Understanding research priorities in home health care: a topic modeling analysis of the health services literature Authors: Mijin Bae, Minsung Kim, Aeri Kim, Hana Lee, Kyungmi Woo Published: 2026 ABSTRACT AbstractBackgroundHome health care (HHC) has become vital amid global aging, rising chronic disease burdens, and shifts toward community-based care. Yet, research priorities remain fragmented. This study aimed to map the evolution of HHC scholarship and identify emerging and underexplored themes.MethodsWe analyzed 39,548 PubMed-indexed HHC publications (1952-April 2025) using Latent Dirichlet Allocation topic modeling. A hybrid Konstanz Information Miner–Python workflow supported preprocessing, parameter optimization, and topic extraction. Model performance was assessed with perplexity, coherence, and clustering validation.ResultsTen coherent topics were identified, spanning clinical, technological, and population-focused domains. Healthcare resource utilization remained the most consistent and dominant theme. Rapidly expanding areas included telehealth adoption, technology-assisted caregiving, and end-of-life services. Temporal analysis revealed increasing attention to digitally integrated, person-centered models since the 2010s, accelerated by the COVID-19 pandemic. Pediatric care and high-risk population screening were notably underrepresented.ConclusionsThis study provides a macro-level mapping of seven decades of HHC research, highlighting evolving priorities and persistent blind spots. Topic modeling offers valuable insights for shaping a more inclusive, evidence-informed research agenda and guiding policy development in home-based care.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12913-026-14831-5.Keywords: Home health care, Topic modeling, Telemedicine, Palliative care, Caregiver support Introduction Home health care (HHC) has become a cornerstone of modern healthcare systems, particularly in the context of global population aging, the rising burden of chronic diseases, and the shift toward value-based, community-centered care. Global population is aging at an unprecedented rate. By 2030, one in six people will be aged 60 years or over, increasing from 1 billion in 2020 to 1.4 billion; this number is projected to double to 2.1 billion by 2050 [1]. This demographic shift is closely linked to a sharp increase in chronic illnesses. In 2019, noncommunicable diseases (e.g., cardiovascular disease and diabetes) caused 41 million deaths globally, approximately 10 million more than in 2000, and now account for roughly 75% of all deaths [2, 3]. As people live longer with multiple chronic conditions, the demand for care in home and community settings has grown. This reflects patient preferences for aging in place and health system efforts to reduce hospital utilization costs. For example, in the U.S., the population aged 65 + is expected to nearly double from 43 million in 2012 to 84 million by 2050 [4]. During the same period, the HHC service industry’s annual revenue grew by over 50% between 2013 and 2020, reaching approximately $103 billion [4]. Research on HHC has also expanded, covering diverse topics such as care coordination, digital health innovations, caregiver support, and end-of-life (EoL) care [5–7]. This growth has produced an increasingly diverse and multidisciplinary body of literature, reflecting the expanding scope of HHC. As the field continues to evolve across multiple clinical contexts and service models, synthesizing overarching trends and identifying long-term research trajectories has become more challenging. Existing bibliometric and scoping reviews on HHC have primarily adopted narrative methods or manually coded approaches. Although valuable, these studies are limited in their scalability and may have overlooked nuanced thematic shifts [8]. Moreover, previous reviews have tended to focus on narrow subdomains. For instance, many studies have focused on areas such as telehealth interventions or pediatric home care rather than on the field as a whole [9, 10]. As a result, significant gaps remain in our understanding of the broader trends and underexplored areas of HHC. It remains unclear how key themes such as the integration of behavioral health, support for family caregivers, and innovative service models have shifted over time, as well as which populations or care delivery models have been neglected. At this juncture, a comprehensive, data-driven synthesis of HHC literature is essential. Such an approach would help identify emerging directions and highlight neglected populations and service models that merit greater scholarly and policy attention. This study is particularly timely, as the last two decades have witnessed substantial changes in home-based care driven by technological innovation, policy reforms, and, more recently, the COVID-19 pandemic’s acceleration of virtual care. HHC is evolving from traditional nurse visits and homemaker services to more complex, technology-integrated models. Together with digital health tools such as remote patient monitoring and smart home devices, telehealth is now becoming an integral part of home care delivery. Particularly, the COVID-19 crisis has dramatically accelerated the adoption of telehealth and remote services in health care [11, 12]. For example, telemedicine use in the U.S. skyrocketed, with telehealth visits accounting for only 0.2% of medical claims in 2019 but rising to 15.4% in 2020 amid the pandemic [13]. In the HHC sector, the overall cumulative prevalence of telehealth among home health agencies rose from approximately 23% before 2020 to over 56% by 2021, marking a remarkable transformation in practice [12]. These developments underscore how HHC has expanded in scope and capability, blurring the lines of traditional facility-based care. Therefore, we must systematically examine how the research landscape has responded to these shifts and identify areas where scholarship may be lagging. To address the identified gaps, the present study employs Latent Dirichlet Allocation (LDA) topic modeling as an unsupervised machine learning technique to map the thematic evolution of HHC research. Rather than conducting a systematic or scoping review to synthesize findings from individual studies, this computational bibliometric approach enables the analysis of thematic patterns and temporal trends in HHC literature. By applying LDA topic modeling, the analysis captures not only surface-level thematic patterns but also latent thematic structures that reflect underlying relationships across studies. Furthermore, this method enables researchers to analyze thousands of abstracts at scale, quantify topic prevalence over time, and detect subtle shifts in scholarly focus, providing insights that are often inaccessible through manual review alone. Building on this approach, this study aims to (1) identify dominant research themes in the HHC literature, (2) explore how these themes have shifted over time, and (3) provide insights for researchers, clinicians, and policymakers seeking to align future care models with evolving evidence. This study contributes across educational, scholarly, and policy domains. Methodologically, it advances the use of LDA topic modeling in health services research by integrating term frequency-inverse document frequency (TF-IDF) refinement, dual-metric evaluation, and a Konstanz Information Miner (KNIME)-Python hybrid workflow, making the approach both rigorous and accessible. At the scholarly level, it highlights persistent gaps in HHC research, promoting a more inclusive agenda that reflects the diverse needs of HHC recipients. From a policy perspective, the findings resonate with evolving initiatives such as telehealth policy updates and home-based care models, underscoring the importance of patient-centered, technology-enabled, and equity-oriented approaches for future care delivery. Methods All preprocessing and topic modeling procedures were implemented using KNIME analytics platform 5.4.3, an open-source visual workflow-based tool widely used in data science and text mining (http://www.knime.com/). KNIME was chosen for its capacity to handle big data, ease of using unstructured data analysis, and preset workflows for machine learning algorithms, all of which fully met the requirements of this study [14]. KNIME enables modular and transparent construction of the workflow, allowing step-by-step manipulation of the corpus, including document filtering, tokenization, lemmatization, stop word removal, TF-IDF calculation, and LDA modeling. Additionally, Python code was executed using Jupyter Notebook version 7.2.2, within the Anaconda distribution, and integrated into the KNIME workflow to complement functionality not natively supported within the platform. Data collection We retrieved articles from PubMed, a leading biomedical literature database, related to HHC. PubMed was selected as the primary data source because it provides comprehensive coverage of biomedical and health services research and uses standardized indexing through Medical Subject Headings (MeSH), which facilitates consistent retrieval of conceptually related publications. The search strategy employed MeSH and entry terms in the query. Specifically, we used the MeSH terms “home care services” and “home health nursing,” as well as entry terms such as “home health care,” “home care,” “home-based care,” and “domiciliary care.” A language filter was applied to include only articles written in English, and only publications containing abstracts were considered. The search terms were developed by researchers and validated by a medical librarian. The literature search yielded a dataset that was refined by excluding entries without abstracts. To prevent duplication, records with identical titles (n = 1,009) and abstracts (n = 24) were excluded. After pre-processing, 39,548 unique articles published up to April 2025 were included in the final corpus. We did not restrict the publication year to ensure comprehensiveness. Text preprocessing The PubMed search was conducted on April 22, 2025 (the complete search query is provided in Supplementary Appendix 1). Records were retrieved directly from PubMed using the Document Grabber node in KNIME, which collects metadata provided by PubMed as document objects within the workflow. The retrieved documents were then processed using the Document Data Extractor node to extract the fields required for analysis, including article title, author information, abstract, document text, and publication year. The extracted data were exported as a CSV file and used for subsequent preprocessing and topic modeling. All publication types retrieved from PubMed were included in the corpus. As stated earlier, because this study aimed to map the thematic landscape of the HHC literature using topic modeling, the analysis included diverse forms of scholarly publications rather than restricting the dataset to specific study designs. Text preprocessing was then conducted to transform the collected abstracts into a structured format suitable for topic modeling. The preprocessing workflow sequentially performed several steps. It began by removing non-alphabetic elements such as language tags (e.g., “Spanish”), then converted all words to lowercase, followed by the removal of punctuation and numerical characters. Words with fewer than three characters were filtered out. Part-of-speech tagging was then performed to annotate each token with its grammatical category (e.g., noun, verb, adjective). From these tagged tokens, only nouns were extracted for further analysis. The text was tokenized using a bag-of-words model based on unigrams, in which each word was treated as an individual token. Common English stop words were removed using the built-in English stop word list provided in the KNIME platform, which relies on predefined language-specific stop word files maintained within the software package. Finally, lemmatization was applied to reduce words to their base or dictionary forms. This process helped unify word variants (e.g., “caring” and “cared” to “care”), thereby improving the consistency and semantic clarity of the text data used for topic modeling [15]. To identify informative and discriminative terms across documents, keywords were extracted using the TF-IDF method. TF-IDF is a widely used weighting scheme in information retrieval and text mining that reflects the importance of a word to a document in a collection [16]. It increases proportionally with the frequency of a term in a document but is offset by how frequently the term appears across all documents in the corpus. This allows for the down-weighting of common but less informative terms while emphasizing terms that are more specific to individual documents [17]. The analysis also included the calculation of TF-IDF values and subsequent filtering based on frequency thresholds to refine the vocabulary used in topic modeling. To remove rare terms, we applied a dynamic threshold based on document frequency; only those terms appearing in at least 0.05% of all documents were retained. After reviewing the distribution of the TF-IDF values across the corpus, the threshold was set at 0.089 based on an iterative expert review process. This selection ensured that the resulting vocabulary maintained high clinical relevance while effectively filtering out non-informative academic terms such as ‘home,’ ‘study,’ and ‘patient (Supplementary Appendix 2). TF-IDF scores were used only as a vocabulary filtering step prior to topic modeling. The initial vocabulary contained 23,319 unique terms and was reduced to 1,806 terms after TF-IDF filtering (approximately 7.7% of the original vocabulary), a level of reduction commonly applied in large-scale text mining studies to improve topic coherence and reduce noise [18, 19]. No documents were removed during this filtering process. To assess the robustness of the selected threshold, we examined topic solutions across nearby threshold values and observed broadly consistent topic structures. Topic modeling LDA topic modeling was conducted to uncover the latent thematic patterns in the corpus of HHC literature. LDA is a generative probabilistic model that assumes that each document is composed of multiple latent topics and that each topic is a distribution over words [20]. Thus, LDA allows themes to emerge inductively from the data, offering an organic, corpus-driven perspective on the structure of knowledge within a field [20, 21]. To enhance transparency in reporting topic modeling procedures, this study followed key recommendations from the LDA Preferred Reporting Checklist [22]. The completed checklist is provided in Supplementary Appendix 2. The two key hyperparameters of LDA are α (alpha), which controls the sparsity of the topic distribution per document, and β (beta), which controls the sparsity of word distributions per topic [23, 24]. In this study, we set both α and β to 0.01 to enforce a conservative prior, encouraging each document to focus on a small number of dominant topics and each topic to consist of a limited set of highly representative terms. This sparse configuration was intended to enhance the interpretability of topic structures by minimizing noise and topic overlap. LDA is particularly suitable for this study because it enables the unsupervised extraction of interpretable themes from large-scale textual data [5]. KNIME facilitated the modeling process through its intuitive interface and built-in LDA nodes, enabling efficient and reproducible implementation. Moreover, its probabilistic foundation allows the assignment of soft topic memberships to documents, facilitating the observation of mixed thematic patterns over time [25]. Topic modeling was implemented using the Topic Extractor (Parallel LDA) node in the KNIME Text Processing extension, which relies on the Machine Learning for Language Toolkit (MALLET) library and employs parallel Gibbs sampling with the SparseLDA algorithm for topic inference. The model was trained for 1,000 iterations, with a fixed random seed of 42 to ensure reproducibility of the results. Model convergence was determined by reaching the predefined maximum number of iterations specified in the algorithm. These parameter settings were selected to promote stable estimation of topic distributions across the corpus. The fitted LDA model produced a probability of distribution of topics for each document. For interpretability in reporting topic prevalence, each document was assigned to its dominant topic, defined as the topic with the highest probability (hard assignment). No additional probability threshold was applied, and each document was assigned to the topic with the maximum posterior probability. Examination of the maximum topic probability across documents indicated that dominant topics were generally well separated, supporting the use of dominant-topic assignment for summarizing topic prevalence. Identification of optimal K and its validation To advance the analysis, we needed to determine the optimal number of topics (K) for the model. We evaluated model performance using both perplexity and coherence scores. Perplexity, which reflects model generalizability, was calculated in two stages. First, a broad search (K = 2–80) was conducted to identify the stabilization zone. A finer-grained evaluation (K = 5–20) was performed to determine the optimal range [26, 27]. Coherence, which assesses the semantic consistency of topics, was computed within the same range (K = 5–20) using the Gensim library in Python [27]. To enable a direct comparison between these two metrics, coherence and perplexity values for each candidate topic number K were normalized using min–max scaling: Because lower perplexity indicates better model fit, perplexity was reverse-normalized so that higher values represented better performance. A composite score was then calculated as: These quantitative metrics were used to identify a reasonable range of candidate topic numbers. The final K was selected based on the researchers’ qualitative assessment of topic interpretability and semantic distinctiveness across candidate models. Following this combined evaluation, K = 10 was selected as the optimal solution. The number of selected topics was validated using the elbow method with K-means clustering. The underlying assumption was that if LDA reveals a meaningful topic structure, a similar number of clusters should emerge through unsupervised clustering [28]. To test this, we applied K-means to the same TF-IDF matrix used for LDA, varying K from 1 to 20 [29]. The inflection point, determined by examining the within-cluster sum of squared errors, aligned with the K selected through perplexity and coherence scores, thereby reinforcing the validity of the topic model structure. To enhance transparency and reproducibility, a detailed overview of the workflow, including preprocessing, topic number evaluation, and LDA modeling procedures, is provided in Supplementary Appendix 3. Results A total of 39,548 articles published between 1952 and April 2025 were included in the final analysis. The overall trend reveals a substantial increase over the past few decades, particularly since the 1980s. Starting in the 2010s, the number of publications rose sharply, with over 1,000 articles published annually and a consistent upward trajectory. The most pronounced peak was observed in 2024, when more than 1,800 articles were published. This represents the highest annual count in the dataset (Fig. 1). Temporal Distribution of Publications on Home Health Care (n = 39,548). Legend: The figure shows the yearly distribution of articles related to Home Health Care. Data for 2025 include publications up to April only Based on the evaluation of coherence, perplexity, and inertia (the elbow method), the optimal number of topics for the LDA model was determined to be 10 (K). The intertopic distance map, generated using t-distributed stochastic neighbor embedding (t-SNE), further supported the choice of K = 10 by visually illustrating well-separated and distinct topic clusters (Fig. 2). Furthermore, to validate the decisiveness of these assignments, we analyzed the distribution of maximum topic probabilities across documents (mean = 0.7589, median = 0.8659, interquartile range [IQR] = 0.6350–0.9275), which indicates that dominant topics were generally well-separated and that documents were assigned to their respective topics with high certainty. Intertopic distance map generated using t-Stochastic Neighbor Embedding (t-SNE) (K = 10). Legend: This technique reduces the high-dimensional topic distributions into two dimensions, preserving semantic relationships. Each circle represents a topic, and the spatial distance between circles reflects the dissimilarity in topic-word distributions. Clear separation supports the coherence of the 10-topic solution (Maaten & Hinton, 2008). Axes (Dimension 1 and Dimension 2) are derived from t-SNE projection and do not carry inherent interpretive meaning Ten topics were extracted from the HHC literature corpus using LDA topic modeling. Each topic was assigned a theme based on the top ten highly weighted terms. These topics reflect diverse areas, including clinical care, digital health, high-risk populations, and health service delivery in home-based settings. Table 1 summarizes each topic and the corresponding themes and keywords. Table 1Key topics extracted from home health care literature and Top 10 Keywords per Topic Key topics extracted from home health care literature and Top 10 Keywords per Topic Topic 1 highlighted the variations in healthcare utilization across populations and systems. Top terms, such as utilization, resource, hospitalization, and pattern, indicated a focus on comparative service research and resource allocation. Topic 2 examined strategies for implementing telehealth and the barriers hindering its adoption. The key terms included implementation, barrier, telehealth, veteran, and strategy, reflecting the need to overcome the challenges in remote care delivery. Topic 3 focused on assessing infectious and mental health risks in high-risk populations. Terms such as infection, depression, HIV, and instruments suggest the use of standardized tools in community-based health assessments. Topic 4 emphasized care transitions, clinical decision-making, and patient preferences at EoL. Terms included transition, preference, participation, and choice, indicating patient-centered decision making in long-term or palliative care. Topic 5 explored health perceptions, knowledge, and behaviors among children, adolescents, and their families. Terms such as parent, infant, childhood, adolescent, and attitude indicate research interest in parental influence, early health education, and developmental literacy. Topic 6 covered the use of technology to support caregivers and older adults. Key terms such as technology, caregiving, Alzheimer’s, and application reflect technology-enabled care environments and caregiver burden relief. Topic 7 centered on telemedicine and devices for managing symptoms in chronic or post-acute care. Terms included injury, ulcer, wound, asthma, and telemedicine. Topic 8 focused on team-based care and patient-clinician interactions. Terms such as worker, aide, satisfaction, and relationships indicate the attention paid to care dynamics and perceived care quality. Topic 9 examined the impact of HHC on surgical outcomes, recovery, and broader health endpoints. Keywords such as readmission, day, outpatient, and mortality indicated research focusing on reducing hospital readmissions, improving outpatient recovery pathways, and enhancing survival after procedures. Topic 10 addressed the provision of HHC for individuals undergoing dialysis, awaiting transplantation, or recovering from such procedures. Terms such as dialysis, survival, transplantation, and center indicate the research interest in patient outcomes, treatment settings, and differences in care delivery across facilities. To analyze topic distribution and temporal trends in the HHC literature, we used two visualizations: a bar chart of the overall percentage of documents per topic (Fig. 3) and a heatmap of annual topic proportions (Fig. 4). The analysis covered the years 1952–2025 for the bar chart and 1980–2024 for the heatmap; earlier years and 2025 were excluded from the heatmap because of insufficient publication volume. The bar chart shows that Topic 1 dominated the literature (15.5%), followed by Topics 2, 8, and 6, indicating their central roles in the research landscape. The heatmap reveals that Topic 1 maintained a consistently high share, while Topics 2 and 8 rose markedly in the 2010s, reflecting a growing interest in digital transformation and patient-centered care. In contrast, Topics 3, 5, and 7 remained relatively stable, while Topic 4 grew gradually from the mid-2000s. Topic 6 was more prominent in the late 1980–1990 s before leveling off. Percentage of documents assigned to each topic (1952–2025). Legend: Each bar represents the proportion of documents assigned to each topic among the analyzed corpus. Data for 2025 include publications up to April only. Topic percentages were calculated based on the dominant topic assigned to each document from the LDA model, rather than weighted topic probabilities. The percentage for each topic was computed as the number of documents assigned to that topic divided by the total number of documents included in the analysis Temporal Heatmap of Topic Proportions in Home Health Care Literature (1980–2024). Legend: The heatmap shows the annual proportion (%) of documents assigned to each topic. Each document was assigned to a single topic using dominant-topic (hard) assignment, defined as the topic with the highest posterior probability in the LDA model. Topic percentages were calculated as the proportion of documents assigned to each topic within each publication year. Topic numbers correspond to the thematic labels described in Table 1 Discussion This study applied LDA to 39,548 HHC publications (1952–2025) and identified ten thematic areas. Healthcare resource utilization was the most dominant topic, and themes such as telehealth implementation, technology-assisted caregiving, and person-centered care have grown in recent years. In contrast, areas including pediatric and high-risk screening groups remained underrepresented. These patterns provide a concise overview of the evolving research priorities in HHC. Healthcare resource utilization has remained consistently prominent in HHC research for decades, reflecting enduring research interests and policy relevance. This persistence is partly attributable to population aging and the rising burden of noncommunicable diseases, which have intensified the demand for home-based services [1–4]. To explain variations in HHC utilization, researchers often rely on Andersen’s behavioral model, which is a widely used framework that categorizes factors influencing service utilization into predisposing, enabling, and need [30]. In HHC, utilization is shaped by clinical needs and social determinants of health, including age, insurance status, living arrangements, and caregiver support [31, 32]. The prominence of this topic reflects not merely the frequency of utilization but also the substantial body of research aimed at explaining variations and inequities in service access. Such inquiries are closely aligned with management and policy priorities for equitable service access and efficient resource distribution [33]. Moreover, HHC has emerged as a strategic tool for utilization management, reducing rehospitalization and lowering overall costs, particularly for Medicare beneficiaries [34, 35]. Together, these findings explain why this topic remains highly prominent, serving as a measurable indicator shaped by multiple determinants and a central strategy for optimizing home-based care. Following healthcare resource utilization, two other prevalent themes in the HHC literature are telehealth adoption and technology-assisted caregiving. Research in these areas has substantially expanded, with growth further accelerating during the COVID-19 pandemic [11, 12]. These findings highlight telehealth’s role in chronic disease management, reducing missed appointments, and supporting medication adherence [13]. In the United States, the annual new adoption rate of telehealth among home health agencies more than doubled, increasing from 10.2% in 2019 to 26.0% in 2020 [12]. This surge directly explains the dramatic jump in the overall cumulative prevalence, while outcomes and patient satisfaction remained largely comparable to those of in-person care [36]. In Europe, approximately 78% of the WHO/Europe member states have incorporated telehealth into national health policies or strategies [37]. Between March 2020 and July 2022, over 118 million telehealth services were delivered to over 18 million patients by nearly 95,000 practitioners in Australia, demonstrating the rapid integration of digital health into routine care [38]. Parallel advances in technology-assisted caregiving such as remote patient monitoring, smart home health technologies, and mobile applications have improved safety, early detection of deterioration, and caregiver responsiveness [39]. While such tools promote independence and reduce fall risk, barriers such as usability and access inequities remain [40]. Informal caregivers are also increasingly using everyday technologies (e.g., smartphones, video conferencing) to coordinate care and maintain support networks [41]. Together, these trends indicate a shift toward technology-integrated, flexible, and accessible models of home-based care, with digital innovations playing an increasingly central role in supporting aging and reducing reliance on institutional care. Although technology-driven innovations enhance HHC, the human dimensions of care remain a core determinant in achieving high-quality outcomes. A growing body of research emphasizes that technology alone cannot guarantee positive outcomes unless it is embedded in strong interpersonal relationships [42, 43]. Patient-service provider interactions highlight the central role of trust, continuity, and team-based approaches. Trust between patients and service providers, particularly nurses and home care providers, is consistently associated with higher perceived care quality, satisfaction, and improved adherence to care plans [44]. Continuity in the care workforce, such as consistent visits from the same clinician, enhances reliability and emotional comfort for patients and families [45]. Further, team-based home care models contribute to improved outcomes by fostering coordinated communication, shared accountability, and stronger connections between patients and their care teams [46, 47]. These findings indicate that technology must be designed and implemented not to replace but to complement and reinforce the relational aspects of care that are central to patient well-being. Another human-centered theme in HHC research is patient-centered EoL care, which builds on these interpersonal aspects. Since the mid-2000s, patient-centered EoL care has gained momentum, driven by demographic aging, expanding palliative care policies, and greater recognition of patient preferences [1, 6, 7]. Research has highlighted the importance of integrating advance care planning (ACP) into home and community settings [48, 49]. By facilitating discussions about patients’ preferences, ACP improves care transitions, especially from skilled home health to hospice care, and supports the development of home-based palliative models that build trust, provide holistic care, support caregivers, and foster interprofessional collaboration [6, 7, 49]. These trends reflect a shift from institution-centered approaches to community-integrated, value-aligned models that facilitate smoother transitions, respect patient values, and enhance caregiver support. Although the overall proportion of publications on EoL care remains relatively low, the recent growth of this field underscores its increasing alignment with broader health system shifts toward patient-centered care. Nonetheless, further research is still needed to strengthen the evidence base and inform sustainable practice models in this domain. Moreover, research focusing on screening high-risk groups and pediatric care has remained scarce, with limited growth over time. This persistent gap reflects broader structural and policy dynamics. Key factors are workforce shortage and policy limitations, particularly in pediatric home care for children with medical complexities [8]. These challenges are compounded by reductions in specialized training programs that threaten the future supply of qualified clinicians [50]. Beyond workforce issues, ethical and methodological barriers further contribute to the persistent underrepresentation of pediatric populations in HHC research. At the clinical level, caring for technology-dependent children in home settings often raises complex ethical dilemmas for clinicians, as highlighted in classic discussions on pediatric home care ethics [51]. At the research level, involving children in clinical studies is subject to stringent requirements for consent, assent, and protective safeguards, which create additional hurdles for researchers [52]. Together, these systemic factors help explain why pediatric and high-risk populations continue to be underexplored. Given these constraints, future research should focus on community-based pediatric models, youth mental health support, and equity-driven interventions to ensure timely and appropriate home and community care [53]. This study has several limitations. First, although TF-IDF-based filtering improved thematic distinctiveness, it may have removed contextually important terms that appeared frequently across documents. Second, despite using perplexity and coherence scores to select the optimal number of topics (K), interpretive judgment was still required, introducing subjectivity into the model calibration. Third, the study relied solely on PubMed for literature collection; nonetheless, with 39,548 articles spanning 1952–2025, the dataset offers substantial breadth for identifying macro-level trends in HHC. Fourth, limiting the analysis to English-language abstracts may have excluded region-specific perspectives, and analyzing abstracts rather than full texts may have reduced contextual depth. Future studies should consider multi-database research, inclusion of full-text data, and hybrid analytic platforms to enhance robustness and inclusivity. Finally, the topic structures identified in this study may exhibit thematic overlap with adjacent home-based clinical domains, such as home dialysis. Because these clinical boundaries are not universally standardized and are subject to diverse cultural and policy-driven contexts, interpreting these topics from a global perspective requires caution. The definition and scope of ‘home health care’ may vary significantly depending on local healthcare structures. Building on these findings and acknowledging the study’s limitations, we propose three directions for advancing HHC research and practice. First, at the educational level, the expansion of technology-based interventions and patient-centered EoL care underscores the need to strengthen workforce competencies. Nursing and allied health curricula should integrate digital health literacy, interprofessional collaboration, and communication skills for advanced care planning, consistent with recommendations from the International Council of Nurses and WHO [54, 55]. Second, from a scholarly perspective, future topic modeling and bibliometric research should extend beyond biomedical databases to include sources such as CINAHL, Scopus, and ProQuest to capture the interdisciplinary nature of HHC. The combination of computational analyses with expert reviews or case studies can yield more contextually relevant insights [56]. Greater scholarly attention should also be directed toward underrepresented areas such as pediatric home care and adolescent mental health. Finally, for political implications, future research should focus on how HHC can better serve high-risk populations such as children, frail older adults, and immunocompromised individuals. Priority areas include evaluating equity-focused HHC models in underserved settings and assessing the cost-effectiveness, implementation barriers, and sustainability of telehealth and remote monitoring in real-world contexts [11, 41, 57]. Conclusions This study’s topic modeling analysis of seven decades of HHC literature revealed the evolution of research trends and identified significant gaps. HHC research has steadily expanded, focusing on healthcare utilization, technology-enabled care, and person-centered EoL care. Conversely, services for pediatric and adolescent populations, mental health, and medically or socially vulnerable groups have remained insufficiently researched and discussed. This study’s findings underscore the need for policymakers and practitioners to prioritize underserved areas, offering an evidence-based foundation rooted in the research shift toward digital and community-integrated models to design more equitable, effective, and patient-centered healthcare systems. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Acknowledgements This work was supported by the BK21 four project (Center for World-leading Human-care Nurse Leaders for the Future), Mijin Bae received a scholarship from the BK21 education program (Center for World-leading Human-care Nurse Leaders for the Future). Abbreviations Home Health Care End-of-Life Latent Dirichlet Allocation Term Frequency-Inverse Document Frequency Konstanz Information Miner Medical Subject Headings t-Stochastic Neighbor Embedding Advance Care Planning Author contributions MB: Conceptualization, study design, investigation, methodology, data curation, and writing—original draft, reviewing, and editing. MK: Investigation, methodology, data curation, and writing—original draft, reviewing and editing. AK: Methodology and writing—original draft, reviewing, and editing. HL: Methodology and writing—original draft, reviewing and editing. KW: Supervision, data curation, and writing—reviewing and editing. Funding This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No.RS-2024-00334672). Data availability The bibliographic data analyzed in this study was retrieved from the PubMed database, which is publicly available. The search strategy used to generate the dataset is described in the Methods section and Supplementary Materials, allowing the dataset to be reproduced. Declarations Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Data Availability Statement The bibliographic data analyzed in this study was retrieved from the PubMed database, which is publicly available. The search strategy used to generate the dataset is described in the Methods section and Supplementary Materials, allowing the dataset to be reproduced.