Mapping the Landscape of Healthcare Supply Chain Management Through an NLP-Driven Systematic Review

Abstract

Background: Healthcare supply chains (HSCs) are critical socio-technical systems that ensure the timely delivery of pharmaceuticals, medical devices, and electromedical equipment, yet they face increasing complexity due to regulatory constraints, demand uncertainty, and the growing digitalization of healthcare systems. This study aims to systematically map the HSC literature and identify its main thematic structures and research gaps. Methods: A systematic literature review was conducted following PRISMA guidelines, analyzing 705 peer-reviewed articles retrieved from the Web of Science database (PROSPERO registration: CRD42024605761). Natural language processing techniques were applied to support the analysis, including topic modeling, term frequency–inverse document frequency for keyword relevance, and Keyword in Context analysis for semantic interpretation. Results: The analysis identified six main thematic clusters and revealed a fragmented research landscape, characterized by limited integration across supply chain tiers, uneven attention to technological innovations, and marginal consideration of sustainability and implementation issues. The findings also highlight a gap between conceptual developments and real-world applications. Conclusions: This study provides a data-driven overview of the HSC research domain, highlighting key gaps and opportunities for more integrated, resilient, and efficient supply chain management.

1. Introduction

Supply chains refer to the interconnected network of organizations and processes involved in producing and delivering a product or service from its origin to the end customer [1]. Effective supply chain management (SCM) is widely recognized as an integral part of business success across industries, ensuring that goods flow efficiently to meet consumer demand [2]. Indeed, in sectors such as manufacturing and retail, well-managed supply chains enable industrial systems to reduce costs, improve efficiency, and reach international markets. Therefore, without a robust supply chain, goods would not efficiently reach the customers, underscoring their role in the global economy [3]. Thus, in every industrial system, supply chains form the backbone of operations by aligning suppliers, manufacturers, distributors, and retailers toward the common goal of customer satisfaction and market competitiveness [4].

In healthcare, supply chains play an important role since the industry depends on the timely availability of medical goods to save lives. HSCs encompass the provision of pharmaceuticals (such as medications and vaccines), medical devices (from syringes to implants), and electromedical equipment (such as ventilators and diagnostic machines). Ensuring that these items are available at the right place and time is a life-saving imperative. However, managing HSCs is uniquely challenging due to strict regulatory requirements, the need for temperature-controlled logistics (e.g., for certain medicines and vaccines), and highly uncertain demand patterns (such as sudden disease outbreaks or a pandemic/epidemic event) [5]. Growing cost pressures due inflation, rising demand, and increasing interdependencies in healthcare delivery have driven efforts to optimize efficiency, which in turn has made healthcare supply networks more complex [6,7]. As complexity increases, so does risk—a HSC with many interconnected players and just-in-time practices is inherently more vulnerable to disruptions [8]. Moreover, the multi-tiered nature of HSCs, involving several agents such as manufacturers, distributors, hospitals, pharmacies, and regulators, means that a breakdown at one point can have cascading effects. A delay or shortage in the supply of a critical item, for instance, a vital medicine or surgical equipment, can impact patient care by delaying treatments or forcing the use of less effective substitutes. In short, while efficient supply chains are important in any domain, healthcare systems are highly significant, with failures directly affecting patient outcomes.

Recent events have underscored the importance of strengthening HSCs. The COVID-19 pandemic brought unprecedented stress to medical supply networks globally [9]. During the pandemic’s peak, healthcare systems experienced severe stock shortages of personal protective equipment (PPE), ventilators, and essential pharmaceutical goods [10]. Demand for certain items surged drastically, while manufacturing and distribution were disrupted, leading to a shortfall of critical supplies in many regions. These shortages highlighted how vulnerable HSCs can be and the dire consequences of disruption: hospitals in many countries were unable to procure enough masks or oxygen equipment for their patients, illustrating that supply chain failures in healthcare can quickly become public health crises [11]. Even outside of global emergencies, HSCs face persistent challenges, such as drug shortages and supply interruptions. For example, an analysis by the U.S. Food and Drug Administration found that the single largest cause of drug supply disruptions under routine conditions was manufacturing quality failures, accounting for 62% of supply issues [12]. Such statistics emphasize that issues in production and logistics, whether due to quality control problems, capacity constraints, or other bottlenecks, can ripple through the HSC and ultimately jeopardize patient care. Consequently, there is a critical need for resilient and well-managed supply chain practices in healthcare to ensure that pharmaceuticals, devices, and equipment remain continuously available to those who need them [13].

A systematic synthesis of the existing research is therefore valuable to map the landscape of HSC studies and to identify major focus areas. To this end, the present study conducts a systematic literature review (LT) of the HSC domain, using a robust methodology to ensure comprehensive coverage and unbiased analysis. The manuscript followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to guide the search and selection process [14]. By adhering to this method, the LT searched through the Web of Science (WoS) database. This process yielded a final corpus of 705 contributions that met the criteria for inclusion in our analysis. These 705 peer-reviewed studies, published over several decades, form the basis of the evidence considered in this review.

This study enhances a traditional manual LT by incorporating natural language processing (NLP) techniques to identify thematic patterns in the HSC literature. Using an unsupervised machine learning (ML) approach, specifically latent Dirichlet allocation (LDA), the authors clustered 705 academic publications into topic-based groups, allowing for a data-driven exploration of dominant research themes [15,16,17]. In this context, the study can be framed within a data-driven analytical perspective that shares key characteristics with big data approaches. The dataset analyzed in this research includes 705 peer-reviewed articles, covering a wide range of topics, methodologies, and application domains within healthcare supply chains. This level of scale and heterogeneity makes traditional manual literature review approaches increasingly difficult to apply in a consistent and systematic way. The diversity of the corpus, combined with the continuous growth of scientific publications in this field, requires the adoption of automated and scalable analytical techniques capable of extracting meaningful patterns from large volumes of unstructured textual data. In this sense, natural language processing and machine learning methods provide a suitable framework to support the analysis. Techniques such as LDA topic modeling, TF-IDF analysis, and semantic exploration through KWIC and collocation analysis enable the systematic processing of complex textual datasets, reducing subjectivity and enhancing the reproducibility of the results. Therefore, the contribution of this study lies not only in mapping the literature, but also in demonstrating the use of data-driven analytical approaches to explore complex and heterogeneous research domains. To deepen the analysis, the study applied TF-IDF to highlight distinctive terminology within each cluster [18], and Keyword in Context (KWIC) analysis to interpret how critical terms such as “resilience” are used across different textual settings [19]. Additional n-gram and collocation analyses were conducted to surface commonly co-occurring phrases and domain-specific terminology [20].

Finally, a knowledge graph was built using Obsidian to visualize connections among key topics and concepts, showing the relational structure of the literature and revealing both central and peripheral areas of research [21]. This comprehensive NLP-enhanced review not only maps the current research landscape but also provides a strategic framework for future studies in HSC management.

By combining a systematic review methodology with advanced NLP-driven analysis, this study provides a comprehensive overview of the current state of research on HSCs. The introduction of thematic clusters and a knowledge graph is particularly aimed at helping researchers and decision-makers understand not only what has been studied, but also how the various research streams relate to each other. In the remainder of this paper, we detail the findings from our analysis and discuss their implications for both theory and practice. This study contributes to the healthcare supply chain literature by providing a comprehensive and data-driven mapping of the research field. Unlike traditional literature reviews, which typically focus on specific sub-domains or a limited number of contributions, this work analyzes a large corpus of 705 peer-reviewed articles to capture the overall structure of the domain. By adopting a systematic and scalable approach, the study aims to identify not only the main research themes but also the relationships among them, providing a broader and more integrative understanding of the healthcare supply chain landscape. Despite the growing body of literature on healthcare supply chain management, the field remains characterized by a high degree of fragmentation and a lack of conceptual integration across different research streams. The existing studies often focus on specific sub-domains—such as pharmaceutical logistics, hospital operations, or technological applications—without providing a comprehensive and system-level understanding of the field. This fragmentation limits the ability to identify overarching patterns, research gaps, and interconnections among different themes. In this context, the present study addresses the need for a more structured and data-driven mapping of the healthcare supply chain literature. By integrating a systematic review approach with natural language processing (NLP) techniques, the study aims to uncover latent thematic structures and provide a comprehensive overview of the domain. From a theoretical perspective, this contributes to a better conceptualization of the field by identifying its main research streams and their relationships. From a practical perspective, the study supports decision-makers by highlighting key challenges, gaps, and opportunities for improving the design and management of healthcare supply chains. Specifically, the review is structured to address the following key research questions:

• RQ1: How can the existing literature on healthcare supply chain management be systematically structured to identify the main thematic clusters within the field?
• RQ2: To what extent do these thematic clusters address the key challenges faced by healthcare supply chains, such as ensuring the availability of pharmaceuticals, medical devices, and electromedical equipment?
• RQ3: What gaps and limitations emerge from the current thematic structure of the literature, and how can these inform future research directions and practical improvements in healthcare supply chain management?

2. Methodology

This section outlines the fundamental methodologies employed to conduct the review and subsequent analyses. The initial phase of the research applied the PRISMA method to ensure a rigorous and well-structured review of the academic contributions related to HSCs. This systematic review complies with the PRISMA guidelines and was registered in PROSPERO, an international prospective register of systematic reviews in health and social care (CRD42024605761, “Supply chain in healthcare”). Following the systematic review, ML algorithms were implemented to support the analysis of the identified contributions, enabling a deeper exploration of emerging patterns and thematic structures within the literature.

2.1. Primary Sources Selection

The LT has been structured to ensure that sources are both pertinent and trustworthy to extract the contributions. The research achieved this by following the PRISMA method and exclusively utilizing the WoS database. The choice of the Web of Science database is motivated by its high-quality standards, structured indexing, and wide coverage of peer-reviewed journals, which ensures the reliability and consistency of the dataset. Figure 1 shows the flow diagram that describes the three macro-phases of the PRISMA method and the selected contributions, extracted, screened and included in the analysis. Adhering to the PRISMA method allowed us to adopt a systematic, transparent, and rigorous review for selecting studies that directly aligned with our research objectives, thereby ensuring their relevance. At the same time, the sole reliance on the WoS guaranteed that only peer-reviewed publications were included, confirming the reliability and academic rigor of the sources. This dual approach highlights the commitment to basing the review on literature that is not only directly related to the study, but also of high verified quality, ultimately enhancing the credibility and robustness of the findings. No restrictions were imposed in terms of time span, journal fields, or research areas in order to ensure the broadest possible coverage of the literature and to avoid introducing selection bias in the dataset. As a result, the final corpus includes contributions spanning multiple disciplines and application domains within healthcare supply chains, reflecting the heterogeneity of the field. To further enhance transparency and allow for a detailed inspection of the dataset, the full list of the 448 articles included in the analysis, together with their main characteristics, is provided in the Supplementary Materials.

This LT has been registered on PROSPERO, an international database that prospectively records systematic reviews in health and social care to ensure transparency and reduce duplication of efforts. The registration details are as follows: ID number: CRD42024605761, name of the dataset: “Supply chain in healthcare”, link to consult the dataset: https://www.crd.york.ac.uk/PROSPERO/view/CRD42024605761 (accessed on 13 December 2025). On 24 October 2024, a search was conducted on the WoS research engine to retrieve original articles and journal reviews in English, without imposing any time limits. This search strategy aimed to capture all possible variations in the relevant topics. In more detail, the query sought to identify all possible declinations of the terms “Supply Chain” and “Healthcare” in texts. The query is reported below:

TS = (“supply chain”) AND TS = (design) AND TS = (healthcare)

Specifically, the query was executed on the publication database to exclusively retrieve articles pertaining to the fields of supply chain, design, and healthcare. The search query was designed to balance specificity and coverage, ensuring the inclusion of relevant studies while limiting the introduction of unrelated records. The selected keywords reflect the core concepts of the research domain. The complete literature matrix can be consulted in the Supplementary Materials.

To ensure transparency and reproducibility of the systematic review process, the inclusion, eligibility, and exclusion criteria adopted in this study are summarized in

Table 1. Inclusion and eligibility criteria for the systematic review.

Criteria Type	Description
Inclusion Criteria	Peer-reviewed journal articles
Inclusion Criteria	Articles written in English
Inclusion Criteria	Studies indexed in the Web of Science (WoS) database
Inclusion Criteria	Studies addressing healthcare supply chains or related topics
Inclusion Criteria	Articles containing keywords related to “supply chain”, “design”, and “healthcare”
Inclusion Criteria	Availability of title and abstract for analysis
Eligibility Criteria	Articles providing sufficient information for thematic and textual analysis
Eligibility Criteria	Studies relevant to the research questions of the review
Eligibility Criteria	Articles suitable for NLP-based analysis (textual content available)
Exclusion Criteria	Conference papers, book chapters, editorials, and non-peer-reviewed documents
Exclusion Criteria	Articles not written in English
Exclusion Criteria	Studies not related to healthcare supply chain management
Exclusion Criteria	Articles lacking sufficient textual data (e.g., missing abstract)
Exclusion Criteria	Duplicate records

.

Table 2. Extract from the Excel file containing the ID and Abstract of the retrieved articles.

ID	Abstract
1	Healthcare and disaster supply chain have become a more important and popular research issues recently. However, only a few papers are known about the current issues both healthcare and disaster supply chain especially in natural disaster case. This paper is a preliminary report of research on healthcare and disaster supply chain. The paper intends to review and analyze several papers on above topic published during the last ten years. Published papers on healthcare and disaster supply chain research from 2005 and 2014 were classified into three main themes: (1) healthcare supply chain, (2) disaster supply chain and (3) healthcare supply chain in natural disaster. The topic issues in each main themes include operational management, information technology, inventory and control management, strategic management, and service management. Besides, the type of research methods contains empirical study, case study, modelling and simulation, literature review, and conceptual theory. Result of the review will provide the basis for the direction of future research in these three themes.
2	Healthcare remains a very crucial sector for every economy. Medicines or drugs serve as essential consumables for the treatment of ailments and hence making them essential commodities in healthcare delivery. The hospital pharmaceutical supply chain of most emerging economies has not been given the requisite attention in relation to enhancing visibility among the stakeholders. The lack of visibility within the supply chain leads to drug shortages in hospitals. Some studies have identified the problem of visibility in pharmaceutical supply chains. However, a few of these studies tend to offer solutions to deal with this crucial issue. From a design science perspective, the current study relies on information systems research framework to design an architecture for a class of systems aimed at enhancing supply chain visibility and in effect help to mitigate the issue of drug shortage resulting from ineffective supply chain management.
3	Supply chain management in healthcare is evaluated with a particular focus on the distribution of medicines from a wholesaler to clinics. Currently, there are issues with service levels to clinics that need addressing. The value of the paper arises from providing a detailed analysis of a healthcare supply chain in the developing world and diagnosis of the parameter involved in inventory.
4	This study was designed primarily to assess the relationships among the implementation of e-commerce, the external collaboration and supply chain performance in the healthcare industry of Taiwan. To examine the impact of e-commerce on SCM in the healthcare industry, its relationship with external collaboration and supply chain performance were empirically tested. The statistical results of analyses showed that the implementation of e-commerce, the external relationship between trading partners and supply chain performance were highly correlated to each other. Supply chain performance is also highly correlated with customer satisfaction.
…	…

shows the obtained results, which were exported in XLS file format to facilitate subsequent data manipulation to create a literature matrix for further analysis. The initial step involved preparing the matrix for processing; to achieve this, a unique ID field was added to ensure direct and unambiguous referencing of the articles during data manipulation. Subsequently, an Excel file was created containing a table with two main columns: ID and Abstract.

The initial extraction consisted of 705 articles. Rather than processing the full content of each article, the proposed methodology focused the first analysis exclusively on the abstracts and titles, as these were considered to provide a sufficiently comprehensive overview for the intended analysis. This strategy streamlined the analytical process while preserving a clear and adaptable framework for subsequent developments. The adoption of the PRISMA approach enhances the transparency and reproducibility of the study, providing a structured and traceable selection process that reduces potential biases in the dataset construction. To facilitate further processing, each row in the Excel file was converted into a Markdown-formatted text file. Figure 2 depicts the evolution over time of scholarly publications referring to the “supply chain in healthcare”.

The graphical data indicates that discussions of HSCs have increased since 2015, likely due to the Ebola outbreak in Africa that emerged during this period and lasted until 2019. From 2020 onward, research in this field has surged significantly, with the exception of 2022, which saw a slight decline compared to the previous year. The objective was to identify the main thematic categories addressed in the analyzed articles as a foundation for applying the PRISMA method. Initially, the following thematic categories were defined: (i) drugs, (ii) medical devices, and (iii) electromechanical equipment.

2.2. Analysis Using ML Algorithms to Analyze the Data Extracted in the LT

Therefore, after the descriptive analysis of the obtained paper extracted in the LT, the following section provides a brief description of the ML algorithms used to analysis the extracted data for the PRISMA method. Figure 3 depicts the three macro-phases of the proposed method adopted for the systematic review and the subsequent textual analysis of the LT on HSCs.

The initial dataset was structured in an Excel file containing article content and unique identifiers. A Python-based script (3.14.3 version) was developed to convert each row into a Markdown-formatted file, preserving the article ID (Appendix A). These Markdown files were then transformed into plain text to prepare them for natural language processing tasks. Preprocessing routines included the removal of stop words, punctuation, special characters, and isolated numbers, along with text normalization through lowercasing [22,23]. The preprocessing phase is a critical step in text analysis, as it reduces noise and ensures that the extracted features accurately represent the semantic content of the documents, improving the reliability of the subsequent analysis. A modular codebase was developed to automate the processing pipeline. This included the extraction and transformation of text data, keyword tagging through pattern matching with a predefined lexicon, and keyword extraction using cosine similarity-based vector techniques. Additional modules were employed to generate a term frequency–inverse document frequency (TF-IDF) matrix, execute LDA topic modeling, and determine the optimal number of clusters using the elbow method. Textual data in plain text format were further examined using a specialized corpus analysis tool. This enabled detailed linguistic investigation through Keyword in Context (KWIC), n-gram frequency analysis, and collocation analysis. These functions were configured with adjustable context windows and case sensitivity settings to capture subtle patterns and semantic relationships [24]. The TF-IDF technique was employed to quantify the importance of terms within each document relative to the overall corpus. This involved computing term frequency within a document and scaling it by the inverse frequency across all documents. The resulting matrix was used to support clustering and to identify key distinguishing terms across the thematic landscape [25].

Therefore, LDA was used to uncover latent topics in the document corpus by modeling each document as a mixture of topics, and each topic as a distribution over words. An iterative algorithm inferred the underlying probabilistic structure, enabling each document to be characterized by a set of weighted keywords. The outputs were visualized using intertopic distance maps and word clouds to enhance interpretability [26,27]. To determine the optimal number of clusters for topic grouping, the elbow method was applied. This unsupervised learning approach involved plotting the sum of squared errors (SSE) against various cluster counts. The “elbow point” in the graph where adding further clusters yielded diminishing returns in error reduction was identified as the ideal number of clusters, which in this case was six [28]. To support exploration and visual analysis, a Markdown-based note management system was used. This software enabled graph-based visualization, where articles were represented as nodes and thematic tags as connecting edges. This visual mapping allowed for intuitive exploration of relationships among the documents [8,29,30,31]. Although more advanced optimization techniques, such as meta-heuristic algorithms, can be employed for cluster selection, the choice of the elbow method in this study is consistent with the exploratory nature of the analysis and the characteristics of the dataset. The primary objective of this research is not to achieve a mathematically optimal partition of the dataset, but rather to identify a meaningful and interpretable thematic structure of the literature. In this context, the elbow method provides a transparent and widely adopted approach that allows for a clear balance between model complexity and interpretability. Furthermore, the selection of the number of clusters was not based solely on the elbow criterion. Additional validation steps were considered, including coherence measures and qualitative evaluation of topic interpretability. The results obtained with six clusters showed a consistent and semantically meaningful representation of the literature, avoiding both excessive fragmentation and oversimplification.

More complex optimization approaches, such as meta-heuristics, are typically more suitable for large-scale numerical optimization problems where the objective is to minimize a specific cost function. However, in text mining applications such as topic modeling, interpretability and semantic coherence are equally important criteria. Therefore, the combination of the elbow method with coherence evaluation and qualitative assessment was considered appropriate for the purposes of this study.

This approach ensures that the selected number of clusters is not only supported by quantitative indicators but also aligned with the interpretability and practical relevance of the results.

An unsupervised learning technique, the elbow method, was applied to determine the optimal number of clusters for the document set [28,32]. A plot was constructed with the number of clusters on the x-axis and the sum of squared errors on the y-axis. The point at which additional clusters did not yield a substantial decrease in error was identified as the optimal cluster count. From the graph, it can be seen that according to the elbow method, the ideal number of clusters is 6 (c.f. Figure 4).

The analytical workflow is structured as a multi-stage pipeline. The input data consist of bibliographic records retrieved from the Web of Science database, including titles and abstracts of the selected articles. These data are first pre-processed through cleaning procedures, including stopword removal, normalization, and text transformation into plain text format. Subsequently, keyword extraction and tagging procedures are applied, supported by cosine similarity and TF-IDF techniques. Topic modeling is then performed using LDA to identify latent thematic structures within the corpus. The use of LDA enables the identification of latent thematic structures in an unsupervised manner, allowing the analysis to capture underlying patterns in the literature without imposing predefined categories, thus enhancing the objectivity of the results. Additional corpus analysis methods, such as KWIC, n-gram, and collocation analysis, are employed to refine the interpretation of results. The outputs of the analysis include thematic clusters, keyword rankings, intertopic distance maps, word clouds, and a knowledge graph representation, which collectively provide both quantitative and qualitative insights into the structure of the literature. From a methodological perspective, this study introduces an integrated analytical framework that combines a PRISMA-based systematic literature review with natural language processing (NLP) techniques. The proposed pipeline leverages multiple tools, including LDA topic modeling, TF-IDF analysis, KWIC, n-gram and collocation analysis, as well as knowledge graph visualization, to enable a data-driven exploration of the literature. The research design combines systematic literature review and data-driven analytical techniques, providing a robust and reproducible framework for exploring complex research domains. This approach enhances the reproducibility and scalability of the analysis, allowing the systematic identification of latent thematic structures while reducing the subjectivity typically associated with manual literature reviews.

3. Results

This LT employed two distinct NLP methods. One approach was based on the principles of LDA, a probabilistic topic modeling technique designed to extract latent thematic structures from textual data. The other approach involved a tagging strategy implemented through software that creates and edits Markdown files, facilitating the construction of a knowledge graph to visualize and organize key concepts. The two methodologies were integrated in the development of a TF-IDF table, which provided a quantitative basis for comparing term frequency and relevance across the corpus.

Before initiating text manipulation, preliminary operations were executed to guarantee that the data were free from errors and extraneous information. A PowerShell script was implemented to remove all pre-existing tags from the Markdown (.md) files. The removal process ensured that only tags defined during the current study were retained. Special care was taken to preserve the text immediately following each tag to maintain potentially significant keywords. A separate PowerShell script was then utilized to extract statistical data concerning text length distribution. The analysis produced the following results: an average of 240 words per file, a mode of 251 words, a maximum of 1391 words, and a minimum of 61 words per file. These statistics provided critical insight into the variability of text sizes and informed subsequent text analysis processes, such as tokenization. A Python script was subsequently employed to conduct text pre-processing. Its functions included the removal of English stop words, the elimination of punctuation and isolated numbers, and the conversion of all text to lowercase. These steps were essential in reducing noise and ensuring that the textual data were suitably refined for further analysis. For example, failure to pre-process the text would result in the inclusion of high-frequency words (e.g., “what”, “he”, “the”) that lack semantic significance and could distort the analytical outcomes. Two distinct types of analysis were then performed on the pre-processed texts, allowing for a comparative evaluation of the results.

3.1. Tagging for Visualization Using Markdown-Based Tools

The tagging method comprises a series of iterative procedures designed to incorporate meaningful tags into a corpus of texts, with the objective of automating the annotation process as extensively as possible. Following the elimination of common English stop words, the Markdown files were imported into a dedicated text management platform to verify that the dataset was free of unintended tags or inadvertent linkages. This preparatory step ensured that only deliberately assigned annotations were present, thereby enhancing the precision and reliability of subsequent tagging operations (Figure 5).

Examination of the resulting graph in Figure 5 revealed no edges connecting the nodes, confirming that the files are devoid of unintended associations and are suitably prepared for subsequent processing via the tagging procedure. A Python-based keyword extraction routine was applied to the pre-processed files, ensuring that no potentially relevant clusters were omitted from the analysis (c.f.,

Table 3. Result of the extraction of keywords.

Text ID	Keyword	Score
1	disaster supply chain	0.78
1	chain disaster supply	0.72
1	supply chain disaster	0.66
1	healthcare disaster supply	0.65
1	supply chain healthcare	0.62
2	pharmaceutical supply chain	0.65
2	pharmaceutical supply chains	0.63
2	visibility pharmaceutical supply	0.63
2	hospital pharmaceutical supply	0.59
2	problem visibility pharmaceutical	0.57
3	healthcare supply chain	0.80
3	analysis healthcare supply	0.68
3	distribution medicine wholesaler	0.68
3	medicine wholesaler clinics	0.67
3	supply chain	0.66
4	commerce external collaboration	0.56

). The routine was configured with specific parameters to optimize the extraction process. The parameter controlling the range of words per key phrase was set to allow one to three words. This configuration was chosen because, in anticipation of frequent multi-word terms (such as “supply chain” and “supply healthcare”), limiting extraction to two-word phrases could result in the omission of significant information. Allowing for three-word phrases increased the likelihood that at least one substantive word would be captured within each key phrase. A further parameter was established to extract five key phrases per document. This decision followed an analysis of the average and mode of word counts across the articles, which indicated that extracting five phrases per file provided a manageable data volume while reducing redundancy in the conceptual representation of each document. The resulting output yielded a balanced and informative set of key phrases that served as the basis for subsequent clustering and thematic analysis.

The generated resulting list was manually examined to assign a reference topic to each extracted keyword. Keywords deemed irrelevant or off-topic were excluded from further clustering. A Python-based cosine similarity tool was then employed to streamline the manual assignment process by identifying semantically related terms within the dataset. For instance, during the review of extracted keywords, the term “vaccine” was recognized as a potential cluster candidate. Cosine similarity analysis subsequently identified ten synonymous terms related to “vaccine” within the dataset. These synonyms were cross-referenced with the original list of keywords using Excel’s filtering functions, ensuring accurate labeling of the keywords pertinent to the “vaccine” cluster. Figure 6 depicts this combined approach, integrating manual evaluation with automated similarity analysis, which enhanced the dataset’s refinement and improved the precision of the clustering process (c.f.,

Table 4. Association of keywords with general clusters.

Text ID	Keyword	Score	Cluster
1	disaster supply chain	0.78	Disaster
1	chain disaster supply	0.72	Disaster
1	supply chain disaster	0.66	Disaster
1	healthcare disaster supply	0.65	Disaster
1	supply chain healthcare	0.62	Design
2	pharmaceutical supply chain	0.65	Pharmaceutical
2	pharmaceutical supply chains	0.63	Pharmaceutical
2	visibility pharmaceutical supply	0.63	Pharmaceutical
2	hospital pharmaceutical supply	0.59	Pharmaceutical
2	problem visibility pharmaceutical	0.57	Pharmaceutical

).

A graph view was utilized on the note management platform with label filtering enabled, allowing evaluation of the inserted tags and the interconnections among the files. Figure 6 shows an in-depth analysis of the resulting network highlighted the prominence of several topics, including blockchain, logistics, pharmaceutical, vaccine, and medical equipment. The cluster labeled as design was observed to be overly indexed, rendering it less informative for the purposes of this study, while other clusters appeared in only a few documents due to their highly specific nature.

The described approach enabled an initial clustering that provided sufficient detail to identify key terms and the corresponding topics present in the corpus. Although the classification remained somewhat rudimentary, it established a foundation for refining the analysis further by pinpointing the terms that warrant closer attention and by offering insight into the overall thematic distribution within the texts. Subsequent to this evaluation, a second iteration of the method was undertaken. This phase focused on enhancing the precision of cluster identification using AntConc (4.3.1 version), a corpus analysis software, to fine-tune the search for clusters and improve the accuracy of the thematic categorization.

3.2. Corpus Analysis

A high frequency of non-informative keywords, such as “supply”, “chain”, and “healthcare”, was observed in the initial extraction. This observation prompted a second analysis on texts that had been further refined to remove both generic stopwords and additional low-utility words, as prepared before the second iteration of the LDA analysis. A subsequent word count analysis on these refined texts provided the following statistics: an average of 110 words per file, a mode of 98 words, a maximum of 251 words, and a minimum of 18 words per file. These figures informed the adjustment of parameters in a Python-based keyword extraction script. The configuration was modified to extract key phrases composed of one to two words (keyphrase_ngram_range = (1,2)) and to select the top three key phrases per document (n = 3). The resulting Excel file, which contained the newly extracted keywords, was then processed to cluster only those keywords with a score greater than 0.5. This threshold was set to exclude terms that exhibited low semantic alignment with the corresponding texts. A detailed analysis of the refined keywords followed, aimed at interpreting their semantic significance and determining whether they should be subdivided into subcategories or merged with previously identified clusters. The functions for collocate and n-gram analysis provided by AntConc were employed to examine the contextual relationships and co-occurrence patterns among the keywords, thereby supporting a more precise interpretation of their relevance. Analysis of the data revealed that the term “design” frequently appears alongside words such as “network”, “methods”, “green”, “science”, “efficient”, and others. This pattern indicates that “design” functions as a highly generic and polysemous term, thereby reducing its effectiveness for precise and meaningful categorization.

A Keyword in Context (KWIC) analysis was conducted to further examine the usage context of these ambiguous terms (Figure 7). This approach provided detailed insights into the various semantic environments in which “design” and similar words are employed, facilitating a more accurate interpretation of their relevance and application within the corpus. Figure 7 presents the KWIC analysis by illustrating how the hit feature “design” appears across various textual sources. Therefore, entries from 598.txt predominantly associate “design” with academic programs in industrial engineering, emphasizing its role in senior capstone projects, instructional methods, and curriculum development. Terms such as “course”, “engineering”, “capstone”, and “senior” frequently appear, indicating a strong link between design and structured learning outcomes. In contrast, entries from other files (e.g., 68.txt, 114.txt, 615.txt) show a shift toward more technical domains, with phrases like “science describes design”, “evaluated methodology”, and “initiated documenting business case”, reflecting design’s relevance in research and applied science. This distribution reveals the polysemous nature of “design”, which bridges pedagogical contexts and domain-specific innovation, and underscores the importance of contextual analysis for disambiguating key terms in text mining studies.

Therefore, to validate the results and guarantee a robust analysis of the LT, a second iteration was iterated. Then, the keywords and their associated clusters were extracted. A simple macro, written in a scripting language, was employed to expand the table by automatically generating the inverse pair for each keyword combination while preserving the original cluster assignment.

Table 5. Excerpt of the keywords table with associated clusters.

Keyword	Score	Cluster
additive manufacturing	0.55	Additive manufacturing
manufacturing additive	0.55	Additive manufacturing
printing	0.55	Additive manufacturing
ai	0.51	AI
deep learning	0.55	AI
learning deep	0.55	AI
deep reinforcement	0.58	AI
reinforcement deep	0.58	AI
optimizer mogwo	0.51	AI
mogwo optimizer	0.51	AI
analytics	0.52	Analytics
cps	0.56	Big data
dea	0.61	Big data
digital twin	0.56	Big data

shows a bidirectional mapping that enhances the dataset’s structural integrity and facilitates further semantic analysis.

A PowerShell-based tagging process was re-executed using the newly generated Excel file as input and a working directory containing Markdown files that had been refined to remove stopwords and low-utility terms. This procedure produced an updated knowledge graph that reflects the enhanced interconnections among the documents based on their associated keywords. Figure 8 presents an in-depth analysis of the resulting network, highlighting the prominence of several key topics in the second iteration, including blockchain, logistics, pharmaceuticals, vaccines, and medical equipment. The cluster labeled design was found to be overly indexed, making it less informative for the purposes of this study. Additionally, some clusters appeared in only a few documents due to their highly specific or niche nature.

Compared to the first iteration, the second benefited from a more refined text preprocessing phase, which included the removal of generic terms and the application of stricter semantic filtering using TF-IDF thresholds and KWIC analysis. As a result, the clustering outcomes in the second iteration were clearer and more meaningful, with distinct thematic groups emerging around prominent topics such as blockchain, logistics, and vaccines.

Moreover, a Python script was executed on the processed files to evaluate the distribution of tags throughout the corpus by computing the TF-IDF matrix. The standard TF-IDF calculation, which multiplies term frequency (TF) by inverse document frequency (IDF), was modified with a scaling factor of 1000. Figure 9 introduces the multiplier enhanced the visibility of differences in tag scores, thereby facilitating a more robust comparison of the relative significance of tags across the dataset. The TF-IDF analysis provides insights into the most relevant terms within the corpus, highlighting the concepts that are most characteristic of the healthcare supply chain literature. Rather than representing isolated values, the identified terms reflect the prominence of specific topics and research directions within the field.

In particular, the presence of terms related to logistics, supply, and healthcare operations confirms the central role of operational management in the literature. At the same time, the distribution of terms suggests a strong focus on efficiency and resource allocation, while more strategic and integrative aspects, such as system-wide coordination and sustainability, appear less prominent.

Therefore, the TF-IDF results contribute to understanding the thematic priorities of the literature and support the identification of dominant research areas, as well as potential gaps that are further explored in the subsequent analysis.

Figure 10 illustrates an initial round of topic modeling conducted on pre-processed texts using LDA. Visualization techniques included an intertopic distance map (IDM) and word cloud representations. The IDM provided a clear depiction of the separation or overlap among detected topics, while the word cloud method enabled the immediate identification of key terms associated with each topic.

A more comprehensive examination of the texts was performed to refine the dataset for effective categorization. Markdown files were converted to plain text using a Python script, facilitating compatibility with corpus analysis software. The resulting corpus was imported into AntConc, which was employed to extract the 500 most frequent words using its word frequency function. A manual review then reduced this list to 100 terms by eliminating words deemed non-informative for the analysis. A subsequent Python-based process identified and incorporated synonyms for these selected words, resulting in a final exclusion list of 1000 words. A further Python script applied this refined exclusion list to the Markdown files that had already undergone stop word removal, thereby enhancing the dataset for subsequent analyses.

At this stage, the pre-processed files were prepared for another round of topic modeling. An analysis of the intertopic distance map from the initial run revealed that the spatial distribution of topic clusters suggested the presence of six distinct clusters. Topic modeling using LDA was performed again on the refined dataset, this time setting the number of topics to six. This adjustment aimed to achieve clearer thematic distinctions within the corpus and enhance the interpretability of the results shown in Figure 11.

The refined analysis produced more distinct topic clusters. Figure 12 depicts the word cloud visualizations associated with each cluster, displaying key terms that were representative of the underlying themes. An initial experiment with ten topics had been conducted to verify that no significant latent concepts were overlooked; the broader topic set helped to identify any potentially relevant ideas that might have been missed. However, the considerable overlap among topics in that trial indicated that reducing the number of topics was necessary for improved clarity.

A Python-based implementation of the elbow method was then applied to validate the optimal cluster count. The graph generated, consistent with the methodology outlined in previous sections, confirmed that six clusters were ideal. The plot of the sum of squared errors against the number of clusters exhibited an “elbow” at six, thereby supporting the decision to adopt six topics. As mentioned, the LDA model was implemented using the Gensim library, with the number of topics set to six based on the exploratory analysis supported by the elbow method and interpretability considerations. The model was trained using multiple passes over the corpus (passes = 15) to ensure the convergence and stability of the results.

In order to assess the validity of the topic modeling results, both quantitative and qualitative criteria were considered. From a quantitative perspective, model performance was evaluated through coherence measures, which assess the semantic consistency of the topics, and perplexity, which provides an indication of the model’s predictive performance. While perplexity tends to decrease with increasing model complexity, coherence scores were used as a more interpretable metric to support the selection of the final number of topics.

From a qualitative perspective, the interpretability of the topics was carefully evaluated by examining the most relevant terms associated with each topic and their consistency with the domain of healthcare supply chain management. The selected configuration (six topics) provided a meaningful and coherent representation of the thematic structure of the literature, balancing statistical validity and interpretability.

4. Discussion

The identification of six clusters reflects a balance between analytical robustness and the interpretability of the results. The selection was supported by the elbow method, which indicated a point of diminishing returns at six clusters, suggesting that additional clusters would not significantly improve the explanatory power of the model. At the same time, the six-cluster configuration enabled a clear and meaningful interpretation of the thematic structure, avoiding excessive fragmentation or overlap among topics. From an analytical perspective, the clustering is based on the representation of documents through term-weighted vectors derived from the corpus. Specifically, textual data were transformed into a document-term matrix using TF-IDF weighting, where each term represents a variable contributing to the semantic characterization of the documents. This approach allows the identification of latent patterns in the literature without the need for predefined variables, as the relevant features emerge directly from the data through statistical significance and frequency-based filtering. The resulting clusters therefore reflect underlying thematic structures in the literature, where each topic is characterized by a set of co-occurring terms. The choice of six clusters proved to be effective in capturing the main research areas while maintaining a level of abstraction that supports a comprehensive interpretation of the healthcare supply chain domain. The results of the analysis provide a system-level interpretation of the healthcare supply chain literature, revealing structural patterns that are not easily identifiable through traditional review approaches. In particular, the identification of thematic clusters highlights the presence of fragmentation and limited integration across different research streams, as well as a gap between technological innovation and practical implementation. A deeper interpretation of the results enables the identification of key challenges that characterize the healthcare supply chain domain as emerging from the clustering and NLP-based analysis. These challenges do not arise from isolated contributions, but rather from the aggregated patterns observed across the corpus, reflecting structural issues in the literature, and by extension, in the systems it represents. The analysis highlights a persistent lack of integration across supply chain tiers, as research tends to focus on specific segments such as manufacturing, distribution, or hospital operations without adequately addressing end-to-end coordination. This tendency is reflected in the weak connectivity between clusters and in the absence of cross-cutting themes capable of linking different parts of the supply chain into a coherent system-level perspective. At the same time, a clear misalignment emerges between technological innovation and practical implementation, with topics related to digitalization, blockchain, and advanced analytics being widely discussed at a conceptual level but rarely connected to operational or managerial contexts. The clustering results show that technology-oriented studies often remain isolated from research streams focused on implementation, suggesting a gap between innovation and real-world application. The issue of resilience also appears as a critical yet insufficiently structured dimension, as resilience-related concepts are dispersed across multiple clusters without forming a cohesive thematic area, indicating the absence of a unified framework capable of capturing interdependencies across the entire system. Similarly, sustainability emerges as a marginal and fragmented topic, with environmental and social aspects receiving limited and inconsistent attention, often treated as secondary considerations rather than as integral components of supply chain design and management. The analysis further reveals challenges related to data sharing and interoperability, as digital tools and data-driven approaches are frequently addressed in isolation, without a systemic perspective on how information should be integrated and exchanged among different actors, thereby limiting visibility and coordination. In addition, regulatory complexity appears as a cross-cutting but unevenly explored issue, with a strong focus on pharmaceutical regulations and comparatively limited attention to other domains such as medical devices and broader healthcare logistics, resulting in a partial and fragmented understanding of regulatory constraints. Overall, these patterns depict a research landscape characterized by fragmentation, limited integration, and an imbalance between conceptual developments and practical application, highlighting the need for more systemic, interdisciplinary, and implementation-oriented approaches to healthcare supply chain management.

To clarify how research gaps have been derived from the clustering and thematic analysis, it is important to specify the analytical criteria adopted in this study. The identification of research gaps is not based solely on the frequency of keywords or the size of clusters, but rather on a combined interpretation of quantitative and qualitative indicators emerging from the analysis.

First, the relative weight of themes was evaluated through TF-IDF scores and topic distributions. Themes characterized by low frequency or weak representation across the corpus were interpreted as underexplored areas, suggesting potential gaps in the literature. However, frequency alone was not considered sufficient to define a research gap.

Second, the analysis examined the structural relationships among clusters through the knowledge graph and topic modeling outputs. The absence of connections between clusters, as well as the lack of bridging concepts (e.g., resilience, interoperability, or integration), was interpreted as evidence of fragmentation. In this sense, research gaps emerge not only from what is missing, but also from the limited integration among the existing research streams.

Third, a semantic analysis based on KWIC, n-grams, and collocation techniques was used to understand how key concepts are contextualized within the literature. Concepts that only appear in narrow contexts or are weakly connected to operational or managerial applications were identified as areas where further development is needed. Research gaps were interpreted in relation to the alignment between technological, operational, and strategic dimensions. When clusters showed a strong focus on conceptual or technological aspects without corresponding evidence of implementation or practical integration, this was considered indicative of a gap between theory and practice. Therefore, research gaps in this study are identified through a multi-dimensional approach that combines thematic relevance, structural connectivity, semantic interpretation, and practical applicability, rather than being based solely on the number of occurrences of specific keywords or themes. These findings contribute to a more comprehensive understanding of the field, supporting the identification of research gaps and informing future directions for both academic research and practical applications.

4.1. Implications for Theory

Based on the criteria described above, the following section discusses the main implications and research gaps emerging from the analysis. The HSC literature is vast and varied. It spans diverse topics, from detailed inventory logistics to high-level policy and governance issues. This broad scope spanning products, functions, and regions makes it difficult to achieve thematic coherence. For example, one study might analyze pharmaceutical inventory management in a major hospital, while another examines rural clinic supply strategies in a developing region. These studies rarely cite each other, reflecting the lack of integration across contexts. Rather than contributing to a unified discourse, the literature is fragmented into isolated thematic areas. The conceptual mapping confirms this pattern: research themes form distinct silos with minimal overlap. Even without formal cluster analysis, one can sense the separation. Some research clusters share similar topics or methodologies, but connections between them remain weak. Each cluster functions independently, with limited interaction across themes. For instance, the literature on pharmaceutical supply chains clusters tightly around quality control, regulation, and manufacturing issues. Those papers use common keywords (compliance, drug quality, safety standards, regulation) and often cite the same regulatory guidelines. In contrast, papers on medical device supply chains emphasize innovation and technology adoption. These device-focused studies discuss topics like new design methods, sensor integration, and Internet-of-Things applications. They cite largely different sources, emphasizing Industry 4.0 themes rather than drug quality. This highlights the lack of cross-disciplinary engagement among researchers focused on different product categories. Each major product category and context appears to occupy its own research niche. The literature on vaccines, for example, focuses on immunization logistics and cold chains, separate from general drug or device supply studies. Likewise, emerging areas like telemedicine or personalized medical devices appear only in a few papers. As a result, these emerging topics remain on the periphery of the research landscape. Generic operational themes (logistics planning, inventory control, supply reliability) form their own clusters as well. One strand of work focuses on distribution network optimization, and another focuses on inventory modeling. These papers use abstract language (routes, warehouses, lead times, safety stock) without being tied to specific items. For example, a logistics model may discuss minimizing travel distance without specifying whether it moves vaccines or implants. Similarly, an inventory policy paper might propose an ordering model without addressing factors like drug expiration or device obsolescence. These operational clusters run parallel to the product-specific ones. They address general supply chain concepts, but they do not bridge the gaps between product domains. This split in the literature highlights a core disconnect between innovation and implementation. Emerging technology topics are prominent, but they stand apart from practical supply chain discussions. In our mapping, blockchain and other digital solutions emerge as a major theme. Many papers discuss blockchain’s potential for traceability in healthcare; these form a clear cluster of innovation research. However, the literature contains few studies on how to actually use these technologies in real-world healthcare systems. We found little on how hospitals would integrate blockchain into the existing IT infrastructure or how regulations would adapt. Instead, the blockchain papers mostly cite other tech-focused works. In effect, researchers treat cutting-edge solutions as conceptual ideas, largely disconnected from operational reality. Tools like AI and the Internet of Things appear only tangentially. Mentions of ML, predictive analytics, or smart sensors are scattered across different papers, but they did not coalesce into distinct themes [33,34]. There is no dedicated AI cluster, for example, despite many authors suggesting its promise. This uneven prioritization suggests that some innovations dominate the discourse while others receive scant attention, which could reflect hype cycles or novelty bias in publication. Similarly, proven technologies like RFID tracking or lean management get surprisingly little focus. In summary, certain trendy technologies (e.g., blockchain) are heavily studied, while other potentially valuable innovations remain underexplored. Regulatory and policy issues are handled unevenly. The knowledge graph shows that regulatory themes are strongly tied to drug-related clusters: studies on pharmaceuticals frequently mention FDA, EMA, and compliance details. In contrast, discussions of device approvals or hospital accreditation are nearly absent in other clusters, even though they also affect supply chain operations. It is as if regulations only matter for medications. This siloing is a structural weakness. HSCs everywhere operate under strict rules, but the literature does not present an aligned view of these constraints. For example, no single theme emerged to cover medical device safety standards or international trade regulations; each is treated in isolation. The result is a piecemeal discussion of policy that fails to connect across the system. Environmental sustainability concerns are similarly fragmented. Topics like waste reduction, energy efficiency, and carbon footprint do not form cohesive themes in the healthcare literature. The analysis found that terms associated with sustainability appear only marginally. For example, the word “green” surfaced mainly as part of a generic design cluster, not in any dedicated environmental topic. This suggests that sustainability is treated as a side note. One paper might mention a recycling initiative in passing, but the main focus remains efficiency or cost. In short, the literature lacks an integrated focus on sustainable supply chain practices in healthcare. Even when environmental issues are addressed, they are largely appended to other themes rather than addressed holistically. Terminological ambiguity further obscures the landscape. We observed that fundamental terms can be too generic to differentiate topics. Words like “supply”, “chain”, and “healthcare” appear in nearly every document, yielding no specific thematic signal. Such high-frequency keywords were removed from consideration. Even after filtering them out, other broad terms remain problematic. For instance, the word “design” appears in many contexts (network design, product design, process design). Because “design” is so polysemous, it cannot anchor a clear topic. Such semantic noise means two papers might use similar generic words but actually discuss different problems or use different words for the same issue. For example, one study might refer to “demand forecasting” while another refers to “demand estimation”, and these end up separated in the analysis. These linguistic inconsistencies hamper the clustering process, causing key ideas to be scattered across multiple clusters.

The concept clusters themselves mirror the fragmentation. The knowledge graph shows a few central hubs and many isolated peripheries. Central hubs include topics like blockchain, logistics optimization, pharmaceutical distribution, vaccine delivery, and medical equipment supply. Each hub corresponds to a well-defined theme with many publications. For example, the logistics optimization hub contains dozens of papers on routing, warehousing, and inventory management, forming a dense subnetwork. The pharmaceutical hub contains many papers on drug manufacturing and quality control. These hubs accumulate internal citations and form the backbone of the research landscape. By contrast, the periphery contains only a few papers on niche subjects. The “design” cluster itself was down-weighted due to its overly broad nature. Other niche clusters had only a couple of documents each, often on very narrow case studies. For instance, separate small clusters might cover “supply chain risk in a single hospital” or “customized prosthetics manufacturing”. Because these topics are so specialized, they have not merged into mainstream discourse. They highlight gaps in the literature, areas that have attracted very little research attention. The absence of bridging topics is also notable. One might expect overarching themes like resilience or interoperability to connect these areas, but none clearly emerged. Instead, related concepts are scattered [35,36]. For example, “resilience” appears in some disaster-preparedness studies and a few on production planning, but no unified resilience cluster formed. Similarly, interoperability and data-sharing appear in isolated informatics papers rather than as major supply chain topics. This lack of cross-cutting themes underscores how disjointed the field is. Each fragment remains loosely connected to the rest, if at all.

4.2. Implications for Practitioners and Policymakers

At the system level, these patterns manifest as operational miscoordination. HSCs are inherently multi-tiered and interdependent. They involve manufacturers, distributors, hospitals, pharmacies, regulators, and more, all linked in complex networks. Yet the research largely ignores end-to-end integration. Nearly every study focuses on a single tier or segment. For example, one paper might optimize hospital pharmacy orders, while another model’s pharmaceutical distribution. Very few papers connect those stages. Important cross-level processes are under-studied. For instance, aligning inventory policies across multiple tiers did not emerge as a unified topic. Synchronizing supply and demand across organizations is not addressed by any major theme. In essence, the literature treats each link in the chain separately, rather than as parts of a coordinated system. This lack of coordination has real implications. Efficiencies discovered in one silo may not improve the whole system. For example, a distribution center could optimize its delivery routes, but if hospitals continue ordering independently, system-wide performance remains suboptimal. Many optimization models in the literature treat inputs from other segments as fixed parameters, ignoring interdependence. Similarly, risk analyses often focus on a single event (such as a factory shutdown) without connecting to the broader chain or patient outcomes. Without a holistic view, improvements in one area do not automatically yield benefits elsewhere. Based on the findings of the analysis, a set of high-level recommendations can be derived to address the main gaps identified in the healthcare supply chain literature.

Table 6. High-level recommendations for healthcare supply chain management based on the literature analysis.

Area	Identified Gap	Recommendation	Expected Impact
System Integration	Fragmentation across supply chain tiers and lack of end-to-end coordination	Develop integrated, multi-tier supply chain models that connect manufacturers, distributors, and healthcare providers	Improved coordination, reduced inefficiencies, and enhanced system resilience
Digital Technologies	Strong focus on conceptual technologies (e.g., blockchain) but limited real-world implementation	Promote applied research and pilot projects for integrating digital solutions into healthcare systems	Increased adoption of digital tools and improved traceability
Resilience	Lack of unified approach to resilience across different supply chain stages	Develop cross-cutting resilience frameworks that consider the entire supply chain	Enhanced preparedness for disruptions (e.g., pandemics, crises)
Sustainability	Sustainability is marginal and not treated as a core theme	Integrate environmental sustainability into supply chain design and decision-making processes	Reduced environmental impact and alignment with global sustainability goals
Data and Interoperability	Limited integration of data-sharing and interoperability across actors	Develop standardized data-sharing protocols and interoperable information systems	Improved visibility and decision-making across the supply chain
Regulatory Alignment	Regulatory aspects are addressed mainly for pharmaceuticals, not across all domains	Extend regulatory analysis to include medical devices and broader healthcare logistics	Improved compliance and reduced operational risks
Research Integration	Strong thematic silos with limited cross-disciplinary research	Encourage interdisciplinary studies combining logistics, policy, and technology perspectives	More comprehensive and systemic solutions
Implementation Focus	Gap between theoretical models and practical applications	Increase collaboration between academia and practitioners to test models in real contexts	Greater applicability and impact of research findings

summarizes these recommendations, linking the key challenges to potential directions for both research and practical implementation. To clarify the relationship between the results of the analysis and the recommendations presented in Table 6, it is important to explicitly link the identified thematic clusters and analytical outputs to the derived implications. The recommendations are not introduced as independent propositions, but are directly grounded in the empirical evidence emerging from the clustering and NLP-based analysis. In particular, each recommendation stems from a combination of (i) thematic relevance identified through LDA topic modeling, (ii) term importance derived from TF-IDF analysis, (iii) semantic patterns observed through KWIC and collocation analysis, and (iv) structural relationships among topics highlighted by the knowledge graph. More specifically, the identification of gaps associated with system integration arises from the limited connectivity observed between clusters, indicating weak links among different supply chain tiers. Similarly, the recommendation related to digital technologies is derived from the presence of technology-oriented clusters (e.g., blockchain) that remain largely disconnected from operational and implementation-oriented themes, revealing a gap between conceptual innovation and practical application. The resilience-related recommendations are based on the dispersed occurrence of resilience-related terms across multiple clusters without forming a coherent thematic structure, suggesting the absence of a unified framework. In parallel, sustainability is identified as a marginal topic due to its low frequency and lack of a dedicated cluster, indicating an underdeveloped research area. Furthermore, the analysis highlights inconsistencies in the treatment of regulatory aspects and data interoperability, which appear in isolated contexts rather than as integrated themes, leading to recommendations aimed at improving alignment and standardization. Finally, the observed fragmentation of the literature, characterized by isolated thematic clusters and limited cross-disciplinary connections, directly informs the recommendation to promote more integrative and interdisciplinary research approaches. Therefore, Table 6 represents a synthesis of the main outcomes of the analysis, translating the identified structural patterns and thematic gaps into actionable research and managerial implications. Each recommendation is thus explicitly linked to evidence derived from the clustering results and the supporting analytical techniques.

These findings depict an overall picture of a research field that is highly compartmentalized. (i) Innovation-related themes are separated from implementation concerns; (ii) sustainability topics are fragmented and marginal; (iii) regulatory issues are handled selectively; (iv) the technology focus is uneven; (v) and operational challenges are studied in isolation. The HSC literature remains a body of work rich in individual insights but weak in integration. This structural disconnect highlights the need for more integrative, cross-cutting research approaches. The existing literature frequently considers technical solutions and policy or organizational strategies as distinct areas, hindering the development of scalable and cohesive solutions. Promoting greater interdisciplinary research and fostering practical collaboration could result in more efficient and equitable HSC management. The findings of this study provide several important implications for both theory and practice in the field of healthcare supply chain management. From a theoretical perspective, the identification of distinct thematic clusters highlights the fragmented nature of the literature, suggesting the need for more integrative frameworks that bridge currently disconnected research streams. In particular, the limited interaction between technological innovation, operational management, and policy-oriented studies indicates that future research should adopt a more systemic and interdisciplinary approach.

From a methodological perspective, the study demonstrates the potential of combining a systematic literature review with natural language processing techniques to analyze large and heterogeneous bodies of literature. This approach enables the identification of latent patterns that are difficult to capture through traditional qualitative reviews, contributing to a more objective and scalable analysis of complex research domains. From a managerial perspective, the results underline the importance of adopting an integrated view of healthcare supply chains, moving beyond isolated optimizations toward system-wide coordination. The identified gaps in areas such as sustainability, resilience, and the implementation of digital technologies suggest that decision-makers should prioritize strategies that enhance end-to-end visibility, data sharing, and cross-organizational collaboration. This is particularly relevant in contexts characterized by high uncertainty and critical resource availability, such as healthcare systems. The study highlights the gap between theoretical advancements and practical implementation, indicating the need for closer collaboration between academia and practitioners. Bridging this gap is essential to translate analytical models and technological innovations into effective solutions capable of improving the performance, resilience, and sustainability of healthcare supply chains.

4.3. Limitations and Future Research Directions

In terms of methodological limitations, this study relies on a single database, namely Web of Science (WoS), which, although ensuring high-quality and peer-reviewed sources, may not fully capture the breadth of the relevant literature across all disciplines. Other databases, such as Scopus or Google Scholar, may include additional contributions, particularly from emerging research areas or interdisciplinary fields, which could enrich the analysis. Furthermore, the search strategy, while carefully designed, is based on a specific set of keywords that may not fully represent the diversity of terminology used in healthcare supply chain research. As a result, some relevant studies may have been excluded, particularly those adopting alternative expressions or focusing on closely related domains. From a methodological perspective, the use of natural language processing techniques introduces additional limitations. Although these methods enable a scalable and systematic analysis of large corpora, their outcomes are inherently influenced by preprocessing choices, such as stopword removal, term filtering, and keyword selection. In addition, the rapidly evolving vocabulary of healthcare supply chains, as well as the presence of domain-specific terminology, may affect the accuracy and consistency of text-based analyses. The topic modeling approach adopted in this study, based on latent Dirichlet allocation (LDA), involves a degree of subjectivity in the interpretation of the resulting topics. While quantitative indicators and validation procedures were considered, the labeling and interpretation of clusters remain partly dependent on the researchers’ judgment. The study provides a cross-sectional analysis of the literature, focusing on the current state of research without explicitly considering its temporal evolution. Future research could address these limitations by incorporating multiple databases, refining keyword selection strategies, integrating expert validation, and adopting longitudinal approaches to better capture the dynamics of the field over time.

5. Conclusions

This study identifies six central thematic clusters in the research on HSCs by offering valuable insights into both current challenges and emerging trends. Through a robust application of NLP techniques, including topic modeling and TF-IDF analysis, this research clarifies key focus areas and highlights gaps that guarantee further investigation to explore how to manage. The analysis shows that while resilience, digitalization, and emergency logistics have become increasingly central, areas like sustainability and cross-sector integration remain underexplored. This imbalance highlights the need for research that bridges tactical operations and broader strategic coordination to address long-term challenges in HSCs. The research questions outlined in Section 1 are addressed through the combined application of a systematic literature review and NLP-based analysis. With respect to RQ1, the study identifies six main thematic clusters in the healthcare supply chain literature through LDA topic modeling, providing a structured representation of the domain. Regarding RQ2, the analysis links these clusters to key challenges in healthcare supply chains, such as the availability of pharmaceuticals, medical devices, and electromedical equipment, highlighting how different research streams address operational, technological, and organizational issues. Finally, RQ3 is addressed by interpreting the identified thematic clusters to uncover research trends, gaps, and opportunities, including the fragmentation of the literature, the limited integration across supply chain tiers, and the underrepresentation of sustainability and implementation-focused studies. Overall, the findings provide a comprehensive understanding of the current state of research and offer a basis for future developments in healthcare supply chain management. In particular, the findings underscore the critical need for resilient and flexible supply chains capable of responding to unforeseen disruptions, such as those triggered by pandemics or natural disasters. While logistics, traceability, and flexible production systems are recognized for their transformative potential, they remain relatively underexplored in terms of comprehensive research focus. The research also reveals that the existing literature tends to concentrate on isolated operational functions such as warehousing or distribution, often in response to demand or crises, while systemic design and coordination receive less attention. The key challenges in ensuring the availability of critical medical products, such as pharmaceuticals, medical devices, and electromedical equipment, have been prominently featured in the literature, with a significant focus placed on improving inventory management, distribution efficiency, and quality control. However, issues like long-term system integration and anticipatory design remain underrepresented, despite their critical role in building robust HSCs. The analysis also points out some opportunities for future research that could address these gaps, particularly in the areas of cross-sector integration and the development of sustainable, data-driven solutions. Moreover, longitudinal studies could offer deeper insights into the evolution of HSCs over time and their responses to systemic changes. The implications of this research are far-reaching, offering both academic and practical insights. By providing a structured overview of HSC research, the study suggests a clear pathway for future innovation efforts. The focus should be on developing adaptable and integrated infrastructures that can cope with both expected and unexpected challenges. As HSCs become more interconnected and complex, it is essential to prioritize strategies that enhance resilience, sustainability, and efficiency. This study lays the groundwork for a more comprehensive and adaptive approach to managing HSCs, ultimately contributing to more effective, patient-centered healthcare systems.