Harmonizing Vietnamese traditional and western medicine: a digital health blueprint and the IMKE architectural framework

1 Introduction

Healthcare systems globally face an escalating burden of chronic non-communicable diseases (NCDs), which account for nearly 75% of all mortality [1]. While Western biomedicine has achieved extraordinary clinical success in acute trauma, surgical intervention, and infectious diseases through targeted pharmacological protocols [3], there is growing recognition that this reductionist approach struggles to fully address the complex, systemic nature of chronic morbidities [2]. By isolating individual biological targets, modern biomedicine often overlooks the underlying physiological dysregulations driving conditions like metabolic syndrome, autoimmune disorders, and neurodegeneration [4, 5].

Consequently, patients increasingly turn to Eastern medical traditions, such as traditional Chinese medicine (TCM) and its deeply localized derivative, Vietnamese traditional medicine (VTM). These holistic systems emphasize the harmonization of the human body through personalized, multi-compound herbal formulas targeting functional “patterns of disharmony” rather than isolated molecular biomarkers [6, 7]. The World Health Organization (WHO) actively advocates for the integration of traditional medicine into national healthcare systems, recognizing its role in evidence-based complementary care for billions globally [8]. In Vietnam, this integration is a highly active clinical reality. VTM is practiced alongside Western biomedicine in over 90% of provincial hospitals, supported by national policies officially mandating dual-paradigm treatments for high-prevalence chronic diseases [9, 10].

Despite this widespread clinical co-utilization, integrative care at the point of delivery remains dangerously fragmented, creating severe barriers to patient safety. This fragmentation is driven by three foundational gaps. First, a knowledge gap: VTM therapeutic wisdom remains trapped in unstructured textual formats—ranging from ancient classical manuscripts to modern, handwritten practitioner notes—rendering it inaccessible for digital health integration and modern clinical workflows [11, 12]. Second, an evidence gap: a profound epistemological incompatibility exists between Western medicine’s demand for molecular biomarkers and VTM’s phenomenological diagnoses, hindering health data interoperability [13, 14]. Third, and most critically, a safety gap: the underreported risk of herb-drug interactions (HDIs). Epidemiological surveys reveal that over 42% of patients concurrently using pharmaceutical drugs and traditional herbal remedies do not disclose this co-administration to their primary physicians. This creates a massive “shadow pharmacovigilance” crisis, leaving clinicians prescribing in the dark [15, 16].

Digital health technologies, particularly Artificial Intelligence (AI), offer a viable bridge for these epistemological and structural divides. The convergence of large language models (LLMs) [17], biomedical knowledge graphs (KGs) [18, 19], and graph neural networks (GNNs) [20] now provides the computational infrastructure necessary to unify divergent medical data. Predictive analytics driven by deep learning on electronic health records [21] and big data machine learning [22] have already catalyzed a fundamental shift in evidence-based medicine. Furthermore, recent dual-paradigm clinical guidelines have demonstrated the clinical viability of integrated protocols [23]. Recently, significant methodological advancements have been proposed to organically integrate TCM and Western medicine clinical practice guidelines [84, 85]. However, while prior research has successfully applied AI to isolated traditional medicine tasks—such as TCM symptom mapping or molecular docking for active compounds [24, 25]—these efforts remain siloed. Existing platforms function primarily as static pharmacological databases for researchers rather than dynamic clinical tools, critically lacking the capacity for real-time safety inference and explainable clinical decision support (CDSS) at the point of care [26, 27].

Figure 1 illustrates the conceptual role of the IMKE as a bridge between the VTM and Western biomedical paradigms.

To address these urgent clinical bottlenecks, this paper presents a comprehensive review of AI in integrative healthcare and proposes the Integrative Medicine Knowledge Engine (IMKE). Rather than presenting an isolated algorithm, this paper introduces a comprehensive digital health blueprint designed to safely harmonize VTM and Western medicine in clinical practice. The contributions of this work are threefold:

State-of-the-art analysis: We critically evaluate existing traditional medicine AI, identifying the critical absence of end-to-end, cross-paradigm safety monitoring systems.

Architectural blueprint: We conceptualize the IMKE framework, a three-stage clinical informatics pipeline comprising an advanced Natural Language Processing (NLP) module for unstructured VTM knowledge extraction, an integrative knowledge graph (IKG) for cross-paradigm ontology alignment, and an AI-driven module for predictive HDI safety monitoring.

Simulated operational scenarios: We illustrate the architecture’s logical data flow and reasoning capacity through three simulated clinical use-cases. These demonstrate how the IMKE translates complex data into explainable, actionable CDSS alerts that augment physician judgment and mitigate clinical alert fatigue.

The central hypothesis of this work is that fusing VTM phenomenological concepts with Western pharmacological data into a multi-relational knowledge representation, combined with relation-aware computational reasoning, can improve the sensitivity and interpretability of novel HDI detection compared with single-paradigm databases.

By formalizing this architecture, this work provides a scalable, mathematically rigorous, and culturally sensitive template for AI-mediated integrative healthcare, offering a robust foundation for patient safety and future empirical deployment in Vietnam and globally.

The remainder of this paper is structured as follows. Section 2 presents the state-of-the-art in integrative medicine AI. Section 3 details the clinical and technical design rationale of the IMKE architecture. Section 4 illustrates the framework through simulated clinical scenarios. Section 5 discusses open clinical challenges, regulatory perspectives, and the future development roadmap. Section 6 concludes the paper.

2 State-of-the-art in integrative medicine AI

To establish a rigorous clinical informatics and architectural foundation for the IMKE framework, this section comprehensively reviews the current landscape of artificial intelligence (AI) applied to traditional Asian medicine and Western integrative practices. Guided by clinical insights and expert consultations with practitioners from the Hospital of Traditional Medicine,¹ Ho Chi Minh City, this review critically evaluates existing computational frameworks across divergent medical paradigms to identify pivotal gaps in care delivery and patient safety.

2.1 Search strategy and divergent paradigms

A comprehensive literature search was conducted across IEEE Xplore, PubMed, and Scopus, including articles published from January 2014 to May 2024, restricted to English and Vietnamese languages. The precise search string utilized combinations of terms related to artificial intelligence, traditional medicine, integrative medicine, knowledge graphs, clinical decision support, and herb–drug interactions. The fundamental clinical challenge identified across the literature is the epistemological divide between Western and Eastern medicine.

Western medicine’s epistemology is rooted in Cartesian reductionism, decomposing biological complexity into measurable components (genes, proteins) targeted with precise interventions [28]. While this paradigm has generated transformative clinical advances validated by randomized controlled trials (RCTs) [29], it struggles with chronic, multifactorial diseases where causality is distributed across systemic biological networks [4]. Conversely, Vietnamese traditional medicine (VTM)—systematized in the 18th-century text Hải Thượng Y Tông Tâm Lĩnh—operates on a holistic framework [6]. Diagnosis is pattern-based (e.g., “liver Yang rising”), integrating phenomenological descriptors that resist quantification without substantial ontological translation [5]. As Monnais (2019) observes, this creates a fundamental language barrier that persists in contemporary clinical workflows and regulatory frameworks, hindering collaborative care [14].

2.2 Taxonomy of AI applications in traditional medicine

The application of AI to traditional medicine has evolved from rudimentary digitization to deep-learning-driven multi-domain reasoning [11]. To provide a comprehensive technical overview, Table 1 summarizes the primary machine learning architectures currently applied in integrative medicine, highlighting their representative studies and inherent limitations.

Table 1. Artificial intelligence and machine learning techniques in integrative medicine research.

The application of AI to traditional medicine has evolved from rudimentary digitization to deep-learning-driven multi-domain reasoning [30]. Based on our comprehensive review, we categorize the current research landscape into three primary technological pillars (Figure 2):

Natural language processing (NLP) for medical text mining: The foundational step involves extracting traditional concepts (herbs, formulas, symptoms) from unstructured texts, a task fundamentally challenged by the complexity of biomedical entity and adverse effect recognition [31]. Transformer architectures [32] have revolutionized this domain. Domain-specific pre-training—demonstrated by models such as BioBERT [33], ClinicalBERT [34], and SciBERT [35]—has substantially improved performance on biomedical NLP tasks. However, while these models excel in English, VTM presents a unique challenge spanning three linguistic registers—modern Vietnamese, Sino-Vietnamese, and classical Han-Nom. Although PhoBERT provides a robust foundation for Vietnamese NLP [12], domain-specific relation extraction for VTM remains an unsolved challenge.

Knowledge representation and graph engineering (KG): KGs are heterogeneous networks representing real-world entities and typed semantic relationships [36]. In biomedicine, KGs like the UMLS metathesaurus and SPOKE have become central architectural patterns for reasoning over clinical knowledge [18]. For traditional medicine, the core challenge lies in ontology alignment—mapping pattern-based diagnoses to Western biomedical standards (e.g., ICD-10) and utilizing graph embedding techniques (e.g., TransE, R-GCN) to preserve semantic structures [37, 38]. Recent comprehensive reviews [39] and methodological advances [40] in TCM knowledge graph construction underscore the immense clinical value of these technologies, yet domain-specific ontology alignment for VTM remains largely unsolved.

Clinical Decision Support Systems (CDSS): The ultimate translational goal of AI is to augment human clinical judgment [27]. While AI has been applied to diagnostic instruments like tongue and pulse analysis [41], there is a critical lack of end-to-end CDSS frameworks that monitor cross-paradigm safety, specifically herb-drug interactions (HDIs).

Based on our literature review, we categorize the current research landscape into three primary technological pillars (Figure 2):

2.3 State-of-the-art frameworks: a comparative analysis

Over the past decade, significant efforts have been made to digitize traditional Chinese medicine (TCM) and explore network pharmacology [30]—a paradigm first established by Hopkins to transition drug discovery from reductionist single-target models to systems-level, multi-target analyses [42]. Landmark databases such as TCMID [43], SymMap [25], and HERB [24] have provided extensive mappings between herbs, chemical compounds, and biological targets. Beyond static databases, recent conceptual advances have moved toward artificial-intelligence-based precision TCM using network pharmacology [86] and complex-systems evidence frameworks [87]. However, as noted by clinical experts, these platforms and conceptual frameworks do not yet constitute dynamic cross-paradigm clinical safety tools.

Table 2 compares these databases, evidence frameworks, and guideline methodologies with the proposed IMKE architecture.

Table 2. State-of-the-Art comparison of integrative medicine databases, evidence frameworks, guideline methodologies, and the proposed IMKE architecture.

The four recent conceptual contributions clarify the specific research gap addressed by IMKE. Zhang et al. position network pharmacology as a foundation for AI-enabled precision TCM by linking herbal components to molecular targets and efficacy-oriented predictions [86]. Wei et al. advance this systems perspective at the evidence level through a four-phase framework intended to organize complex, multi-source Chinese medicine evidence [87]. These contributions strengthen the theoretical basis for systems-oriented traditional medicine research, but neither specifies an end-to-end clinical informatics pipeline for extracting unstructured VTM records, aligning them with Western clinical terminologies, and generating patient-level HDI safety alerts.

Jin et al. and Xie et al. address a complementary implementation layer by proposing methodological principles for the organic integration of traditional Chinese and Western medicine clinical practice guidelines [84, 85]. Their work provides an important governance and guideline-development foundation, whereas IMKE focuses on the computational operationalization of integrative guidance at the point of care. Accordingly, IMKE’s proposed novelty lies not in replacing these evidence and guideline frameworks, but in connecting them through a technology-agnostic architecture that combines multilingual NLP, many-to-many ontology mapping, multi-relational HDI inference, and explainable CDSS delivery.

2.4 Herb-drug interaction safety: the critical clinical gap

Herb-drug interactions (HDIs) represent the most consequential challenge at the interface of integrative medicine. HDIs encompass both pharmacokinetic interactions (e.g., CYP450 enzyme modulation) and pharmacodynamic interactions (synergistic or antagonistic effects) [44]. Epidemiological studies indicate that clinically significant HDIs affect 5%–10% of patients using concurrent therapies, with a vast majority of cases undetected by routine pharmacovigilance [16].

Foundational data-driven approaches have long demonstrated the power of predicting unknown drug effects and interactions from large-scale clinical databases [45]. Building on this, while recent advances have utilized graph neural networks (GNNs) to predict drug-drug interactions (DDIs) with high accuracy [20, 46, 47], and data mining techniques have successfully extracted interactions from clinical text sources [48], HDI prediction specifically tailored to VTM remains sparse. Although machine learning classifiers have shown promise in predicting TCM herb-drug pairs [49], the complexity of herbal formulations—where multiple active compounds simultaneously modulate different pathways—renders simple rule-based detection inadequate, necessitating multi-hop graph reasoning capabilities [37].

Recent studies published in IEEE Access have demonstrated the immense potential of graph neural networks in modern healthcare [50]. Specifically, state-of-the-art architectures have successfully integrated Transformer models with biomedical knowledge graphs (KITE-DDI) [46] and utilized Graph Attention Networks (KGAT) [47] to accurately predict complex drug-drug interactions, providing a strong methodological baseline for the IMKE’s safety module. The complexity of herbal formulations—where multiple active compounds simultaneously modulate different pathways—renders simple rule-based detection inadequate, necessitating multi-hop graph reasoning capabilities [37]. To illustrate the clinical severity of this gap, Table 3 details key examples of highly consequential herb-drug interactions that the IMKE framework is designed to detect and prevent.

Table 3. Representative herb–drug interactions requiring GNN-Based detection in the Integrative Medicine Knowledge Engine (IMKE).

2.5 Policy, cultural dimensions, and institutional alignment

The integration of traditional and Western medicine is highly dependent on institutional and regulatory frameworks. Vietnam’s Ministry of Health (MoH) actively promotes dual-paradigm clinical protocols. A landmark expression of this policy is the official MoH Guidelines (Decision No. 3890/QD-BYT, 2020; updated 2026), which establish standardized protocols combining VTM pattern-based diagnosis with ICD-10-aligned Western differential diagnosis for high-prevalence conditions [9, 10].

Crucially, these guidelines mandate that VTM practitioners must not rely solely on traditional medicine but must combine it with Western diagnostic methods to ensure patient safety. This regulatory foundation provides the IMKE with an authoritative, operationally tested schema for knowledge graph construction. However, sociological reviews confirm that even well-designed policies often produce institutional coexistence rather than genuine collaboration [13, 51]. Discussions with practitioners highlight that without computational mediation, the risk of undetected HDIs remains high.

2.6 Identification of research gaps and IMKE’s positioning

Synthesizing the literature review and comparative analysis (Table 1), we identify three pivotal gaps that the IMKE framework is designed to bridge:

The end-to-end processing gap: Existing databases (TCMID, SymMap) lack automated, VTM-specific NLP pipelines capable of dynamically extracting multi-relational knowledge from unstructured, multi-script Vietnamese texts.

The cross-paradigm safety gap: Current platforms utilize the principles of network pharmacology [42] primarily to prove therapeutic efficacy (e.g., identifying target overlap for disease treatment), but largely ignore predictive safety. No existing traditional medicine platform features automated, GNN-driven prediction of novel herb-drug interactions spanning both VTM and Western pharmacopeia.

The clinical translation gap: Despite the existence of MoH integrative guidelines [9], no computational system has operationalized these rules into a machine-readable format integrated with an explainable CDSS.

The IMKE addresses these gaps by treating the MoH guidelines as the authoritative seed for its integrative knowledge graph (IKG), grounding all mappings in regulatory-grade evidence before extending into probabilistic GNN-based HDI discovery. The architectural blueprint to achieve this is detailed in Section 3.

3 Methodology: the IMKE framework

As a Hypothesis and Theory article, the IMKE defines functional requirements and testable methodological pathways rather than prescribing a fixed implementation stack. Technologies named below—including PhoBERT, Neo4j, R-GCN, SHAP, Docker, and particular optimization functions—are illustrative candidate instantiations selected to demonstrate how the conceptual requirements might be operationalized. The equations, query examples, interface mockups, and evaluation dimensions are therefore explanatory artifacts rather than a finalized protocol, trained system, or claim of achieved performance. Future empirical studies must compare alternative encoders, graph stores, relation-aware learning methods, explanation techniques, and deployment architectures before any implementation choice is fixed.

The IMKE framework is designed as a modular, three-stage architectural pipeline that transforms unstructured Vietnamese traditional medicine (VTM) knowledge into actionable, computable clinical guidance. The architecture is conceptualized to operate in a real-time clinical environment, processing multi-modal patient data—combining Western electronic health records (EHR) and VTM practitioner notes—through successive layers of NLP processing, knowledge graph reasoning, and graph-neural-network-based safety prediction.

Figure 3 summarizes the proposed functional architecture and its three interconnected layers.

3.1 Stage 1: NLP-based knowledge extraction and design rationale

The foundation of the IMKE is a specialized NLP pipeline that transforms unstructured VTM texts into machine-readable structured triples: (Entity_1, Relationship, Entity_2). This pipeline addresses the unique linguistic complexity of the VTM corpus, which spans three registers: modern Vietnamese, Sino-Vietnamese, and classical Han-Nom scripts.

3.1.1 Data sources and corpus construction

The framework’s primary corpus leverages regulatory-grade documents to ensure clinical reliability. The core structured knowledge source is the Vietnamese Ministry of Health’s (MoH) official integrative medicine guidelines (2020) [9] and its 2026 expansion [10]. These protocols provide institutionally validated dual-paradigm diagnostic criteria, bat cuong (eight-principle) classifications, classical herbal formulae, and ICD-10-aligned Western guidance. This authoritative seed is supplemented by canonical VTM classical texts [7, 52], contemporary peer-reviewed pharmacological studies [16], and the Vietnamese National Formulary.

3.1.2 Named entity recognition (NER): context-aware Vietnamese encoders

The NER module identifies five entity classes: Herbs, Symptoms, VTM Patterns of Disharmony, Western Diseases, and Therapeutic Actions.

Architectural considerations: The encoder backbone of the NLP pipeline necessitates an advanced, context-aware Transformer architecture to handle linguistic complexity. For instance, localized foundation models such as PhoBERT-base [12] serve as illustrative options alongside general multilingual models such as mBERT. Because Vietnamese occupies only a subset of multilingual training corpora, generic tokenization may increase out-of-vocabulary fragmentation, whereas models using syllable-aware byte-pair encoding are theoretically better aligned with Vietnamese linguistic structure. Consequently, fine-tuning a domain-specific contextual encoder—for example, following principles demonstrated by BioBERT [33]—is hypothesized to improve the representation of semantic boundaries and enable attention over clinically meaningful concepts rather than fragmented syllables. The optimal encoder and tokenization strategy remain empirical questions to be evaluated against multilingual, language-specific, and domain-adapted alternatives.

3.1.3 Relation extraction and triple generation

Following NER, a joint entity-relation extraction architecture [53] classifies entity pairs into predefined relation types derived from VTM theory and Western pharmacology. Table 4 outlines the proposed ontology relation labels extracted by the NLP module to construct the IKG.

Table 4. Proposed relation extraction labels for the IMKE ontology.

For example, from unstructured text, the system extracts: (Glycyrrhiza uralensis)—[HARMONIZES]→ (Toxic_Compounds). These structured triples form the raw input for the next stage.

3.2 Stage 2: integrative knowledge graph (IKG) construction

The IKG is the central data structure of the IMKE, formally defined as a directed, labeled multigraph G=(V,E,R,ψ), where V is the set of entity nodes, E⊆V×R×V is the set of typed relational edges, R is the set of relation types spanning both paradigms, and ψ:E→[0,1] is a confidence weight [36].

3.2.1 Ontology schema design and data fusion

To enable cross-paradigm reasoning, the IKG enforces a strict ontology schema connecting macroscopic VTM phenomena to microscopic Western biomedicine:

VTM Sub-graph: (VTM_Pattern) [TREATS]- (Herbal_Formula) -[INCLUDES] (Herb)

Bridge Layer: (Herb) -[CONTAINS] (Active_Compound)

Western Sub-graph: (Active_Compound) -[TARGETS] (Gene/Protein) [INHIBITS/ACTIVATES]—(Western_Drug)—[TREATS] (Western_Disease)

Ontology alignment is performed via lexical matching, semantic embedding similarity, and expert manual curation [37]—a necessary multi-strategy approach given the global need for a standardized nomenclature bridging East and West [54]. The IKG fuses VTM triples with established external databases, including DrugBank [55], Gene Ontology, KEGG, and the UMLS Metathesaurus, utilizing an entity resolution algorithm to collapse synonymous nodes.

It is crucial to note that VTM patterns and Western ICD-10 diagnoses do not exhibit a direct one-to-one equivalence. The IKG is specifically architected to represent these relationships as many-to-many, evidence-qualified mappings, reflecting the complex clinical reality where a single Western disease may manifest as multiple VTM patterns, and vice versa.

Figure 4 illustrates the ontology schema and the proposed cross-paradigm alignment of traditional and biomedical entities.

3.2.2 Graph traversal and illustrative query pathway

When instantiated within a graph database management system (e.g., Neo4j), the IKG enables explicit topological reasoning. To identify whether a prescribed herb shares a biological target pathway with a pharmaceutical drug, a future CDSS could execute targeted graph traversals. For instance, to detect potential interactions between Uncaria rhynchophylla (Câu đằng) and Lisinopril, an illustrative Cypher Query Language (CQL) implementation could be structured as follows:

// IMKE Cypher Query for Cross-Paradigm Interaction Detection

MATCH path = (h:Herb {name: "Uncaria rhynchophylla"})

-[r1:CONTAINS]->(c:Compound)

-[r2:TARGETS]->(t:Target)

<-[r3:INHIBITS]-(d:WesternDrug {name: "Lisinopril"})

RETURN c.name AS Active_Ingredient, t.name AS Shared_Target, type(r2) AS Herb_Action, type(r3) AS Drug_Action

This query illustrates how multi-hop traversal could identify candidate mechanistic overlaps not represented as explicit pre-recorded interaction pairs. Equivalent traversals could be implemented using other graph query languages or storage systems.

3.3 Stage 3: clinical decision support system (CDSS)

The CDSS represents the proposed clinical interface to the IKG [27]. Its core functional requirement is to surface candidate Herb-Drug Interactions (HDIs) not captured by deterministic graph traversal alone, regardless of the specific predictive algorithm eventually selected.

3.3.1 Relation-aware interaction prediction

Because biological interactions are inherently probabilistic, the IMKE framework proposes relation-aware graph learning as one methodological pathway for predictive safety monitoring. Architectures capable of handling heterogeneous edge types—such as Relational Graph Convolutional Networks (R-GCN) [38]—are illustrative candidates alongside standard GCNs [56], attention-based models [57], and inductive frameworks such as GraphSAGE [58]. Methods that treat all edges equivalently may fail to preserve the clinically opposite meanings of relations such as [TREATS] and [CAUSES_ADVERSE_EVENT]. An R-GCN-like architecture addresses this requirement by allocating relation-specific transformations to different edge types. One possible mathematical representation of this updating process is:where hi(l+1) is the hidden representation of node i at layer l+1, R denotes the set of relation types, Nir represents the set of neighbor indices of node i under relation r, Wr(l) is a relation-specific weight matrix, ci,r is a problem-specific normalization constant (e.g., computed based on node degrees), W0(l) represents the self-loop weight matrix preserving the node’s original features, and σ is a non-linear activation function.

hi(l+1)=σ(∑r∈R∑j∈Nir1ci,rWr(l)hj(l)+W0(l)hi(l))(1)

Equation 1 illustrates how relation-specific neighbor information and a self-loop contribution can be combined to update a node representation. Figure 5 illustrates relation-aware message passing using R-GCN as one possible model, rather than a required implementation.

3.3.2 Illustrative objective function for HDI prediction

The empirical training of a future predictive module would require optimizing an objective function against known HDI databases [16]. One standard option for link prediction is Binary Cross-Entropy Loss with Negative Sampling, although ranking, contrastive, and calibrated probabilistic objectives should also be compared. Under this illustrative option, the probability of an interaction edge can be modeled as: e=(vherb,INTERACTS\_WITH,vdrug)where Ω+ is the set of true interactions and Ω− represents negatively sampled non-interacting pairs.

L=−∑(s,r,o)∈Ω+logσ(f(s,r,o))−∑(s,r,o′)∈Ω−log(1−σ(f(s,r,o′)))(2)

Equation 2 illustrates a possible training objective combining positive interaction examples and negatively sampled pairs; it is not a prescribed component of the conceptual framework. Figure 6 presents an illustrative CDSS dashboard showing how candidate interaction alerts and their supporting explanations could be communicated to clinicians; it does not display empirical model outputs.

3.3.3 Explainability and integration pathways

To prevent alert fatigue and build clinician trust [59, 60], every IMKE alert should be accompanied by a reviewable computational rationale. Post-hoc feature-attribution frameworks, such as SHAP [61, 62], represent one illustrative option that could be applied to learned graph representations alongside explicit path tracing (e.g., identifying a candidate CYP450 inhibition pathway). Other explanation methods, including counterfactual, rule-based, and inherently interpretable graph approaches, remain eligible for comparison. For eventual deployment, the IMKE is envisioned to integrate with existing EHRs through interoperability standards such as HL7 FHIR APIs [63]; containerized microservices (e.g., Docker) are one possible deployment pattern for reducing healthcare IT integration friction [64].

It must be explicitly noted that attribution scores—including those produced by SHAP or alternative methods—describe computational rationale but do not inherently establish biological or causal mechanisms. All generated pathways and explanations would require review by clinical pharmacologists before clinical acceptance.

4 Conceptual framework illustration through simulated scenarios

To demonstrate the IMKE framework’s architectural design, reasoning capacity, and clinical utility, we present three simulated conceptual case studies spanning distinct chronic disease domains. Because this paper proposes an architectural blueprint, these scenarios do not represent live clinical trials but serve as theoretical step-by-step executions to demonstrate the data flow pipeline—from raw text input to NLP processing, graph traversal, GNN inference, and explainable CDSS output.

Each case study is designed to demonstrate a different category of herb-drug interaction mechanism (pharmacodynamic, pharmacokinetic, and toxicological). Table 5 details the specific interaction mechanisms mapped across the three scenarios and provides a comparative summary of the CDSS pipeline outputs.

Table 5. Comparative summary of IMKE simulated executions.

4.1 Scenario 1: essential hypertension (pharmacodynamic synergism)

4.1.1 Clinical profile

“Mr. Nam,” a 62-year-old male, presents with chronic headaches and dizziness. Western diagnosis: Essential Hypertension (ICD-10: I10), BP 150/95 mmHg, prescribed Lisinopril 10 mg (ACE inhibitor, DrugBank DB00722). VTM presentation: dizziness (huyễn vựng), tinnitus (nhĩ minh), irritability; pulse diagnosis: wiry and rapid (huyền sác); tongue: red with yellow coating. VTM pattern: Liver Yang Rising (Can dương thượng cang)—an upward surge of physiological energy due to Yin deficiency [28]. From a pharmacological perspective, in vitro and in vivo studies summarized by Ndagijimana et al. confirm that rhynchophylline, a major alkaloid in Uncaria rhynchophylla, exhibits significant calcium-channel-blocking activity [82]. This provides empirical support for potential additive hypotension when co-administered with ACE inhibitors.

4.1.2 IMKE step-by-step execution pipeline

Step 1 (Clinical Input): The physician inputs the patient’s EHR data (Lisinopril) and unstructured VTM clinical notes into the CDSS interface.

Step 2 (NLP Processing): A context-aware Vietnamese clinical-language encoder would identify the key entities: [Western_Drug: Lisinopril], [VTM_Symptom: huyễn vựng], and [VTM_Pattern: Can dương thượng cang]. PhoBERT is one illustrative candidate for this function rather than a prescribed component.

Step 3 (Graph Traversal & Ontology Mapping): The entity-linking function would map the VTM pattern to Essential Hypertension through evidence-qualified shared phenotypes. A standards-compatible graph traversal—illustratively expressible using Neo4j/Cypher—could retrieve the MoH 2020 official integrative protocol [9] and the formula “Thiên ma câu đằng ẩm” (Gastrodia and Uncaria Decoction) [65], including [Herb: Uncaria rhynchophylla].

Step 4 (Relation-Aware Inference): A heterogeneous graph-learning method could evaluate the retrieved herb against the prescribed drug and generate a candidate high-risk alert from the topological evidence surrounding the edge (Uncaria rhynchophylla, INTERACTS_WITH, Lisinopril). R-GCN represents one possible instantiation.

Step 5 (Explainability & CDSS Output): An explanation layer could combine explicit path tracing with a suitable attribution method (e.g., SHAP) to identify graph evidence contributing to the alert. An illustrative mechanistic path is: Uncaria rhynchophylla → [CONTAINS] → Rhynchophylline → [INHIBITS] → Calcium Channel (L-type) → [VASODILATION].

Illustrative System Output: In this conceptual scenario, the CDSS would display a high-priority additive-hypotension alert and present monitoring or substitution options for clinician review rather than issue an autonomous recommendation.

4.1.3 CDSS output

CLINICAL ALERT (High Priority): Potential Synergistic Hypotension. Mechanism: Rhynchophylline (Uncaria rhynchophylla) exhibits calcium channel blocking activity. Concurrent use with Lisinopril (ACE inhibitor) may produce additive blood pressure reduction, elevating risk of symptomatic hypotension (dizziness, syncope). Recommendation: (a) Monitor BP continuously during initiation; (b) Consider dose reduction of herbal formula; (c) Alternative formula substitution: Qi Ju Di Huang Wan (without Uncaria) may be appropriate for milder patterns.

4.2 Scenario 2: type 2 diabetes (therapeutic synergy discovery)

4.2.1 Clinical profile

“Mrs. Lan,” a 58-year-old female, five-year history of poorly controlled Type 2 Diabetes (ICD-10: E11). Fasting glucose 9.4 mmol/L, HbA1c 8.7%. Current medications: Metformin 1000 mg twice daily (DrugBank DB00331), Glibenclamide 5 mg once daily (DrugBank DB01016). VTM presentation: fatigue, polydipsia, recurrent UTIs; pulse: deep, thin, weak (tram, te, nhuoc); tongue: pale with thin white coating. VTM pattern: Tiêu khát (wasting and thirsting disorder; Chinese: 消渴), characterized by Kidney-Spleen Qi deficiency. Clinical and pharmacological evaluations indicate that charantin-containing preparations derived from Momordica charantia can promote pancreatic beta-cell activity and insulin secretion [66, 67]. This supports a plausible biological mechanism for synergistic hypoglycemia when combined with sulfonylureas such as Glibenclamide.

4.2.2 IMKE step-by-step execution pipeline

Step 1 (Clinical Input): The physician inputs Western meds (Metformin, Glibenclamide) and VTM symptoms.

Step 2 (NLP Processing): A context-aware clinical-language encoder would extract [Western_Drug: Glibenclamide], [Western_Drug: Metformin], and [VTM_Pattern: Tiêu khát].

Step 3 (Graph Traversal & Ontology Mapping): The IKG bridges Tiêu khát to Type 2 Diabetes (confidence weight ψ=0.82). Traversal of the MoH 2026 guidelines [10] retrieves the base formula Lục vị địa hoàng hoàn, supplemented clinically with [Herb: Momordica charantia] (Bitter melon).

Step 4 (Relation-Aware Inference): A future heterogeneous graph-learning module could evaluate the herb–drug pairs and generate a candidate interaction alert between Momordica charantia and Glibenclamide.

Step 5 (Explainability & CDSS Output): An illustrative combination of path tracing and post-hoc attribution could highlight the following biological pathway: Momordica charantia → [CONTAINS] → Charantin → [ACTIVATES] → Pancreatic Beta Cells → [INCREASES] → Insulin Secretion → [OVERLAPS MECHANISM WITH] → Glibenclamide [66, 67].

Illustrative System Output: In this conceptual scenario, the CDSS would present a dual-channel output: a potential compounded-hypoglycemia warning and a hypothesis-generating synergy flag requiring specialist assessment before any medication adjustment.

4.3 Scenario 3: rheumatoid arthritis (pharmacokinetic toxicity)

4.3.1 Clinical profile

“Ms. Hoa,” a 55-year-old female, eight-year history of Rheumatoid Arthritis (ICD-10: M05). CRP 24 mg/L (elevated), ESR 58 mm/hr. Current medication: Methotrexate 15 mg/week (DrugBank DB00563). VTM presentation: bilateral joint pain worsened by cold and damp weather; pulse: wiry and tight (huyền khẩn); tongue: pale with white greasy coating. VTM pattern: Wind-Cold-Damp Bi Syndrome (phong hàn thấp tý), a pattern of external pathogenic obstruction of the musculoskeletal channels. Pharmacokinetic studies identify CYP3A4 as the predominant enzyme responsible for triptolide metabolism and show that reduced CYP3A4 activity can increase triptolide exposure and liver injury [83]. Thus, co-administration with hepatotoxic agents such as Methotrexate raises concern for additive liver toxicity, although a CYP3A4-mediated reduction in Methotrexate clearance has not been established.

4.3.2 IMKE step-by-step execution pipeline

Step 1 (Clinical Input): Input of Methotrexate prescription and VTM musculoskeletal symptoms.

Step 2 (NLP Processing): Extraction of [Western_Drug: Methotrexate] and [VTM_Pattern: Phong hàn thấp tý].

Step 3 (Graph Traversal & Ontology Mapping): The IKG conceptually maps the pattern to RA (ICD-10: M05) [9]. A graph query—with Cypher serving only as an illustrative syntax—could retrieve Quyên bí phương (Juan Bi Fang, 4), which may clinically include [Herb: Tripterygium wilfordii] (Thunder God Vine).

Step 4 (Relation-Aware Inference): A heterogeneous graph-learning method could flag a candidate toxicological risk based on shared metabolic pathways between the herb and the drug.

Step 5 (Explainability & CDSS Output): A candidate explanation method could highlight the following potential toxicological path: Tripterygium wilfordii [CONTAINS] Triptolide [METABOLIZED BY] Hepatic CYP3A4 [REDUCED CYP3A4 ACTIVITY] Increased Triptolide Exposure [ADDITIVE HEPATOTOXICITY WITH] Methotrexate [68, 83].

Illustrative System Output: In this conceptual scenario, the CDSS would present a high-priority hepatotoxicity concern, identify the supporting evidence path, and defer avoidance or substitution decisions to specialist review.

4.3.3 CDSS output

CLINICAL ALERT (High Priority): Potential Additive Hepatotoxicity. Mechanism: CYP3A4 is a major metabolic pathway for triptolide; reduced CYP3A4 activity may increase triptolide exposure and compound the hepatic burden of Methotrexate. This pathway does not establish CYP3A4-mediated inhibition of Methotrexate clearance and therefore requires pharmacologist review. Recommendation: Avoid co-administration unless supported by specialist assessment and close hepatic monitoring. Substitute: Juan Bi Fang reformulated without Tripterygium wilfordii; consider Du Huo Ji Sheng Wan as an alternative Wind-Damp formula with a safer interaction profile.

4.4 Architectural synthesis across scenarios

These three simulated executions collectively illustrate four hypothesized properties of the IMKE architecture. First, Generalizability: the functional pipeline could be instantiated across distinct domains without disease-specific redesign. Second, Mechanism Diversity: relation-aware reasoning may represent synergistic, additive, and toxicological pathways. Third, Bidirectional Value: the framework is intended to support both risk identification and investigation of potential therapeutic synergies. Finally, Explainability: candidate outputs should be accompanied by traceable computational paths that clinicians can review, without treating those paths as proof of causality [59].

5 Perspective and discussion: open challenges and future roadmap

The IMKE framework represents a paradigm shift in the clinical operationalization of integrative medicine. By providing a structured, computable bridge between VTM and Western pharmacology, it transforms integrative practice from an informal art dependent on individual practitioner memory into a systematically supported, safety-conscious clinical discipline. However, transitioning this architectural blueprint into a globally deployable clinical reality requires navigating profound technical, linguistic, and regulatory barriers.

5.1 Clinical and institutional implications

Critically, the IMKE is not designed in a regulatory vacuum. It is explicitly grounded in Vietnam’s Ministry of Health official integrative medicine guidelines (2020, 2026). This alignment provides regulatory legitimacy, ensuring that CDSS recommendations carry the institutional authority necessary to overcome clinician alert fatigue. Furthermore, the MoH guidelines explicitly mandate the combination of Western diagnostic methods with traditional treatments. The IMKE operationalizes this mandate by formalizing cross-paradigm safety: the GNN-based HDI prediction module closes the shadow pharmacovigilance gap, converting a known clinical blind spot—where over 42% of patients concurrently use undisclosed traditional medicines—into a manageable, continuously monitored risk.

Beyond safety, the IMKE acts as an engine for cultural preservation. By digitizing unstructured classical texts and encoding them into a heterogeneous multigraph, it ensures the “art” of traditional healing is preserved in a durable, computable format while maintaining its ontological integrity—representing VTM patterns in their native conceptual vocabulary without arbitrarily collapsing them into Western disease categories.

5.2 Perspective: overcoming open challenges

To realize the full potential of the IMKE architecture, the research and medical communities must collaboratively address three formidable barriers.

5.2.1 The data scarcity and cold-start bottleneck

Unlike Traditional Chinese Medicine (TCM), which benefits from large digitized resources (e.g., TCMID, HERB), VTM suffers from severe labeled-data scarcity. Any data-driven relation-learning approach for HDI prediction would require sufficiently representative positive and negative interaction evidence. Directly applying high-capacity models to sparse VTM data therefore risks a cold-start problem and unreliable representations.

Potential methodological pathways include cross-lingual transfer learning [69], ontology alignment where pharmacological overlap exists, and human-in-the-loop active learning [70]. Future studies could compare transferred representations, weak supervision, retrieval-augmented methods, and expert-curated baselines. Low-confidence predictions could be routed to clinical pharmacologists and VTM practitioners, but the value and safety of iterative model updating must be established empirically.

5.2.2 Semantic ambiguity in classical texts

VTM texts present extreme linguistic complexity. A critical barrier for the NLP pipeline is semantic ambiguity, where a single classical term holds vastly different meanings depending on its context. For instance, the concept of “Khí” (Qi – 汽) can refer to literal air, vital physiological energy, a specific organ’s functional state (e.g., Spleen Qi), or even external pathogenic factors (e.g., Cold Qi).

Static word embeddings, such as Word2Vec [71] or GloVe [72], cannot represent context-dependent meanings with a single fixed vector. Deep contextualized representations [73] therefore provide a plausible methodological direction, potentially supplemented by knowledge-aware encoding [53]. A future implementation could couple a Vietnamese contextual encoder (with PhoBERT as one candidate) to graph-derived context, but whether this improves disambiguation over multilingual, retrieval-based, or rule-assisted alternatives remains an empirical question.

5.2.3 Regulatory frameworks and liability imperatives

Perhaps the most daunting challenge is not computational, but legal: If the IMKE predicts a drug combination is safe, and a patient subsequently suffers fatal hepatotoxicity, who bears the liability? The software developer, the hospital, or the prescribing physician? As Lipton notes regarding the “mythos of model interpretability” [74], and Holzinger et al. argue, the boundary between “decision support” and “clinical authority” becomes dangerously blurred when clinicians lack the expertise to critically evaluate AI outputs [59].

Any future IMKE implementation would need to be evaluated against applicable medical-device and AI governance frameworks [75, 76]. The framework is conceived as clinician-facing decision support rather than an autonomous prescriber, with final clinical authority remaining with qualified practitioners. This governance requirement motivates a method-agnostic explanation layer: attribution, path tracing, counterfactuals, rule-based rationale, and inherently interpretable approaches should be compared for their ability to support independent clinical review.

5.3 Ethical governance and cultural inclusion

The deployment of AI in integrative healthcare introduces profound ethical responsibilities. The medicinal practices of Vietnam’s 54 recognized ethnic minority groups (e.g., the Tay, Nung, H’Mong) are severely underrepresented in existing digitized academic archives. Without deliberate corpus curation strategies targeting these communities, the IMKE’s IKG risks algorithmically marginalizing minority therapeutic traditions, inadvertently entrenching existing health inequities. Mitigation requires targeted digitization partnerships with regional hospitals and community health workers. Furthermore, full clinical utility depends on integrating sensitive EHR and VTM consultation data. Strict adherence to data governance obligations, including localized federated learning techniques to preserve patient privacy [77], is essential. Furthermore, aligning the IMKE with Vietnam’s ongoing digital health transformation [78] requires parallel investments in digital literacy to ensure the health workforce is fully equipped to utilize AI-augmented systems [79].

5.4 Strategic roadmap to clinical reality

Transitioning the IMKE from a conceptual architectural blueprint to clinical evaluation can be organized as a non-prescriptive three-phase research pathway (illustrated in Figure 7).

Phase 1—Feasibility and Knowledge Curation: Define representative clinical use cases, develop ontology and annotation guidance, and compare alternative language-processing, graph-storage, and reasoning approaches on a pilot corpus.

Phase 2—Retrospective Evaluation: Assess data quality, mapping reliability, interaction-ranking performance, explanation usefulness, and workflow feasibility using appropriately governed retrospective records and expert review.

Phase 3—Prospective Clinical Assessment: If feasibility and retrospective evidence are adequate, evaluate safety, clinical utility, human factors, and regulatory requirements in prospectively designed studies before considering adaptation to TCM, Ayurveda [80], Kampo [81], or other traditional medicine systems.

Table 6 outlines conceptual evaluation dimensions that future protocols could operationalize through prospectively defined metrics and acceptance criteria.

Table 6. Conceptual evaluation dimensions for future empirical studies of IMKE.

This table specifies evaluation dimensions rather than achieved results, fixed numerical targets, or a finalized protocol. Metrics, comparators, thresholds, datasets, and candidate technologies must be prospectively justified in future empirical studies.

6 Conclusion

This paper has presented a comprehensive review of Artificial Intelligence in traditional medicine and introduced the Integrative Medicine Knowledge Engine (IMKE) as a novel architectural perspective. By formally addressing the three foundational barriers to safe clinical integration—the Knowledge Gap, the Evidence Gap, and the Safety Gap—the IMKE framework provides a technically rigorous, mathematically grounded, and culturally sensitive blueprint for 21st-century integrative healthcare.

The proposed three-stage functional pathway comprises context-aware clinical-language processing, a heterogeneous Integrative Knowledge Graph (IKG), and relation-aware safety reasoning delivered through an explainable clinical interface. PhoBERT, Neo4j, R-GCN, SHAP, and related technologies are examples of possible instantiations rather than required components. Through simulated scenarios spanning three chronic disease domains, this paper illustrated hypotheses about how the architecture could organize cross-paradigm herb–drug interaction evidence and potential therapeutic synergies; it did not evaluate a trained model or establish clinical performance.

Beyond its immediate technical contributions, the IMKE architecture establishes a computational model for cultural preservation. By algorithmically encoding VTM’s rich therapeutic heritage into a durable, multi-relational knowledge structure, the framework ensures that centuries of phenomenological medical wisdom are preserved with ontological integrity and systematically evolved into a safer, evidence-informed clinical tradition. Furthermore, its agnostic, modular design provides a globally scalable template that can be readily adapted to harmonize other traditional medical systems—from TCM and Ayurveda [80] to Kampo [81] and Unani medicine—with modern Western biomedicine.

Transitioning this conceptual blueprint into clinical reality will require sustained, interdisciplinary collaboration spanning computer science, clinical pharmacology, bioethics, and healthcare policy. As detailed in our perspective, navigating the formidable bottlenecks of VTM data scarcity, linguistic ambiguity, and stringent regulatory liabilities will define the next phase of empirical research. Ultimately, as Artificial Intelligence continues to disrupt global healthcare, the IMKE framework demonstrates that the most profound technological opportunities lie not in replacing human clinical intuition, but in building the interoperable computational infrastructure required to connect, validate, and amplify the full breadth of humanity’s healing knowledge.

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI (e.g., ChatGPT, Gemini) was used to polish the academic language, improve the readability of the text, and assist with formatting the references and manuscript structure. The authors solely conceptualized the research, designed the IMKE framework, provided the clinical scenarios, and thoroughly reviewed all AI-edited content. The authors take full responsibility for the accuracy, originality, and integrity of the final manuscript

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fdgth.2026.1928965/full#supplementary-material.

Supplementary material Supplementary material at the source

Footnotes

Edited by: Shang-Ming Zhou, University of Plymouth, United Kingdom

Reviewed by: Xuxu Wei, Beijing University of Chinese Medicine, China

Avvinish Annant Narine, National Institute of Ayurveda, India

Tena Djuartina, Atma Jaya Catholic University of Indonesia, Indonesia

Footnotes

¹https://www.yhct.vn/.