ABSTRACT
Clinical nutrition studies generate diverse research materials, including study protocols, case report forms, dietary records, laboratory measurements, body composition data, and final reports. However, these resources are often fragmented across storage systems and remain underutilized after study completion. Large language models (LLMs) offer opportunities for knowledge retrieval and documentation, but cloud-based services raise concerns regarding data privacy, institutional governance, and regulatory compliance. This Research Note proposes an institutional local LLM-based framework for secure knowledge retrieval, research documentation, and knowledge reuse in clinical nutrition research. The framework integrates multidisciplinary perspectives from clinical nutrition, human intervention trials, and artificial intelligence. Its architecture comprises five interconnected layers: a clinical nutrition data layer encompassing research documents and structured datasets; a document processing layer for text extraction, metadata generation, and standardization; a knowledge repository layer combining structured and vector databases for metadata management and semantic retrieval; a local LLM layer integrating retrieval-augmented generation with a separate numerical analysis pipeline; and a researcher interface supporting natural language queries and evidence-linked documentation. The proposed framework provides a conceptual foundation for secure, reusable institutional research knowledge management.
-
Keywords: Nutritional sciences; Artificial intelligence; Large language models; Information storage and retrieval
INTRODUCTION
Background
Clinical nutrition trials generate diverse structured and unstructured data throughout the research lifecycle, including study protocols, informed consent forms, case report forms, dietary assessment records, laboratory measurements, body composition data, statistical analysis outputs, final study reports, and scientific publications [
1]. Collectively, these documents constitute a valuable repository of institutional knowledge accumulated through substantial investments of time, funding, and human resources [
2]. Despite their significance, research documents are frequently maintained as separate files across multiple storage locations, including personal computers, institutional servers, laboratory information systems, and cloud-based repositories. Consequently, information generated during previous studies may become difficult to locate, retrieve, and incorporate into subsequent research activities. Such fragmentation can result in duplicated efforts in literature reviews, protocol development, outcome selection, and research report preparation [
3]. Moreover, valuable institutional knowledge may be lost when researchers leave their positions or research teams undergo organizational changes.
These challenges are particularly relevant to clinical nutrition research, in which substantial methodological and outcome-related information is repeatedly generated across studies. Previous protocols, research reports, and scientific publications may provide important methodological guidance for subsequent investigations. Nevertheless, the systematic retrieval and effective utilization of these resources remain inadequate in many research settings.
Current challenges
Recent advances in artificial intelligence (AI), particularly large language models (LLMs), have demonstrated considerable potential for information retrieval, document summarization, question answering, information extraction, and support for scientific writing [
4]. These developments have created new opportunities to manage and use large collections of research documents. Nevertheless, the direct implementation of publicly available, cloud-based AI systems in clinical research settings remains challenging [
5].
Clinical nutrition research involves the integration of dietary intake assessments, nutrient composition data, intervention-specific protocols, and longitudinal clinical outcomes. These data are often characterized by heterogeneous measurement units, dietary assessment methodologies, and nutritional reference standards. Unlike general-purpose document retrieval systems, the proposed framework is designed to connect nutrition-specific data elements with corresponding study protocols and outcome measures, thereby enabling context-aware retrieval of information across previous nutrition intervention studies.
Furthermore, processing clinical research documents through external AI services raises concerns regarding data privacy, confidentiality, intellectual property protection, and regulatory compliance [
6,
7]. Existing AI systems are also predominantly designed for general-purpose information retrieval and may not adequately accommodate the specialized workflows and documentation requirements of clinical nutrition research. Previous studies have demonstrated the application of retrieval-augmented generation (RAG) to clinical information extraction and evidence-based question answering [
8]. Other investigations have explored privacy-preserving deployment strategies for healthcare-related LLM applications, including those relevant to nutrition [
9]. However, the systematic organization, secure retrieval, and reuse of institution-specific clinical nutrition trial documents across research planning and documentation workflows remain insufficiently addressed.
In contrast to conventional digital transformation initiatives, which primarily emphasize the digitization of research workflows, and data interoperability and multimodal integration approaches, which focus on data exchange and harmonization, the proposed framework aims to transform fragmented institutional documents into reusable research knowledge through locally deployed LLMs, semantic retrieval, and evidence-grounded generation. Its principal contribution is the establishment of a link between accumulated clinical nutrition research knowledge and subsequent research planning and documentation activities. Accordingly, there is a growing need for secure AI-assisted systems that facilitate knowledge retrieval and research documentation while ensuring institutional control over research data.
To address this need, this Research Note proposes a locally deployed LLM-based framework for secure knowledge retrieval and research documentation in clinical nutrition trials. The framework is intended to provide a conceptual foundation for developing AI-assisted research infrastructure in clinical nutrition and related health science disciplines.
PROPOSED SYSTEM ARCHITECTURE
The proposed framework is designed to facilitate secure knowledge retrieval, document management, and research documentation throughout the clinical nutrition research lifecycle. Rather than replacing researchers’ decision-making processes, the system is intended to support the efficient retrieval, organization, and reuse of institutional research knowledge accumulated from previous clinical nutrition studies. The framework consists of five interconnected layers, with their respective development stages, computational procedures, and anticipated research applications summarized in
Table 1.
The clinical nutrition data layer serves as the primary information source for the proposed framework. It encompasses both structured datasets (e.g., CSV, XLSX, and JSON files) and unstructured documents (e.g., PDF, DOCX, and HWP files), including study protocols, case report forms, dietary records, laboratory results, and final research reports. Key data elements include study identifiers, intervention characteristics, participant demographics, energy and nutrient intakes, biochemical markers, body composition indices, and clinical outcomes.
These resources constitute valuable research assets containing methodological information, intervention characteristics, outcome measures, and accumulated institutional knowledge. The framework is designed to accommodate these heterogeneous data sources within a unified environment while preserving their original context and traceability. Importantly, it distinguishes unstructured research documents from structured participant-level datasets, enabling each data type to be processed through separate document retrieval and numerical analysis pathways.
Document processing layer
The primary objective of this layer is to convert diverse research documents into standardized, machine-readable formats suitable for subsequent retrieval and knowledge management. To facilitate efficient information retrieval and utilization, uploaded documents undergo a series of preprocessing procedures. Initially, document contents are converted into machine-readable text through text extraction or optical character recognition, as appropriate. The extracted information is then organized through document classification and metadata assignment.
Relevant metadata may include study identifiers, intervention types, participant characteristics, outcome measures, study duration, publication status, and document categories. Where applicable, personally identifiable participant information is removed or replaced through appropriate de-identification procedures to strengthen privacy protection and support compliance with applicable regulatory requirements.
Knowledge repository layer
Following preprocessing, research documents are stored in a knowledge repository comprising structured and vector databases. Structured databases retain document metadata and associated attributes, whereas vector databases facilitate semantic retrieval based on contextual similarity rather than relying solely on keyword matching [
10]. Narrative research documents are divided into section-level text segments (chunks), which are subsequently converted into numerical representations (embeddings) using an embedding model and indexed in the vector database to enable semantic retrieval.
In contrast, participant-level numerical datasets are processed through a separate structured data pipeline that incorporates structured query language-based retrieval and Python-based statistical analysis rather than text-based embedding methods. The resulting numerical outputs can subsequently be integrated with retrieved documentary evidence by the local LLM to generate summaries linked to their respective sources. This dual-repository architecture enables researchers to retrieve relevant documentary and quantitative information while maintaining a clear distinction between semantic document retrieval and numerical data analysis.
Local LLM layer
At the core of the proposed framework is a locally deployed LLM integrated with RAG, an approach introduced by Lewis et al. [
9] that combines information retrieval with language generation. Rather than relying exclusively on information encoded in the model, the system first retrieves relevant evidence from the knowledge repository and then generates responses grounded in the retrieved documents [
11]. For numerical queries, the local LLM is designed to incorporate outputs from the structured data analysis pipeline alongside relevant documentary evidence, rather than independently generating numerical estimates.
Potential functions of this layer include semantic document retrieval, evidence summarization, comparison of previous studies, assistance with protocol drafting and report preparation, and identification of studies involving similar interventions. Importantly, all processing is performed within a local computing environment, without transmitting research documents to external cloud-based AI services. This architecture is intended to strengthen data security and institutional control while retaining the practical advantages of contemporary AI technologies.
Researcher interface layer
The final layer provides a user-facing interface through which researchers can interact with the system using natural language queries. Researchers can request information about previous intervention studies, outcome measures, biomarker assessments, dietary assessment methodologies, and research report preparation.
For example, a researcher may ask, “Which previous probiotic intervention studies evaluated inflammatory biomarkers?” In response, the RAG system retrieves relevant study protocols and research reports containing information on inflammatory outcomes. The local LLM then generates a response summarizing the identified biomarkers (e.g., C-reactive protein and interleukin-6), together with references to the corresponding source documents. The generated outputs can subsequently be used to support protocol development, study planning, and research report preparation.
The interface is designed to deliver evidence-linked responses that facilitate research planning and documentation while maintaining transparency through traceability to the original source documents. This feature enables researchers to verify retrieved information against the underlying research records before incorporating it into subsequent research activities.
SECURITY AND GOVERNANCE
The integration of locally deployed LLMs into clinical nutrition research requires robust governance mechanisms to safeguard sensitive participant information, unpublished research findings, and institution-specific research assets. Accordingly, the proposed framework emphasizes institutional data ownership, secure local processing, controlled access, auditability, and researcher oversight. These five overarching governance principles are further operationalized through ten governance domains, encompassing privacy protection, evidence provenance, regulatory compliance, knowledge stewardship, and model governance, as summarized in
Table 2.
Unlike publicly accessible, cloud-based AI services, the proposed framework is designed to perform all document storage, retrieval, and inference within an institutional computing environment. From a security standpoint, it adopts an on-premises architecture based on a no-cloud-transmission principle. Research documents are stored and processed locally, thereby minimizing the risks associated with external data transfers. Furthermore, user authentication and role-based access control mechanisms can be implemented to restrict access according to project responsibilities and institutional policies. Access logs and audit trails can also be maintained to ensure transparency and accountability in document retrieval and AI-assisted interactions.
The framework additionally incorporates privacy-preserving research practices. Personal identifiers may be removed or replaced before document indexing to minimize the risk of inadvertent disclosure of participant information. However, the secondary use of data from previously conducted studies should be permitted only when consistent with the scope of the original informed consent or otherwise authorized under applicable regulations. Additional institutional review board approval or an appropriate waiver should be obtained whenever required. Importantly, de-identification alone does not constitute authorization for secondary data use. For industry-sponsored studies, the reuse of sponsor-owned or contractually restricted data also requires verification of data ownership, applicable confidentiality agreements, and explicit authorization from the relevant sponsor or data rights holder.
Furthermore, all AI-generated outputs are subject to researcher verification and approval through a human-in-the-loop workflow. Thus, the proposed framework is intended to facilitate scientific decision-making rather than replace professional judgment. For AI-assisted manuscript preparation, the framework should comply with the recommendations of the International Committee of Medical Journal Editors and applicable journal policies. AI-generated content must be critically reviewed and verified by human authors, who retain full responsibility for the accuracy, integrity, and originality of the submitted work. The use of AI must also be appropriately disclosed in accordance with relevant publication requirements. In addition, the reuse of previously published or institutionally generated research materials requires careful scrutiny to avoid inappropriate text recycling, self-plagiarism, and duplicate publication. Accordingly, source attribution, appropriate citation, and originality checks should be incorporated into the researcher review process.
Collectively, these governance mechanisms establish a foundation for the responsible deployment of locally hosted LLMs in clinical nutrition research settings while safeguarding sensitive research data, maintaining institutional control over research assets, and supporting regulatory compliance.
POTENTIAL APPLICATIONS
The proposed framework has the potential to support multiple stages of the clinical nutrition research lifecycle. By integrating secure knowledge retrieval, evidence synthesis, and AI-assisted documentation within an institutionally controlled computing environment, the framework may improve research efficiency, facilitate knowledge reuse, and strengthen institutional research continuity. Several representative application scenarios are described below.
Protocol development
Developing clinical nutrition research protocols typically requires an extensive review of previous intervention studies, outcome measures, participant eligibility criteria, and assessment schedules. However, retrieving relevant institutional documents and unpublished research materials can be time-consuming, particularly when research records are dispersed across multiple storage systems. The proposed framework may facilitate protocol development by enabling researchers to retrieve and compare previous intervention studies maintained in the institutional knowledge repository. Specifically, researchers may identify commonly used inclusion and exclusion criteria, intervention durations, dietary assessment methods, biomarker panels, and statistical approaches. By providing efficient access to accumulated institutional knowledge, the framework may streamline protocol preparation and minimize duplicated efforts during research planning.
Clinical trial documentation
Clinical nutrition studies generate substantial documentation throughout their implementation and reporting phases. Such documentation includes study protocols, monitoring records, interim analyses, final research reports, and scientific manuscripts. Preparing these materials often involves repetitive retrieval and synthesis of information from multiple sources. The proposed framework may facilitate clinical trial documentation by retrieving relevant study information and generating evidence-linked draft content for research reports and manuscripts. Potential applications include drafting methodological descriptions, study summaries, intervention descriptions, and selected report sections based on previously completed studies. Importantly, all AI-generated outputs remain subject to researcher review and approval through the human-in-the-loop workflow to ensure scientific accuracy and consistency with the original research records.
Nutritional intervention planning
Selecting appropriate intervention strategies and outcome measures is a critical component of clinical nutrition research. Researchers frequently need to identify previous studies investigating comparable nutrients, dietary patterns, functional foods, or biomarkers. Through semantic retrieval and evidence synthesis, the proposed framework may help researchers identify commonly used intervention protocols, outcome measures, and assessment schedules. For example, researchers investigating probiotic supplementation may retrieve previous studies evaluating inflammatory biomarkers, gastrointestinal outcomes, or changes in body composition. Access to this accumulated evidence may facilitate informed decision-making during intervention planning while preserving researcher autonomy and independent scientific judgment.
Institutional knowledge preservation
Research institutions accumulate substantial scientific knowledge over time through completed studies, technical reports, unpublished analyses, and internal research documentation. However, these resources are frequently fragmented and may become difficult to retrieve when researchers leave their positions or research projects are completed. The proposed framework may serve as an institutional knowledge preservation system by consolidating heterogeneous research documents into a searchable repository. This approach may minimize institutional knowledge loss, strengthen continuity across research projects, and facilitate knowledge transfer among researchers and research teams. Furthermore, the long-term preservation and systematic utilization of institutional research assets may improve the overall efficiency and sustainability of clinical nutrition research programs.
DISCUSSION
This Research Note presents a conceptual framework based on LLMs for secure knowledge retrieval and research documentation in clinical nutrition studies. Unlike general-purpose, cloud-based AI systems, the proposed framework is specifically designed to facilitate the management, retrieval, and utilization of institution-specific research documents while maintaining data security and governance. Although the framework remains conceptual and has not yet undergone formal implementation or validation, it highlights important opportunities and challenges associated with integrating AI into clinical nutrition research settings.
A major strength of the proposed framework is its emphasis on secure, institutionally controlled AI infrastructure. Clinical nutrition studies frequently involve sensitive participant information, proprietary intervention protocols, and unpublished research findings. By adopting a local deployment architecture and a no-cloud-transmission principle, the framework aims to mitigate risks associated with data privacy, intellectual property protection, and regulatory compliance. This approach may offer a practical alternative for research institutions seeking to leverage AI technologies while retaining control over their research assets.
Unlike previously reported RAG applications in healthcare, which have primarily focused on clinical information extraction and question answering, the proposed framework emphasizes institutional knowledge preservation and research documentation in clinical nutrition trials. Its contribution is therefore primarily application-oriented rather than algorithmic, and its practical effectiveness remains to be established through future implementation and validation. An important implication of this approach is its potential to extend existing healthcare RAG applications beyond clinical information extraction and question answering toward institution-specific research knowledge management. By integrating document-based evidence retrieval with a separate structured data analysis pipeline, the framework may facilitate both the reuse of methodological knowledge and quantitative research documentation in clinical nutrition trials.
Furthermore, the proposed framework may help reduce the documentation burden associated with clinical nutrition research. Through RAG-based information retrieval and evidence-linked document synthesis, researchers may be able to prepare research documents more efficiently while maintaining transparency and verifiability through traceability to the original source documents. This capability may be particularly valuable in research environments where information must be repeatedly retrieved and synthesized across multiple studies and documentation processes.
Despite these potential advantages, several limitations warrant consideration. First, the framework has not undergone formal performance evaluation. Consequently, its retrieval accuracy, response quality, and overall operational effectiveness remain undetermined. Second, no usability assessments have been conducted among researchers, clinicians, or clinical trial coordinators. Thus, the framework’s acceptability, usability, and practical utility in clinical nutrition research settings have yet to be established. Third, the present work describes a conceptual architecture rather than a fully implemented system. Accordingly, the proposed architecture may require further refinement to accommodate differences in institutional infrastructure, regulatory requirements, and research workflows.
Future studies should evaluate the framework using queries specific to clinical nutrition research, including those involving dietary assessment methodologies, intervention dosages, nutritional biomarkers, and clinical outcome measures. Its performance could be evaluated using retrieval precision, source attribution accuracy, and expert-rated response relevance. In addition, usability and practical applicability should be assessed by clinical dietitians and nutrition researchers. Multicenter implementation studies may further establish the framework’s generalizability across diverse research environments and institutional settings. Through iterative implementation, systematic evaluation, and refinement, the proposed framework may provide a foundation for the responsible integration of locally deployed LLMs into clinical nutrition research and related health science disciplines.
NOTES
-
Author Contributions
Conceptualization: JMC. Methodology: JMC, SL. Supervision: JMC. Investigation: SYL, JMC. Visualization: SYL. Writing - original draft: JMC. Writing - review & editing: all authors. All authors read and approved the final manuscript.
-
Conflict of Interest
None.
-
Funding
This research was supported by the ANCHOR program (Global University Program) through the Gyeongbuk ANCHOR CENTER, funded by the Ministry of Education (MOE) and the Gyeongsangbuk-do Provincial Government, Republic of Korea, in 2026 (2026-ANCHOR-15-110).
-
Data availability
Data of this research are available from the corresponding author upon reasonable request.
Table 1.The proposed development stages of a local LLM-based framework for clinical nutrition trials
Table 1.
|
Development stage |
Scientific objective |
Data elements |
Computational procedures |
Research utility |
|
1. Clinical nutrition data layer |
Establish a comprehensive institutional knowledge repository from completed and ongoing clinical nutrition studies |
Study protocols, informed consent forms, CRFs, dietary assessment records, nutrient intake databases, laboratory measurements, body composition data, final reports, and publications |
Secure document acquisition, repository construction, document inventory management, and data governance implementation |
Preservation of institutional research assets and reduction of knowledge fragmentation |
|
Separation of structured numerical datasets and unstructured research documents |
|
2. Document processing layer |
Transform heterogeneous research documents into a structured and interoperable knowledge environment |
Study-level metadata, intervention characteristics, participant information, outcome measures, biomarkers, study duration, and publication status |
Metadata extraction, ontology-based categorization, document classification, terminology harmonization, and schema mapping |
Improved document interoperability and retrieval precision |
|
3. Knowledge repository layer |
Enable semantic representation and contextual retrieval of research documents |
Standardized text documents and metadata annotations |
Section-level chunking, embedding generation, semantic indexing, vector database construction, and source-document linkage |
Creation of a searchable semantic knowledge repository for evidence retrieval |
|
Document embedding and semantic indexing; separate structured query language/Python-based processing of structured numerical data |
|
4. Local LLM layer |
Facilitate context-aware knowledge synthesis while preserving data security |
Researcher queries, retrieved document chunks, metadata context |
Retrieval-augmented generation, semantic similarity search, context augmentation, response generation, and source attribution |
Evidence-grounded responses with traceable document provenance |
|
Integration of retrieved documentary evidence and externally computed numerical results |
|
5. Researcher interface layer |
Support research planning, execution, interpretation, and scientific communication |
Retrieved evidence, AI-generated summaries, source-linked references |
Protocol drafting, literature synthesis, outcome measure recommendation, report generation, manuscript drafting assistance, and human review |
Enhanced research efficiency, reproducibility, and institutional knowledge utilization |
Table 2.Governance framework for the proposed local LLM-based clinical nutrition research system
Table 2.
|
Governance domain |
Governance objective |
Implementation strategy |
Expected outcome |
|
Institutional data ownership |
Maintain institutional control over research assets |
All research documents remain within institution-managed infrastructure and repositories |
Preservation of ownership and prevention of unauthorized external use |
|
Local processing and data sovereignty |
Prevent unintended disclosure of research data |
On-premise deployment and local inference without external cloud transmission |
Enhanced data sovereignty and reduced cybersecurity risk |
|
Privacy protection |
Protect participant confidentiality |
De-identification, pseudonymization, and restricted access to identifiable information |
Reduced risk of participant re-identification |
|
Access control |
Limit document access to authorized personnel |
User authentication, role-based permissions, and project-specific authorization policies |
Controlled access to sensitive research information |
|
Auditability and traceability |
Ensure accountability of AI-assisted activities |
Access logging, query tracking, document retrieval history, and audit trails |
Transparent monitoring of document utilization and system interactions |
|
Evidence provenance |
Maintain transparency of AI-generated outputs |
Source-document linking and citation of retrieved evidence used in response generation |
Improved explainability and reproducibility |
|
Human oversight |
Prevent autonomous scientific decision-making |
Human-in-the-loop review and researcher approval of all generated outputs |
Preservation of scientific judgment and accountability |
|
Regulatory compliance |
Align AI use with research governance requirements |
Compatibility with IRB procedures, institutional policies, and research ethics guidelines |
Responsible integration of AI into clinical nutrition research |
|
Knowledge stewardship |
Preserving institutional research knowledge over time |
Structured metadata management, semantic indexing, and long-term repository maintenance |
Sustainable institutional knowledge management |
|
Model governance |
Ensure responsible operation of local LLM systems |
Periodic system review, document updating, response validation, and governance monitoring |
Long-term reliability and trustworthiness of AI-assisted workflows |
REFERENCES
- 1. Weaver CM, Miller JW. Challenges in conducting clinical nutrition research. Nutr Rev 2017;75:491-9.
- 2. Johnson SB. Clinical research informatics: supporting the research study lifecycle. Yearb Med Inform 2017;26:193-200.
- 3. Sedlakova J, Daniore P, Horn Wintsch A, et al. Challenges and best practices for digital unstructured data enrichment in health research: a systematic narrative review. PLOS Digit Health 2023;2:e0000347.
- 4. Ke W, Zheng Y, Li Y, et al. Large language models in document intelligence: a comprehensive survey, recent advances, challenges, and future trends. ACM Trans Inf Syst 2026;44:1-64.
- 5. Singh K. Artificial intelligence & cloud in healthcare: analyzing challenges and solutions within regulatory boundaries. Int J Comput Sci Eng 2023;10:1-9.
- 6. Adenubi AO, Samuel N, Karimu AY. AI-driven synergistic model for enhancing intellectual property, cybersecurity, and privacy protection in academic research. Kasu J Comput Sci 2024;1:451-64.
- 7. Adeniran AA, Onebunne AP. Leveraging artificial intelligence for intellectual property compliance and global regulatory adherence. Int J Eng Technol Res Manag 2024;8:173.
- 8. Zhang H, Vithanage D, Deng C, Yu P. Optimising clinical information extraction: a comparative study of retrieval-augmented generation techniques in clinical notes. J Biomed Inform 2026;180:105053.
- 9. Lewis P, Perez E, Piktus A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. Adv Neural Inf Process Syst 2020;33:9459-74.
- 10. Taipalus T. Vector database management systems: fundamental concepts, use-cases, and current challenges. Cogn Syst Res 2024;85:101216.
- 11. Abo El-Enen M, Saad S, Nazmy T. A survey on retrieval-augmentation generation (RAG) models for healthcare applications. Neural Comput Appl 2025;37:28191-267.
