While the use of GenAI in healthcare offers substantial promise and is expected to become an integral part of regular clinical practice, its widespread adoption is limited by several critical challenges (Sai et al., 2024). These include fragmented and often inaccessible data silos, significant variability in the quality of real-world data, and complex ethical and legal considerations. Efforts to address these barriers are underway, including the implementation of federated learning approaches, the establishment of robust data governance frameworks that incorporate standards such as HL7 FHIR and FAIR principles, and the development of evolving regulatory measures—such as the European Union AI Act and the General Data Protection Regulation. Collectively, these initiatives aim to foster a landscape of responsible and ethical innovation in healthcare AI (Mons et al., 2017; Busch et al., 2024; Woisetschläger et al., 2024; Ibrahim et al., 2025; Liu et al., 2025). For readers seeking concrete real-world clinical practice use cases of GenAI in healthcare information systems, see Sai et al. (2024) and Reddy (2024).

Real-world data as recorded in EHRs consists primarily of free-text narratives that often contain valuable clinical insights that are not captured by structured data formats (Negro-Calduch et al., 2021; Seinen et al., 2024). However, the synthesis of meaningful health-related text has received little attention in GenAI literature (Murtaza et al., 2023; Ibrahim et al., 2025; Loni et al., 2025; Rujas et al., 2025). This represents a missed opportunity, not only in terms of data synthesis methodology, but also because free-text narratives often contain hidden and nuanced clinical information that is essential for meaningful patient understanding, accelerated early diagnosis, and improved clinical decision making (Negro-Calduch et al., 2021; Seinen et al., 2024; Schut et al., 2025). Moreover, while methods for synthesizing structured EHR data are well established, the generation of synthetic free text from real-world EHRs remains a comparatively underdeveloped area (Murtaza et al., 2023; Ibrahim et al., 2025; Rujas et al., 2025).

High quality synthetic health data generation (SHDG) that emulates real-world clinical documents is emerging as a vital strategy to enable safe, scalable, and privacy-preserving GenAI implementation in health and care settings (Smolyak et al., 2024; Loni et al., 2025). To be precise, SHDG is a privacy-enhancing technology that entails the generation of synthetic data based on real-world datasets. These synthetic datasets are designed to retain the essential statistical patterns and relationships of the original data, without containing any directly identifiable information. The goal is to enable analyses on synthetic data that produce results closely mirroring those obtained from the real data (Murtaza et al., 2023; Drechsler and Haensch, 2024; Ibrahim et al., 2025; Rujas et al., 2025). Ideally fully interoperable and machine-readable, synthetic datasets provide a foundation for the development, testing, and validation of innovative applications ranging from personalized healthcare models to solutions that alleviate administrative workload (Ibrahim et al., 2025; Rujas et al., 2025).

The growing promise and pitfalls of SHDG is best understood in the context of major breakthroughs in natural language processing (NLP) and GenAI, spanning early statistical approaches to today’s transformer-based language models. Synthetic data generation gained momentum in the early 1990s exemplified by Rubin’s (Rubin, 1993) multiple imputation framework and Little’s efforts on statistical disclosure through data masking (Little, 1993). Since 2010 the adoption of machine learning and deep learning has expanded the applications of synthetic data, especially in healthcare (Eigenschink et al., 2023; Murtaza et al., 2023; Drechsler and Haensch, 2024; Goyal and Mahmoud, 2024; Pezoulas et al., 2024). Deep learning models can automatically extract and represent intricate patterns from large datasets, thereby facilitating more accurate and efficient processing of information. Early approaches used models called convolutional and recurrent neural networks, which allowed machines to automatically process unstructured datasets such as images and speech without the need for feature analysis. However, deep learning model development is labor-intensive and complex (LeCun et al., 2015; Goodfellow et al., 2018).

The use of Generative Adversarial Networks (GANs) has led to significant advances in data synthesis and can be classified as both deep learning and GenAI approaches. GANs rely on adversarial training, in which two deep neural networks—the generator and the discriminator—compete with one another until the discriminator can no longer distinguish between real and synthetic data, reaching equilibrium (Goodfellow et al., 2018; Baowaly et al., 2019; Ibrahim et al., 2025). It is the most widely adopted AI technology in the fields of health and healthcare for generating synthetic data. GAN models were originally developed to produce realistic pictures—such as images from magnetic resonance imaging (MRI) or dermatoscopy—as well as to create continuous data. Continuous data includes numerical values that can take on any value within a range, like blood pressure, glucose levels, or patient age. In addition, specialized GAN models have also been used to generate time-series data, such as signals from electrocardiograms (ECG) and electroencephalograms (EEG) (Ibrahim et al., 2025).

The advent of transformer-based architectures revolutionized NLP by significantly enhancing language understanding (Vaswani et al., 2017). Groundbreaking pre-trained language models such as “bidirectional encoder representations from transformer” (BERT) laid the groundwork for this transition by introducing mechanisms for capturing nuanced contextual relationships within text. Building upon this foundation, domain-specific adaptations like BioBERT and Clinical-BERT, trained on biomedical literature and clinical notes, respectively, substantially increased the fidelity with which deep learning could represent the intricate semantics of medical language through vectorization and embeddings of text into numerical expressions (Alsentzer et al., 2019; Lee et al., 2020; Bommasani et al., 2021; Ibrahim et al., 2025). Capturing clinically meaningful concepts through tokenization, vectorization, and text embedding, however, remains a formidable challenge. For example, clinical notes contain valuable contextual information but are characterized by a variety of nomenclatures, abbreviations, misspellings, and synonyms both within and across healthcare disciplines (Cannon and Lucci, 2010; Doan et al., 2014; Meystre et al., 2017; Negro-Calduch et al., 2021; Seinen et al., 2024; Schut et al., 2025).

The evolution of transformer-based architectures culminated in the development of Large Language Models (LLMs) that are not only capable of language understanding—like BERT—but are also capable of language generation—like OpenAI’s GPT3—, called foundational models (Bommasani et al., 2021). LLMs further amplified generative and NLP capabilities—made possible by scaling up model architectures, leveraging enhanced computational power, and training on extremely large datasets. The emergence of novel functionality is associated with the so-called “scaling effect”—a phenomenon that was initially unforeseen and, to this day, remains inadequately understood (Kaplan et al., 2020; Coveney and Succi, 2025). Recent comprehensive surveys underscore the complexity and ongoing debate surrounding the generality of scaling laws, highlighting both their impressive predictive successes and significant limitations—especially when LLMs are applied in situations quite different from the functionality they were originally trained for, or when they face new types of data they have not seen before (Li et al., 2023; Sengupta et al., 2025). One of the most striking emergent properties of LLMs is their ability to respond to prompting—where users provide specific instructions or examples in natural language to guide the model’s output. This phenomenon allows for highly flexible and tailored use of LLMs without the need for additional fine-tuning, thereby broadening their practical utility in clinical document generation and comprehension. For example, prompting allows for zero-shot and few-shot learning, wherein LLMs can generate appropriate responses to tasks or questions with either no examples (zero-shot) or just a handful of provided examples (few-shot), greatly enhancing their versatility to synthesize realistic clinical narratives (Brown et al., 2020).

The rapid development of efficient, fine-tunable Small Language Models (SLMs) and multimodal SLMs is rapidly transforming healthcare. Unlike their bigger LLM counterparts SLMs can operate on-premises servers supporting local deployment, which significantly reduces costs, carbon footprint, and privacy concerns. These advances—as well as techniques like quantization, which reduce model size and accelerate inference time—are making it feasible to bring powerful language understanding and generation capabilities directly to the point of care, enabling clinical decision support, documentation, and patient interaction without reliance on high-end cloud infrastructure (Schick and Schütze, 2021; Dibia et al., 2024; Garg et al., 2025; Kim et al., 2025; Xie et al., 2025).

However, as GenAI becomes increasingly integrated into clinical workflows, it faces unique challenges specific to the medical domain. One of the most persistent and technically demanding issues is the precise segmentation and representation of multi-word clinical terms—such as “low blood pressure” or “low back pain” as well as their common abbreviations like “LBP.” Language models, including both LLMs and SLMs, often struggle to consistently recognize and encode such terms as unified clinical concepts. Inconsistent tokenization or lack of domain-specific context can lead to fragmented or distorted semantic representations, which in turn may compromise the accuracy of clinical information extraction, decision support, or narrative synthesis. This underscores the ongoing need for the development of more sophisticated prompting strategies tailored for healthcare, and the integration of concept embeddings, clinical practice guidelines, and real-world sample data to improve the realism and utility of synthetic clinical data (Beam et al., 2020; Chung et al., 2023; Han et al., 2023; Li et al., 2023).

To overcome the challenges of semantic fragmentation and context loss, recent research has shifted toward collaborative, multi-agent workflows to jointly tackle the complexity of clinical narratives and multimodal data. This new class of GenAI—known as Agentic AI or Generative Agents (GAs)—features multiple autonomous agents, each specializing in a particular task or sensory domain such as vision, language, audio, or touch (Chan et al., 2023; Park et al., 2023). That is, each agent can preside over its own dedicated LLM, prompted to address the unique challenges of its modality and dedicated task. By collaborating iteratively and sharing insights across these different modalities, GAs are able to process and integrate information from multiple sources—text, images, speech—generating content that is coherent and contextually rich (Qiu et al., 2024; Gridach et al., 2025; Hettiarachchi, 2025; Piccialli et al., 2025; Schneider, 2025).

Unlike traditional GenAI, GAs can gather real-time data from various sources, use different tools, design custom workflows, and refine their strategies through feedback, making them highly flexible and context aware. This collaborative approach allows specialized agents to solve complex tasks—such as SHDG—within a single workflow (Qiu et al., 2024). Some advanced workflows also incorporate language-agnostic concept models, which can further enhance the quality and performance of data synthesis (Daull et al., 2023; Barrault et al., 2024; Xie et al., 2024; Jin et al., 2025).

Building on this flexible and context-aware foundation, role-specific GAs are able to distribute cognitive workloads and leverage the strengths of multiple specialized agents within a single workflow (Xie et al., 2024; Jin et al., 2025). This not only allows for more effective problem solving but also enables the application of domain-specific expertise and enhances system robustness across a range of complex real-world tasks. For example, by leveraging advanced language models such as GPT-4, GAs can support workflows in scientific literature review writing (Abdurahman et al., 2025), sentiment analysis (Vasireddy et al., 2024), and the identification of mental health symptoms such as social anxiety from clinical interview data (Ohse et al., 2024). GAs also demonstrate promise in interpreting and structuring unstructured data from EHRs and clinical documentation (Yang et al., 2025). In particular, GPT-4.1 can complement—and sometimes rival—human expertise in global health education and data analysis (Thandla et al., 2024).

Low-code/no-code LLM platforms, such as Flowise,¹ Langflow,² and AutoGen³ (Kumar, 2023), are transforming the way GA-assisted SHDG workflows could be developed (Jeong, 2025). These platforms use drag-and-drop interfaces and natural language prompts, making it much easier to create, modify, and optimize workflows that leverage LLMs without the requirement of in-depth coding expertise. As a result, healthcare professionals and other non-technical users can now play an active role in designing and improving intelligent systems relevant to their needs. Moreover, a recent development called “Vibe-coding”—coined by Karpathy (2025)—illustrates this shift by enabling conversational co-development. Here, developers interact with GenAI tools—like GitHub Copilot—using plain language to iteratively refine and adjust code (Karpathy, 2025; Mayo, 2025).

In what follows, we present a protocolized GA-driven proof-of-concept for SHDG use cases aimed at creating novel synthetic clinical documents that emulate genuine EHRs while safeguarding patient privacy, preserving linguistic integrity, and maintaining informational accuracy under conditions of limited access to authentic EHR datasets.

The protocol described here aims to expand the responsible use and deployment of GenAI healthcare solutions to a broad spectrum of end-users—including those without specialized AI expertise—by providing clear, step-by-step guidance for designing workflows that leverage GAs for SHDG. To further support open-source adoption, reproducibility and practical application of our GA-assisted SHDG workflows we provide a GitHub repository.⁴

At the heart of our methodology is a modular data science infrastructure (DSI) Stack (Figure 1A), designed to organize and streamline the process of generating clinical data by breaking it down into a series of well-defined, interconnected workflows. It starts with ingesting and converting anonymized clinical notes from PDF to Markdown (FLOW01) (see text footnote 4), followed by pseudonymization to protect patient privacy (FLOW02) (see text footnote 4). Synthetic notes are then produced using LLMs, comparing standard prompting (FLOW03) (see text footnote 4) with a Generative Agent approach (FLOW03_AGENT_BASED). The pipeline ends with benchmarking (FLOW04) (see text footnote 4) using metrics for diversity, vocabulary similarity, semantic alignment, and classifier-based machine discernibility—enabling the comparison of the informational and linguistic features of synthetic, anonymized, and real EHRs at both the document and corpus levels. Included is a hands-on guide for responsible deployment of GA-assisted SHDG-workflows,⁵ implemented through open-source data science platforms—Hugging Face Spaces and Flowise—and powered by LLMs accessed via public cloud services such as Azure. This hybrid approach enables rapid prototyping and controlled sharing of workflows, while API key–secured inference endpoints ensure privacy compliance. The GitHub repository also provides lines of example source code (FLOW03) (see text footnote 4) that enables implementation of a privacy-first alternative to public cloud AI services called Ollama, allowing users to run and manage large language models locally while maintaining full data control and privacy compliance.

We start by reporting on a document-level assessment (Section 4.1), examining whether the generated data are contextually and semantically consistent with real EHRs. This is followed by a corpus-level assessment (Section 4.2) to assess whether our GA-assisted SHDG protocol adhered to established clinical data standards and preserved the statistical properties of the sample dataset (for details, see Section 3.5).

Our present work addresses a persistent barrier in digital health GenAI: the lack of accessible, interoperable, and privacy-preserving datasets that capture the diversity of real-world healthcare documentation (Eigenschink et al., 2023; Murtaza et al., 2023; Hernandez et al., 2025; Ibrahim et al., 2025; Loni et al., 2025). Building on our stepwise strategy for constructing a learning health system (LHS) (van Velzen et al., 2023), we argue that a fully integrated LHS will remain unattainable until these data challenges—especially in nursing and allied health (Tischendorf et al., 2025)—are resolved.

Echoing “On the Dangers of Stochastic Parrots” (Bender et al., 2021), we stress that healthcare must ask whether public cloud LLMs can be used in ways that are truly FAIR—Findable, Accessible, Interoperable, and Reusable—while mitigating risks such as high environmental and financial costs, opaque or biased data, and amplification of inequities. This requires curating and documenting high-quality datasets, aligning development with research and stakeholder values, and exploring approaches beyond ever-larger models.

A viable solution is the adoption of compact, fine-tunable SLMs. These small LLM alternatives can be deployed directly within healthcare facilities. Running locally not only reduces reliance on external cloud services but also lowers operational costs, decreases energy demands, and enhances data privacy. SLMs are increasingly capable of powering point-of-care applications—from clinical decision support to patient communication—while ensuring sensitive information remains within institutional boundaries (Schick and Schütze, 2021; Dibia et al., 2024; Garg et al., 2025; Kim et al., 2025; Xie et al., 2025).

By establishing a no-code, protocol for creating GA-assisted SHDG workflows—enabled by rapid prototyping platforms and further enhanced through fully automated, GenAI-assisted co-development—it is achievable to significantly lower the technical threshold for GenAI engagement. Leveraging the DSI stack as a generic blueprint architecture (Figure 1A), our methodological approach illustrates how modular, no-code frameworks can be systematically employed to streamline and democratize the creation of intelligent healthcare workflows.

In particular, the use of GA-assisted workflows to support clinical reasoning for nurse specialists—demonstrated through the “Nandalyse”¹¹ tool at the 2025 ASCENDIO conference—represents a major shift in healthcare technology development and adoption in the Netherlands (Kumar, 2023; Jeong, 2025; Tischendorf et al., 2025). These novel GenAI tools enable clinicians and other non-technical users to easily create and customize workflows using simple, modular, drag-and-drop interfaces, significantly lowering the barriers to participation.

A key aspect of our protocol for GA-assisted SHDG workflows is the explicit bridging of the gap between end users—such as clinicians, quality officers, and researchers—and the technical developers responsible for constructing and maintaining AI systems (as detailed in Figure 1B). In healthcare, this chiasm is often perpetuated by differences in language, priorities, and familiarity with digital tools. The explicit integration of human-in-the-loop (Alemohammad et al., 2024) in no-code GA-assisted SHDG workflows not only supports iterative co-development (Li et al., 2023) but also ensures that the system remains transparent and grounded in real-world clinical needs and documentation standards (Chung et al., 2023).

Notably, our methodology fosters closer collaboration between users and developers by providing open-source GitHub repositories (see text footnote 4), which make design choices, workflow logic, and evaluation criteria more transparent and verifiable. However, realizing the full promise of these platforms will require continued investment in user education, ongoing refinement of documentation and support resources, and the cultivation of communities of practice around open-source synthetic data generation.

The successful adoption of privacy preserving GA-assisted SHDG workflows in clinical practice, healthcare professionals requires more than technical proficiency alone (Kumar, 2023; Chew and Ngiam, 2025; Jeong, 2025); they must also possess a comprehensive understanding of key data privacy principles—such as pseudonymization and de-identification—to ensure patient confidentiality is maintained (Drechsler and Haensch, 2024; Rujas et al., 2025). Especially, clinicians should be aware of the inherent limitations of LLMs, including potential risks of bias, hallucination, and model drift, all of which may impact the fidelity and safety of synthetic health data (Alemohammad et al., 2024; Loni et al., 2025). Harnad (2025), notes that LLMs rely on stochastic patterns by capturing statistical regularities rather than achieving genuine semantic understanding, meaning that even fluent and coherent output may be inaccurate, irrelevant or misleading. Shojaee et al. (2025) further caution that the apparent reasoning proficiency of such models can deteriorate markedly as problem complexity increases—an “illusion of thinking” with significant implications for clinical safety. Collectively, these observations reinforce the imperative for multidimensional evaluation frameworks that integrate both surface-level measures (e.g., document length, lexical overlap) and deep-level metrics (e.g., semantic alignment, diversity) to rigorously assess the fidelity and practical utility of synthetic narratives across diverse, high-stakes clinical contexts (Shannon, 1948; Post, 2018; Zhang et al., 2019).

Importantly, working with GenAI in Healthcare mandates a clear understanding of evaluation metrics (Eigenschink et al., 2023; Abdurahman et al., 2025; Ibrahim et al., 2025). Because no single metric can holistically capture the quality of synthetic clinical text, a comprehensive assessment should integrate document- and corpus-level metrics (for overview see Tables 4, 5, respectively): document length and average word count for surface features, entropy for textual diversity (Shannon, 1948), BLEU for lexical overlap (Post, 2018), and BERTScore for semantic alignment (Zhang et al., 2019). Together, these complementary metrics provide a nuanced perspective on both the linguistic and informative value of synthetic narratives. Importantly, high performance on one dimension (e.g., diversity as measured by entropy) does not necessarily translate to strong semantic faithfulness (as measured by BERTScore). This distinction is particularly relevant when evaluating synthetic data for diverse clinical contexts, such as physiotherapy documentation in high-risk or emotionally fraught scenarios (e.g., cardiac, oncology, or orthopedic surgery). Our results (Table 6) reinforce the necessity of using an ensemble of surface and deep metrics to robustly assess the utility and fidelity of synthetic clinical narratives, supporting their appropriate integration into research and practice.

Our work is not without limitations. We identified four main operational and technical challenges that must be addressed to advance GA-assisted SHDG workflows. First, comparative evaluation of on-premises versus cloud-based AI models (Table 1) is needed to optimize trade-offs between data privacy, computational performance, and cost (Schick and Schütze, 2021; Garg et al., 2025; Xie et al., 2025). Parameter tuning—including the adjustment of prompt temperature and chunk size—must be empirically refined to balance realism and diversity in generated outputs (Peeperkorn et al., 2024). Second, while our GA prompts were intentionally crafted to encourage adherence to established clinical practice guidelines to generate clinically relevant outputs, we acknowledge that real-world clinical practice often diverges from these standards. Notably, intentional non-adherence to guidelines has been observed in up to 65% of cases within EHRs, frequently attributable to specific patient factors such as contraindications, comorbidities, or individual preferences (Arts et al., 2016). Therefore, for GA-assisted SHDG workflows to remain practical and reflective of authentic clinical scenarios, addressing this variability is essential. Consequently, our prompt design may have inadvertently constrained the clinical diversity of the synthesized narratives—a limitation that may have been further amplified by the small sample size of our authentic clinical EHR documentation dataset (N = 13 PDF documents). Future SHDG workflows should explicitly address these limitations through testing various prompt engineering techniques to better capture the variability inherent in real-world clinical narratives, or alternatively, by increasing the size of or sample dataset to encompass a broader range of patient presentations, care contexts, and documentation styles and a wider range of intentional non-adherence to guidelines. Implementing these changes—refining prompt-engineering techniques and expanding the dataset—would improve the representativeness of the generated outputs and better align them with the complexities of actual clinical practice. Third, we identified additional prompt engineering issues (Figure 3). These included the inadvertent introduction of gender bias—all synthetic patients turned out to be female—and inconsistent handling of abbreviations—where sample and markdown files contained frequent abbreviations, but synthetic data did not. We also found another prompt engineering issue that resulted in a one-size-fits-all approach in the generated therapy plans and reduced therapy frequency variability. Document length discrepancies, often resulting from repetition in sample or markdown files, also require further exploration to ensure consistent structural fidelity across datasets. Recent research underscores that prompt engineering is not a value-neutral process and that different components of a prompt can vary significantly in their robustness and susceptibility to bias. Mei et al. (2025) provide a comprehensive survey of “context engineering” strategies for large language models, highlighting how the choice, structuring, and sequencing of contextual elements can systematically influence model outputs. Zheng et al. (2025) further demonstrate that individual prompt components—such as instructions, examples, or delimiters—exhibit heterogeneous adversarial robustness, meaning that some parts are more vulnerable to manipulation or unintended bias than others. Related work on priming effects shows that the initial context or examples provided to a model can strongly condition its subsequent responses, amplifying or dampening biases and shaping output diversity (Zhao et al., 2021; Gallegos et al., 2024). Huang et al. (2025) demonstrate that targeted priming can exploit intrinsic weaknesses in large language models, revealing latent vulnerabilities that may not be apparent under standard prompting conditions. Meinke et al. (2024) further reveal that frontier-scale models are capable of sophisticated in-context behaviors, sometimes strategically adapting to earlier cues in ways that can subtly steer reasoning and decision-making. In addition, Gallegos et al. (2024) synthesize evidence that such biases can emerge not only from pre-training data but also from contextual framing and priming during inference. Together, these findings suggest that the gender bias, abbreviation inconsistencies, and homogenized therapy plans observed in our study may stem not only from prompt content but also from the structural composition, priming effects, and resilience of the prompts themselves. This reinforces the need for systematic evaluation and refinement of both prompt components and priming strategies to mitigate bias and improve variability in generated outputs. Fourth, the normalization of clinical text presents a significant methodological challenge, as variations in terminology—for instance, describing the same condition as “sciatica” versus “lumbosacral radicular pain syndrome”—often reflect individual practitioner preferences as well as institutional conventions (Suominen et al., 2013; U.S.-National-Library-of-Medicine, 2024). This semantic variability can be quantitatively monitored using mutual information-based evaluation metrics (see Section 3.4); in our study, we applied the average bigram pointwise mutual information (PMI) metric. However, the robustness of this metric across more diverse datasets and practitioners from different clinical specialties remains to be fully validated. Thus, relying solely on either surface-level metrics or deep semantic measures is insufficient. Instead, a comprehensive evaluation of the faithfulness of SHDG demands the integration of both approaches (Eigenschink et al., 2023; Smolyak et al., 2024; Chen et al., 2025b; Ibrahim et al., 2025; Scherr et al., 2025).

Based on our findings, we recommend prioritizing the continued development and refinement of GA-assisted SHDG-workflows, ensuring that these toolchains remain accessible, transparent, and customizable for a diverse range of clinical users and researchers through the provision of open-source GitHub repositories (see text footnote 4). Our recommendations align with ongoing efforts by the RUAS Healthcare DataLab to advance innovation in care through collaborative, technology-driven solutions that integrate open, adaptable tools into diverse healthcare contexts.¹² In parallel, our talent program fosters and equips RUAS students with the skills and expertise required to become proficient data science professionals, thereby strengthening the capacity for data-driven innovation within the Dutch healthcare sector. To maximize impact, future initiatives should emphasize robust user education on essential data privacy concepts—such as pseudonymization and de-identification—and foster a deeper understanding of both the capabilities and limitations of no-code GenAI technologies, particularly regarding risks like bias, hallucination, and model drift. Comprehensive evaluation frameworks should integrate both surface-level and deep semantic metrics to robustly assess the linguistic and clinical fidelity of synthetic narratives, while methodological improvements—including enhanced prompt engineering to address issues of bias, guideline adherence, and structural consistency, as well as adherence to normalization standards—will be essential in capturing the diversity inherent in real-world clinical documentation. Comparative assessments of on-premises versus cloud-based deployments, plus empirical tuning of generative parameters, are also advised to optimize privacy, cyber security, cost, and performance trade-offs for varied healthcare settings. Finally, building active communities of practice and cultivating collaborative, iterative engagement between end-users and developers will be critical to realizing truly FAIR (Mons et al., 2017), AI-ready LHS that can adapt to the complexities and dynamic requirements of modern clinical environments (Huerta et al., 2023; Verhulst et al., 2025).

Our protocolization of privacy-preserving, GA-assisted SHDG workflows underscores the critical importance of maintaining transparency by keeping humans actively involved in the process—a principle known as human-in-the-loop. Specifically, we standardized the use of a modular, no-code system in which every step—from designing AI prompts, to providing data inputs, to defining how outputs are applied—can be easily inspected, understood, and refined without programming expertise. This approach enables users to trace how the LLM generates clinical narratives and to make step-by-step improvements over time. In parallel, the integration of an open-source GitHub repository (see text footnote 4), anchored by our modular DSI Stack (Figure 1A), supports the design and deployment of secure, reproducible, and scalable GA-assisted SHDG workflows across both on-premises and public cloud environments. By incorporating Docker-based containerization and API key–secured inference endpoints via Hugging Face Spaces, we ensure controlled access, consistent performance, and community-driven enhancement, while shielding users from underlying infrastructure complexity and maintaining full transparency for healthcare and research applications.