, each document is composed of multiple topics, and the proportions at which these topics are mixed are determined by a regression model influenced by the metadata. This regression model generates the topic distributions, while the word distributions for each topic are configured to reflect interactions with the metadata. During the word generation process, each word in a document is selected from a specific topic, and the word is generated based on the probability distribution of that topic. In STM, the generation process for each word and each topic in each document proceeds as described below. The document-level preference proportions for each topic follow the logistic normal probability distribution:

1

is the covariance matrix, and , where is the coefficient matrix for a normal distribution in the topic occurrence model, and is the document covariate vector. Furthermore, the document-specific distribution of words for each topic , utilizing the document-level covariate , is defined as follows:

2

where is the baseline word distribution, is the topic-specific deviation, is the covariate group deviation, and is the interaction between the topic and covariate groups. Finally, each word in a document is assigned a topic label according to the following distribution based on the document-specific topic distribution :

3

Conditional on the selected topic, observed word is generated from the word distribution of that topic:

4

The final topics are selected by iteratively verifying whether each observed word can be attributed to the respective topic. Based on trained and , a STM can estimate latent variable for document . With , the topic proportion is calculated as follows:

5

where is the total number of topics. Determining the optimal number of topics is crucial for achieving analytical reliability. Four diagnostic metrics guided this process in the present study:

• Held-out likelihood, which is the model fit to validation data.

• Residual, which is the unexplained variance.

• Semantic coherence, which corresponds to the intratopic term relatedness.

• Lower bound, which represents the minimum model likelihood threshold.

Higher values of the held-out likelihood, semantic coherence, and lower bound indicate better model performance, while a lower residual value suggests a superior fit. The following three metrics were used in the comparative analysis of models:

• Appearance probability, corresponding to the word-topic association likelihood, which is limited by a high cross-topic term frequency.

• FREX, which combines frequency and exclusivity to increase the topic distinctiveness.

• Lift, which equals the ratio of the word-topic probability to the corpus-wide distribution, and identifies disproportionate representations.

Results

Topic selection

Figure 2 shows how metrics changed with the number of topics. As the number of topics increased, held-out likelihood and lower bound improved, while residuals decreased, indicating better model performance. However, semantic coherence declined, suggesting that many topics reduce interpretability. A balance was observed at seven or eight topics, where coherence remained stable and other metrics continued to improve. The seven-topic model reflected broad healthcare challenges, while the eight-topic model revealed more specific themes like ethics and stability, with greater semantic independence, making it preferable.

Fig. 2.

Metric values according to the number of topics.

Topic labels

Table 1 presents the high-probability words and the assigned name for each topic, based on term coherence and literature review. Topic 1 focused on the architecture of the information technology infrastructure, encompassing data networks and cloud computing frameworks. Terms such as “network,” “device,” and “architecture” were related to data management systems and cloud computing infrastructure components. Topic 2 emphasized ML paradigms, algorithmic methodologies, and model application purposes. Terms including “learn,” “model,” and “algorithm” reflected advancements in learning frameworks and analytical techniques. Topic 3 centered on metaverse and virtual reality (VR) technologies, with lexical items such as “metaverse,” “virtual,” and “immersive” highlighting experiential dimensions in digital healthcare environments. Topics 4 and 5 addressed personalized medicine in digital healthcare and security solutions in healthcare workflows, respectively. Topic 6 focused on digital transformation in manufacturing sectors, with terminology including “manufacture,” “industrial,” and “transformation” signifying technological advancements related to smart manufacturing and Industry 4.0 initiatives. Topic 7 addressed robotic systems predicated on human-centered design principles, while Topic 8 was related to ethical considerations and future trajectories in healthcare research.

Table 1.

Labels and representative terms of the eight topics.

Topic number	Topic label	Seven terms with the highest probabilities
1	Cloud computing and data network architecture	datum, network, device, enable, present, management, architecture
2	Machine-learning algorithms for application detection and prediction	model, learn, framework, machine, method, propose, simulation
3	Applications of the metaverse and virtual reality in healthcare	technology, virtual, metaverse, application, world, intelligence, healthcare
4	Digital health in personalized medicine	digital, patient, health, medical, personalize, medicine, treatment
5	Security solutions to improve data exchange and communication	healthcare, process, solution, security, improve, thing, analysis
6	Digital transformation in manufacturing sensors	digital, technology, industry, smart, application, manufacture, sector
7	Human-centered design for robotics and assistive systems	system, physical, human, monitor, design, service, provide
8	Ethical challenges and future directions in healthcare	healthcare, research, challenge, potential, study, future, development

We employed FREX scores for comprehensive topic interpretations, visualized in Fig. 3 as a tree map of high-scoring terms for each topic. The FREX methodology balances word frequency and exclusivity parameters to identify terminology that optimally represents thematic content through both its prominence and distinctiveness. The analysis of Topic 1 revealed distinctive terms including “compute,” “cloud,” and “infrastructure,” emphasizing characteristics of cloud computing. Topic 2 featured terms such as “accuracy,” “algorithm,” “detection,” and “prediction,” highlighting concepts used to evaluate ML models. Topic 3 incorporated “reality,” “education,” and “experience,” reflecting research in immersive metaverse applications. Topic 4 included “patient,” “clinical,” and “personalized,” indicating a focus on personalized medicine. Topic 5 emphasized terminology related to secure medical data transmission, focusing on healthcare communications security. Topics 6 and 7 illustrated DT implementation scenarios in industrial contexts, while Topic 8 incorporated “literature,” “ethical,” and “review,” demonstrating a focus on ethical considerations in healthcare research. The thematic content was initially derived from the highest-probability words, with refinement through the FREX analysis. The topic classification and labeling were validated in consultation with two domain experts in healthcare, DTs, and informatics.

Fig. 3.

Tree maps of the topics showing those with the highest FREX scores.

Topic proportion

Figure 4 visualizes the distribution of topic proportions as a bubble chart. Topic 4 focused on personalized medicine in digital healthcare, which constituted 0.17 of the corpus, establishing it as the most-prominent and comprehensively addressed thematic domain in the data set. Topic 8, which explored ethical considerations and future research trajectories, represented 0.15 of the corpus and was the second-most-common topic in the analyzed literature. Conversely, Topics 1 and 5 represented the smallest proportions of the corpus, reflecting that these domains have received less attention in previous healthcare DT research.

Fig. 4.

Bubble chart of the topic proportions.

Figure 5 visualizes proportional temporal shifts in the topic distribution. In each panel, the solid red line represents the estimated temporal trajectory of the topic prevalence across publication years, derived from the STM incorporating year as a covariate. The dashed red lines denotes the pointwise 95% confidence interval bounds associated with this estimate, reflecting the uncertainty at each time point. Topics 2, 3, and 8 demonstrated consistent upward trajectories, indicating progressively increasing scholarly attention in the healthcare DT literature. Conversely, Topics 4 and 6 exhibited discernible declining trends in their proportions over the analyzed period, which suggests a redirection of research focus away from previously dominant domains. Topics 1, 5, and 7 maintained relatively constant proportions throughout the analyzed period, indicating sustained-neither increasing nor decreasing-scholarly interest in these thematic areas. These temporal fluctuations in the topic proportions provide valuable insights into the evolutionary trajectory of research priorities in the healthcare DT domain, highlighting shifting paradigms and emerging areas of scholarly focus.

Fig. 5.

Topic proportions according to year.

Topic comparisons

Eight keywords were selected from the comprehensive topic set to analyze intertopic relationships, with associated weights calculated using the FREX and appearance-probability metrics. The polygons in the radar chart in Fig. 6 reflect both the magnitude and nature of the word-topic relationships.

Fig. 6.

Radar chart of the relationships between topics.

The analyzes demonstrated that the identified topics formed a complex network of interrelationships rather than existing as isolated entities. The data (network, device, and datum) and learning (model, learn, and algorithm) topics shared fundamental technical foundations, with infrastructure supporting data acquisition enabling the development of ML models. This foundation connects to the medicine (personalized medicine and treatment) topic, highlighting the role of data-driven analytics in healthcare contexts, exemplified by patient-specific data facilitating customized therapeutic interventions. The metaverse (metaverse and reality) and industry (smart and industry) topics exhibited strong interconnectivity regarding technological implementation. Immersive platforms in the metaverse environments supported industrial smart manufacturing capabilities and digital transformation initiatives. These domains were connected to the systems (robot and sensor) topic, particularly regarding technologies that simultaneously support metaverse physical interaction and industrial automation. The security (security and blockchain) topic includes critical elements connected across all domains that protect the integrity of medical information, ensure automation reliability, and support virtual-environment functionality. Consequently, security considerations span both data-centric platforms and their application domains. The research (research and future) topic constitutes foundational elements underlying all thematic areas. Technological advancements are driven by continuous research initiatives, while ethical considerations provide directional guidance for development trajectories, exemplified by ethical medical data utilization emerging as significant in both personalized medicine and security research frameworks.

Topic correlations

We visualized topic correlations based on the topic modeling results to examine thematic interrelationships. The network graph in Fig. 7 represents topics as nodes, with connecting edges indicating positive correlations between them. The edges are annotated with correlation coefficients quantifying the thematic associations, while the size and color of each node reflects the topic centrality and significance. The “security solutions to improve data processes and communication in healthcare” topic occupies the central network position with multiple connections, highlighting its fundamental importance. This security-focused topic was strongly correlated (coefficient = 0.41) with the “cloud computing and data network architecture” topic, indicating substantial technical interdependence. The “cloud computing and data network architecture” and “machine-learning algorithms for accurate detection and prediction” topics function as bridging elements between technical domains and medical applications, suggesting significant cross-domain utility potential and facilitating technological implementation in healthcare contexts.

Fig. 7.

Correlations between the topics.

Discussion

This study applied STM with the aim of identifying research trends in the literature on the application of DTs in the healthcare field. Research in this area was first reported in 2018, followed by a gradual increase in the number of publications from 2018 to 2020, since when there has been a marked increase. This trend signifies the growing momentum and increasing scholarly interest in the adoption of DT in the healthcare domain. The availability of crucial healthcare data, a robust cyber technology infrastructure, and scientific expertise combined with the exponential growth of big data and continuous advancements in data science and artificial intelligence (AI) have the potential to significantly accelerate the research and development of DTs³.

Eight topics were identified. Topic 1 is essential for the successful implementation of DT-based healthcare solutions. The cloud computing and network architectures enable real-time data processing, AI-powered analytics, and secure storage^20,21. Liu et al.⁷ proposed CloudDTH, a cloud-based digital twin healthcare framework that uses wearable devices to monitor and predict individual health, especially for the elderly, and enables seamless integration between physical and virtual medical environments. Topic 2 is related to the integration of big-data analytics with AI. Future research should focus on scalable AI models, real-time adaptive learning, and federated privacy- preserving ML for next-generation DT systems^22,23. Hussain et al²⁴ conducted a hybrid attention based ResNet architecture for ICH detection and classification. This research exemplifies Health 4.0 digital twin implementation by creating digital replicas of medical imaging data for simulation, analysis, and prediction. The integration of advanced AI techniques demonstrates the potential for comprehensive digital health twin systems²⁴. Topic 3 refers to the integration of VR, augmented reality, AI, and DT technologies transforming healthcare by enabling life-like simulations, AI-driven diagnostics, and patient-specific treatment models. Medical training and surgical planning are enhanced by AI-powered DTs. AI-driven VR therapy improves pain management, rehabilitation, and mental healthcare. Metaverse-based virtual hospitals bring telemedicine, remote monitoring, and personalized care to new levels²⁵. DT technology in neurosurgery addresses major challenges including the shortage of trained surgeons, training requirements, and instrumental limitations. By generating virtual models of patients’ neuroanatomy and clinical profiles, DT enhances treatment possibilities throughout all phases of surgical management²⁶. Topic 4 is related to clinicians leveraging real-world digital health data. Ongoing advancements in AI, big data, and IoT technologies will result in digital health driving the future of personalized medicine, making treatments more effective, predictable, and patient-centered^27,28. Among the most studied uses of DT are those in oncology, where they enable a precise, highly personalized and dynamic approach to cancer treatment. The creation of digital replicas serves as the foundation for personalized care by allowing a comprehensive understanding of individual cancer cases at the molecular and cellular levels. These virtual models enable healthcare providers to simulate, analyze, and predict cancer progression and treatment outcomes in a virtual environment before implementation in clinical settings. This personalized approach minimizes trial-and-error treatments while maximizing therapeutic efficacy and quality of life for cancer patients^29,30. Topic 5 is essential to protect patient data from unauthorized access, breaches, and misuse while ensuring compliance with regulations such as the HIPAA (Health Insurance Portability and Accountability Act) and GDPR (General Data Protection Regulation). Addressing these challenges will require the implementation of robust security solutions to enhance data processes, secure communications, and protect patient confidentiality in digital healthcare environments²⁸. The relevance of Topic 6 is evident in how DTs are used to simulate complex production processes in biomanufacturing and pharmaceutical industries such as bioreactor conditions, molecule synthesis, and drug formulation. By replicating key process parameters and critical process parameters such as temperature, pH levels, and dissolved oxygen levels, DTs allow manufacturers to simulate and refine operations before their real-world implementation, reducing costs and improving yields²⁸. The evolution of DT technology powered by AI, IoT, cognitive computing, and generative AI is shaping the future of “human-centered design for robotics and sensor systems” (Topic 7). DTs allow intelligent, real-time simulations for optimizing human–robot interactions, adaptive and responsive sensor-driven environments, and autonomous decision-making for the application robotics in healthcare, mobility, and industrial automation^2,22. Topic 8 is interwoven. The rapid advancement of DT in healthcare has introduced ethical concerns. Alhammad et al.³¹. identified significant ethical, regulatory, and governance challenges surrounding DTs, particularly given their dependence on sensitive health data. The authors proposed adapting existing regulatory frameworks from virtual reality and IoT devices to establish comprehensive standards for data ownership, informed consent procedures, and operational transparency. Ringeval et al.¹⁰ recommended aligning these regulatory standards with established international ethical guidelines to strengthen public confidence and ensure responsible implementation of DTs in healthcare settings. While DTs offer transformative benefits in personalized medicine, predictive diagnostics, and patient-specific treatment planning, their development and implementation must adhere to ethical and legal frameworks to prevent misuse and ensure equitable healthcare access^28,30.

The topic proportion quantified the extent to which each topic constituted the dataset, providing insights into the importance or contribution of the topic. Topics 4 and 8, focusing on personalization and ethics, were the most prominent. While technical methodologies have been frequently studied, outcome-related topics were underexplored, suggesting a need for future studies to evaluate the impact of core technologies. Topic 5 was positioned at the center of the network and was strongly correlated with Topic 1, demonstrating a close technical linkage between these two themes. Moreover, both Topics 1 and 2 serve as bridges among the healthcare DT topics. The integration of AI and ML will enhance the efficient analysis of large and complex datasets, significantly improving the predictive accuracy and enhancing model personalization³². The application of DTs in healthcare relies on extensive health data, including highly sensitive personal health information. Therefore, the cloud computing and data network architecture along with the associated data privacy and security are important considerations that need to be addressed.

This study applied STM to healthcare-related DT literature, but it was unclear whether the included articles directly investigated DTs. There was limited information on real-world applications and effectiveness. As STM lacks standardized evaluation guidelines, subjective interpretation may influence results. To address this, classification and labeling were reviewed by two experts. DTs support healthcare innovation by simulating diverse patients and treatments. For widespread adoption, strong governance, data protection, transparency, and adherence to ethical and regulatory frameworks are essential.

Conclusion

This study analyzed research trends in the adoption of DT in healthcare using STM techniques. Eight topics were identified in the topic modeling analysis. It was found that digital health in personalized medicine and ethical challenges in healthcare are the topics that have been studied most extensively. Meanwhile, research into ML, VR, and ethical concerns has received increasing attention, highlighting the integration of AI and data-driven healthcare alongside emerging ethical challenges. The association between security and cloud computing emphasizes the importance of data privacy and network reliability when utilizing DTs. The successful adoption of DTs in healthcare requires privacy concerns, data standardization, and clinical adoption barriers to be addressed in order to realize the full potential of these innovations.

Our analysis was confined to five databases. While these repositories provide broad coverage, they may omit key sources such as conference proceedings and regional publications. To address these limitations, future studies should utilize a more comprehensive set of databases. Future studies need to expand the research on the actual application and evaluattion of DT in the healthcare field and conducting research to verify the practical application effects of DT through systematic literature reviews.

Author contributions

E.K.: Conceptualization, Supervision, Methodology, Result interpretation, Writing Manuscript. Y.L.: Data acquisition, Data analysis, Result interpretation, Writing manuscript.

Funding

This research was supported by a grant from the National Research Foundation of Korea (NRF-2022R1A2C1009890)