Review of Health Knowledge Graphs for Clinical Decision Support: Construction, LLM-Grounded Reasoning, and Applications

1. Introduction

In numerous healthcare applications that contain multi-modal data, graph-based structures are suitable to represent the information in a structured manner. The data structure capturing the relationship between entities and attributes is termed as Knowledge Graph (KG) [1,2]. The information necessary to construct the KG is collected from websites, domain experts, and knowledge bases that are stored as unidirectional (one-way) or bidirectional (two-way) relationships, depending on the directionality of the association between the connected entities [1]. In recent years, KG has been applied to the multidisciplinary domain from the Google search engine to health care [3]. The application of KG models has been studied in a wide range of applications from basic science research [4,5,6], pharmaceutical research [7,8], clinical decision support [9,10], question answering and recommendation systems [11,12,13,14,15]. Post-digitization of healthcare systems, the longitudinal and comprehensive data collected during the delivery of patient care is created and encoded as Electronic Health Record (EHR) data. The National Academy of Medicine defines the core functionality of EHR as health information and data, decision support, electronic communication and connectivity, patient support administrative processes and reporting, and population health management [16]. Meaningful use of EHR has early evidence in better healthcare safety and quality benefits [15,17].

Health-KGs are constructed from intricate medical systems such as single or multiple integrated EHR datasets, medical literature, clinical guidelines, and patient-generated information where the Natural Language Processing (NLP) techniques are used to optimize entity and relation extraction methods [18,19]. The open medical EHR databases such as ‘Medical Information Mart for Intensive Care’ III (MIMIC III-IV) [20,21], and medical terminology for diagnoses, symptoms and procedures for claims processing such as Unified Medical Language System (UMLS) [22], Systematized Nomenclature Of Medicine (SNOMED-CT) [23], ICD-10 [24], and drug database such as DrugBank [25], PharmGKB [26] and KEGG [27] are also useful for Health-KGs construction.

Recent advancements in Machine and Deep Learning (DL) have demonstrated potential for resource allocation via task prediction, patient diagnosis [28,29,30,31], length-of-stay [32], and mortality estimation [33,34,35]. Health-KGs have been widely adopted to improve the data insights and complement EHR modelling [36,37,38]. Pre-trained large language models (LLM) such as BERT [39] and GPT [40] have been implemented for knowledge extraction in Health-KGs construction. The utilization of pre-trained LLM has allowed more comprehensive and accurate medical ontologies and taxonomies, further enabling efficient and effective integration of heterogeneous medical data from multiple resources [41,42,43]. However, there are significant challenges in Health-KGs construction due to their diverse and scattered nature, making integration and effective analysis quite complex [44]. This heterogeneity can also result in incomplete or conflicting data representations within Health-KGs, thereby limiting their utility for various healthcare applications [45,46,47]. Furthermore, the prevailing utilization of domain-specific knowledge graphs may lead to constrained coverage and granularity of knowledge encompassing various medical domains. This limitation hampers the ability to identify correlations and connections among medical concepts spanning multiple areas, thereby impeding comprehensive healthcare insights. This paper presents the results of our comprehensive research analysis on constructing Health-KGs from EHR datasets, the present state-of-the-art from most recent and promising research direction in Health-KGs leveraging LLM to personalized KGs, challenges, and its applications. An extensive literature analysis concerning Health-KGs was carried out at three levels: representation, extraction, and population or integration. The comprehensive summary techniques involved in constructing Health-KGs from scratch and integration are explored. Additionally, the model-based and model-free Health-KGs are also highlighted. Furthermore, an in-depth overview of the diverse applications of Health-KGs in drug development, clinical trial diagnosis, predictions, and treatment recommendations are also explored. It was observed that there is a noticeable gap in research within three critical areas: (i) managing complex real-world patient data in Health-KGs, (ii) leveraging semantic features for knowledge alignment, and (iii) addressing the dynamic nature of healthcare entities and relationships. This study on Health-KGs for healthcare advancement and meaningful application is a practical roadmap and resource for the healthcare community. Our primary research contributions are as follows:

• An extensive study of existing literature is explored to understand the present State-of-the-Art and identify the gaps and challenges in Health-KGs.
• The various methodologies, algorithms, and techniques used in Health-KGs construction, representation learning, reasoning, benchmarking, evaluation, and applications are explored in detail.
• A comprehensive taxonomy of the Health-KGs and its various applications is constructed in Figure 2.
• A dedicated synthesis of LLM-driven Health-KG construction and KG-grounded clinical reasoning (entity/relation extraction, retrieval- and graph-augmented generation, and factuality), an axis not addressed by prior EHR-focused reviews.

The results of this study will benefit a wide range of audiences, from data scientists, medical informatics, artificial intelligence researchers, and healthcare practitioners. The rest of the paper is structured as follows: Section II describes the search methodology; Section III situates this reviews against prior reviews; Section IV covers EHR data; Section V details Health-KG construction; Section VI examines the role of large language models; Section VII reviews applications; Section VIII discusses evaluation, open challenges, and future directions; and Section IX concludes.

2. Search Methodology

We conducted the review following the PRISMA 2020 reporting guidelines [48]. We searched PubMed, Scopus, IEEE Xplore, and the ACM Digital Library for literature published between January 2013 and December 2025, using combinations of the terms "electronic health record," "EHR," "knowledge graph," "medical/clinical knowledge graph," and the task qualifiers "construction," "representation," "extraction," "integration/population," and "large language model." The exact query strings per database are given in Table 1. We also hand-searched the proceedings of ACM SIGKDD, IEEE EMBC/EMBS, IJCAI, and AAAI and the archives of high-impact venues (IEEE Access, Knowledge-Based Systems, Expert Systems with Applications, Journal of Web Semantics, Journal of Biomedical Informatics).

Inclusion criteria: peer-reviewed studies that (i) construct, represent, integrate, or reason over a knowledge graph using EHR or closely related clinical data, or (ii) propose methods or benchmarks directly applicable to Health-KG construction or application. Exclusion criteria: non-English articles, non-clinical knowledge graphs, opinion pieces without methodological contribution, and duplicates across databases. After de-duplication, records were screened by title and abstract, then assessed in full text against the criteria above. Figure 1 reports the counts at each stage of the selection process.

3. Related Reviews

3.1. Related Reviews and Positioning

Knowledge graphs in the biomedical domain have been reviewed from several angles. Nicholson and Greene [1] surveyed the construction of biomedical knowledge graphs and their applications broadly across molecular biology. Ji et al. [2] provided a general (non-clinical) survey of knowledge-graph representation, acquisition, and applications. Abu-Salih [3] reviewed domain-specific knowledge graphs across multiple fields without an EHR focus, and Silva et al. [49] surveyed ontologies and knowledge graphs specifically for oncology research. More recently, Pan et al. [50] proposed a roadmap unifying large language models and knowledge graphs at a general (non-medical) level. None of these jointly addresses EHR-centric construction along the representation–extraction–integration axes, the recent shift toward LLM-based construction and reasoning, and a unified clinical application taxonomy. Table 2 summarizes this positioning.

4. EHR Data: Sources, Standard, Characteristics, and Complexity

The patient’s symptoms, test results, diagnosis, treatment, medication, and follow-up visits are stored in the hospital system as EHR data. The EHR data are comprised of several controlled vocabularies and classification schemas, such as procedure codes (Current Procedural Terminology, statistical Classification of Diseases and Related Health Problems) [51,52], and laboratory observations (such as the Logical Observation Identifiers Names and Codes) [53] to record the patient health-related information. EHR data can be structured in the form of lab reports or unstructured in the form of progress notes. This section discusses the sources of publicly available EHR data, the common standard of EHR management, and the challenges of creating Health-KGs from EHR data.

4.1. EHR Data Sources

The primary source of EHR data is stored within hospitals. The most significant challenge is accessing the EHR data as all organizations generally have strict human research organizational policies due to the presence of Protected Health Information (PHI). However, there are some large-scale publicly available EHR datasets for research purposes. Table 3 summarizes the principal publicly available EHR datasets, their approximate scale, modalities, and access terms. MIMIC-III [20] is a comprehensive dataset of EHR for patients admitted to tertiary care hospitals in the US in critical care units. The database comprises billing (CPT, Diagnosis-Related Group (DRG) codes, and International Classification of Diseases (ICD) codes), demographic details, dictionary, interventions, laboratory, medications, notes, physiologic, and reports (vital sign assessment). MIMIC-IV [21] database constructed from EHR records from Beth Israel Deaconess Medical Center (BIDMC) and Massachusetts Institute of Technology (MIT) comprising patient diagnoses, free-text clinical notes, orders, measurements, treatments, and procedures. The national NLP clinical challenges (n2c2) [54] are also available, subject to the data user agreement. The field’s heavy reliance on the MIMIC family means most reported Health-KG results are ICU-centric and single-region; broader external validation on multi-center resources (eICU-CRD) and non-ICU populations remains scarce, which limits claims of generalizability.

4.2. Healthcare Ontologies and EHR Standard

The EHR standards such as Health Level Seven (HL7) [55], openEHR [56], and CEN/ISO EN13606 [57] cover the multiple aspects of clinical records management. These standards cover integrating medical devices, exchanging patient data between healthcare institutions, and interfaces with clinical decision support systems. HL7 supports health service evaluation, delivery, and clinical practice and management. The structure and semantics of the EHRs are standardized using the Clinical Document Architecture (CDA) document. The HL7 RIM [58], HL7 DT is used to define the machine derivable CDA and using the SNOMED CT [23] it facilitates clinical documents semantic representation and interoperable with other standards. The OWL/RDF automatic rule-based inferencing is explored for HL7 RIM and HL7 DT representation. Furthermore, with the advancement of Web technologies, HL7 has also been extended to Fast Healthcare Interoperability Resources (FHIR) [59].

The primary aim of the openEHR and CEN/ISO EN13606 is to standardize the management, storage, retrieval, and exchange of health data in EHRs. The European Committee for Standardization (CEN) stated the norm for exchanging data between EHR systems known as Electronic Health Record Communication (EN13606). However, a subset of openEHR can be viewed as a stand-alone standard. In the past, the reference model and data types and structures of the openEHR OWL ontology following the official openEHR were developed [60]. Further, an ontology was formed recently to provide semantic interoperability between EN13606 and openEHR standards [61].

4.3. Multimodal EHR Data

The patient health records include demographics, laboratory test results, vital symptoms, histopathological test results, radiological test reports with different imaging modalities, and current and follow-up diagnoses. In a recent approach to model development, researchers have considered including most of the modalities despite the complexity of the nature of the data. Integrating multi-modal data in Health-KGs construction generally involves two primary phases;

• Using representation learning to properly represent multi-modal data from structured knowledge and language from many sources to overcome structural disparities [65,66]
• Depending on the data representation structure, adapting multi-modal learning for improved reasoning, classification, and prediction [67].

Recently, DL-based methods have been applied to tabular domain data, where a multi-modal transformer framework based on self- and semi-supervised learning frameworks was proposed to explore value imputation and data augmentation was studied using real-world MIMIC-III data [68]. Given the complexity of raw real-world data, the underlying explicit feature interactions from EHR data were modelled using Deep Cross Network (DCN) [69] and the novel pipeline to derive usable hourly time-series data from the raw MIMIC-III data [70]. In a recent article, ‘FIDDLE’, a systematic pre-processing technique for processing EHR data, was proposed [71], utilizing for medication [72] and diagnosis predictions [73]. The existing diagnosis predictions using a learning-based algorithm were improved by leveraging the multi-modal nature of the EHR data [74,75]. A multi-modal attention neural network trained to combine MIMIC-III clinical notes and codes was proposed [76].

4.4. Heterogeneous Temporal Data in EHR

The heterogeneous medical data such as patient ID, progress and current diagnosis, lab reports, imaging results, and medications are stored in EHR [77]. More importantly, constructing and utilizing Health-KGs to predict clinical decision support tasks incorporating clinical data collected over time with cross-departmental information provided enhanced, in-depth, superior longitudinal analyses and patient trajectories of clinical events. However, the modelling and integration of heterogeneous data is challenging due to the complex nature of the data. Besides, temporal KGs are proposed by integrating time stamps in the model, which better captures the underlying complex temporal dynamics of the entities and relationships in EHR data [78].

4.5. Semantic Complexity of EHR

Improved contextual knowledge utilization for improved interoperability, reduced duplication, and missing entity identification are made possible by the semantic knowledge representation of the combined and linked entities. The main challenges faced by the researcher in constructing Health-KGs is the complex semantic patterns, which frequently manifest as terms and medical disorders in clinical records. In a recent work, the conceptual graphs were integrated with health data. To separate the contextual information from the complicated semantic textual data, they employed inference mechanisms and textual pattern mining for knowledge retrieval [79]. The absence of prior subject expertise was the reason for the unimpressive outcomes. Removing repetitive semantic structures from EHR was proposed using syntactic trees from sentences to search similar sub-trees and labelled sentences by employing word embedding algorithms [80]. Learning medical knowledge graphs from EHR data is more complicated than learning from structured sources such as textbooks and articles. However, there are a lot of associated relationships that a KG created from real patient information infers. The benefit of creating a Health-KG from an EHR is the minimum post-processing requirements.

5. Health-KGs Construction: Representation, Extraction, and Integration

To understand the significant features of EHR using Health-KGs, we analyzed the existing approaches used to construct Health-KGs. Figure 2 illustrates the comprehensive taxonomy for Health-KGs in four primary categories: representation, extraction, integration, and applications. The primary step of Health-KGs construction is identifying data from sources such as open databases, scientific publications, and real-world EHR [81]. The data, once identified, are primarily available in the form of structured data (such as demographic information, vital signs, laboratory test results, and diagnostic codes) and unstructured data (such as clinical notes, medical images, and radiology reports) [82]. The subsequent sections below explain more comprehensive details on the construction of Health-KGs.

5.1. EHR Representation and Learning

The heterogeneity, temporality, and knowledge intensity are the primary challenges in representing EHR data [83]. The most common approach to representing data is in the form of two entries and a relation connecting these entries, i.e., triplet format. The triplets are extended to a quadruple format, which includes additional disease or patient-related properties [84]. The representative sample of triplet and quadruple format is shown in Figure 3. The main focus on existing Knowledge Representation Learning (KRL) is to learn the semantic relations, relational triplets (represented as head entity, relation, tail entity), and attribute triples (delineated as entity, attribute, attribute value) [85].

The existing KRL approaches are divided into two categories. The first involves modelling the structural information among entities, while the other involves using texts or semantic information retrieved from triples in KGs. The proposed Trans series models are computationally easier to train, representing the anterior approach yet work in one-to-one relations [86]. In the TransH model, the head and tail nodes are projected onto the same vector in the relation plane, which addresses the 1-to-n and n-to-1 problem between entities and relations [87]. The TransE and TransH models project the head and tails in the same vector space, which was addressed using the TransR model [88]. The other improved models in the Trans series models included PTransE [88], TransD [88], and TransSparse [88]. The different text semantic pertinences between different entities can be modelled using the neural tensor network model [89]. The targeted text is extracted via embedding of target entities through multiple convolutional filters using end-to-end fully connected convolutional neural network (CNN) proposed in the ConMask model [90]. The further optimization in the form of the conditional probability of the objective function proposed in KR-EAR methods demonstrated superior performance in entity, relation, and attribute prediction tasks, respectively [91]. The triplet prediction and attribute value regression were modelled using a multi-task neural network approach [92], while the Description-Embodied Knowledge Representation Learning (DKRL) performs the representation learning of similar task via a weighted combination of the TransE model and hybrid model, which was a combination of Continuous Bag of Words (CBOW) model and deep CNN [93]. The comprehensive structural and semantic relational and attribute triples information was utilized using KG via Attribute-embodied Knowledge Representation Learning (AKRL) [84].

5.2. Ontological Representation

The sharing of the EHR across different healthcare providers securely and efficiently is an ongoing challenge. Integrating medical ontology models for constructing Health-KGs from heterogeneous EHR data across various sources assists in achieving interoperability and semantic visualization, which is often beneficial for clinical research.

Several works have been proposed for creating medical ontologies. The properties of tumorigenesis and cancer cell description are determined in OncoCL ontology [94], which has been built by using Cell Ontology (CL), UBERON, BRENDA Tissue Ontology (BTO), Pathway Ontology, Phenotype And Trait Ontology (PATO), Coding Process Ontology (CPO), and Sequence Ontology (SO). The drug target prediction is determined from the CRC ontology [95] by reusing the PharmGKB ontology. The POCS ontology [96] was developed, reusing the Friend Of A Friend (FOAF) ontology for cancer survivor engagement. The HemOnc ontology was constructed to systematize hierarchic cancer treatment [97]. Besides, the survival outcome measures of clinical trials in oncology were proposed in TOCSOC ontology, which was constructed using Experimental factor ontology (EFO), CCTOO, Interlinking Ontology for Biological Concepts (IOBC), and National Cancer Institute Thesaurus (NCIT) [98]. Newer terminology-focused and Semantic-focused application ontologies are available in [49].

5.3. Graph Embedding

Graph Embedding (GE) converts the KG into a low-dimensional vector to represent the data. GE preserves the semantic information by computing the entity similarity in the form of distributed embeddings as Knowledge Graph Embedding (KGE) [99,100]. The primary aim of KGE is to represent the non-semantically connected entities and relations in similar vectors. In general, the neural network is trained over a KG to learn the KG via three primary steps:

• Encoding entities into dispersed points within the semantic space, with relations encoded as vectors
• Employing a scoring function or model-specific function to amalgamate information derived from a triplet
• Implementing an optimization procedure guided by the loss function. This optimization aims to define and minimize the objective during the knowledge graph embedding, ensuring effective representation and learning from the encoded entities and relationships [101].

The categorization of embedding techniques typically encompasses translational distance (TD), semantic matching (SM), neural network-based models (NNM). TD utilize distance-based score functions, while SM employ similarity-based score functions. These mechanisms calculate the distances or similarities between entities and relationships, forming the basis for generating embeddings in knowledge representation [100,102]. Table 4 compares the principal KGE models by category, relation patterns handled, relative complexity, and key limitations.

In a recent work, the projection of a single entity into several projection vectors in various relations was made possible by a relation-dependent hyperplane constructed via mapping the subject and object embedding vectors [103] while the other work, in contrast to hyperplane, it was focused on relation-specific space [104]. The probabilistic model learned from Health-KGs was proposed to predict the missing link [105]. Assessing the formed triplets’ validity involves using semantic or distance-based scoring functions. Subsequently, to enhance the model knowledge, the representations of facts are encoded with a suitable neural model, incorporating additional information such as entity types and attributes descriptions, & relational paths [2]. Semantic matching models that capture the inherent semantics of triples include RotatE [106], DistMult [107], and RESCAL [108], ComplEx [109]. The ConKB and ConvE [110,111] utilize convolutional neural networks (CNN) based architecture to feature analysis of triples. The extraction of semantically relevant terms using word embeddings trained on EHR and medical literature has been prevalent. The Multiple BERT version SciBERT [112,113], ClinicalBERT [114], and BioBERT [115] are proposed in the clinical domain. Integrating KGE into clinical KGs representation has demonstrated significant success in the clinical pharmacy domain in applications like DL-based drug discovery [116,117] and prediction of drug-drug interactions [118]. Node embedding plays a crucial role in testing various link prediction methods within biomedical KG [119]. For instance, the node2vec [120] algorithm generates node embeddings from a KG centred around drug reactions [121]. Case studies conducted on COVID-19 and liver injury drugs showcased the algorithm’s effectiveness [122]. Furthermore, utilizing graph-based representation to learn the concept embeddings from SNOMED-CT in healthcare analytics yielded promising link and patient state prediction [123]. The joint embedding-based representation model, which incorporates descriptions of semantic medical entities and triples, was developed to measure patient similarity. The joint embedding-based representation model was developed using the Siamese CNN model utilized MIMIC-III dataset [20], yielding superior results in patient clustering [124]. Using GE to construct specific disease knowledge graphs in multi-modal reasoning has shown promising outcomes. However, KGE is in the early development stage to provide universal and new reliable knowledge in the disease-specific field [4]. In Health-KGs the choice of embedding is governed less by raw link-prediction accuracy than by the relational structure of clinical data, where many-to-many and antisymmetric relations (e.g., drug–interacts-with–drug, disease–progresses-to–disease) dominate; this favors RotatE and ComplEx over the original TransE, while CNN-based models such as ConvE trade interpretability for expressiveness.

5.4. Health-KGs Integration

The KG integration is crucial in constructing Health-KGs to avoid repetitive work and construction from scratch. Integrating or merging multiple Health-KGs are effectuated into single Health-KGs [125,126,127]. The dissimilar data format, terminologies, and schemas make the process of Health-KGs integration more challenging. In recent work, various methods have been proposed to address this challenge. The fusion algorithms like schema alignment [128,129], entity and conflict resolution [130,131,132], and ontology matching [133,134] provides robust integration and reconciliation to integrate different Health-KGs. In earlier proposed work, Health-KGs were integrated with DL-based methods to diagnose thyroid disease [135]. The integration of context retrieved from the large corpus of cancer pathology reports using KGE and external medical terminology ontologies was used to train a multi-task CNN for cancer phenotyping [136]. The partially overlapping multiple EHR institutional codes were integrated with the textual data using the multiview incomplete Health-KGs integration algorithm. The results demonstrated superior cross-institutional data harmonization and improved semantic accuracy related to COVID-19 EHR studies [137].

5.5. Extraction

In the context of knowledge extraction, Health-KGs construction involves the extraction of entities and the relationships between them. The extraction involves three main tasks [138]: i) Named Entity Recognition, which involves recognition of organizations, events, individuals, location and other entities from (un)structured or semi-structured data; ii) Named Entity Disambiguation, aimed at resolving the ambiguity of an inferred entity by mapping it to its real-world counterpart; and iii) Named Entity Linking, assigning a unique Internationalized Resource Identifier to the disambiguated entity.

5.5.1. Named Entity Recognition

Named Entity Recognition (NER) is a popular research domain in the scientific community. Most approaches have been domain-specific and limited to reviews, news, and health care. We categorize the literature into three main categories.

• The rule-based KG approaches, in this category, most studies have implemented hand-crafted rules-based approach [138,139]. This approach does not require Large-scale annotated data since they mostly rely on lexical resources. Additionally, due to the nature and domain-specific approach, very high precision is often achieved [140]. However, the most common disadvantages are the unavailability of lexicon resources, domain dependence, and the cost of expert resources for many languages.
• The second approach is learning-based approaches, which are categorized into supervised, semi-supervised, and unsupervised. Machine learning (ML) models like Support Vector Machines, Conditional Random Fields, Hidden Markov Models, and decision trees are trained on input samples and used as classifiers [141]. More end-to-end models via bootstrapped and unsupervised methods are also proposed, requiring minimal training or seed ontologies [142,143,144].
• The feature-inferring neural-network approaches is the most recent and last category in this domain [39,145,146,147,148,149]. In contrast to the rule- and learning-based methods, these approaches leverage ML but distinguish themselves by automatically deducing features through DL. Recent studies suggest that they surpass earlier methods [141,142,150]. Unlike the aforementioned approaches, these methods do not necessitate seeds, ontologies, or domain-specific lexicons, rendering them more domain-independent. Additionally, they capitalize on the precision of their inferred features. However, constructing the large-scale dataset is required to develop robust models using these approaches.

5.5.2. Named Entity Disambiguation

In conventional Named Entity Disambiguation (NED) research, the computation of similarity between a mention and its candidate entities often relies on manually crafted features [151,152]. These investigations can be categorized into independent (or individual), collective, and Neural Network based approaches.

• Independent methodologies employ semantic similarity techniques to arrange candidate entities based solely on their lexical similarity and/or observed co-occurrence with mentions [153,154]. Each mention undergoes autonomous disambiguation in these approaches, framing entity disambiguation as a ranking challenge that selects the entity with the highest confidence. Confidence levels are determined by amalgamating handcrafted features drawn from the mention’s context, including surrounding words and textual portrayals of the candidate entity, often extracted from sources like Wikipedia. Various similarity metrics, such as BOW-based cosine similarity, have been suggested. Nevertheless, owing to the limited textual information inherent in handcrafted features, the performance decreases in intricate scenarios [152]. Moreover, a more complex model-based approach is required to accommodate better the intricate and interconnected nature inherent in natural language.
• Collective methodologies utilize semantic similarity considering that text generally revolves around the same subject [155]. Consequently, co-occurring entities are expected to frequently exhibit semantic relationships [156]. The associations between candidate entities are thus established using collective approaches, and the probabilistic graph method is utilized to build the mention-entity pairs. The Babelify [157], AGDISTIS [158], and TagME [159] leverage graph connectedness benefits from semantic relationships among disambiguation candidates outlined in a knowledge base. Most of them rely on BOW, which is computationally expensive.

Neural network-based models have been prevalent and have gained significant success. Recently, to reduce the time in NED recognition in Health-KGs applications, the Long short-term memory (LSTM)-conditional random field (CRF) model was utilized. The proposed model was a hybrid combination of rule-based and dictionary-based and achieved a significant recognition rate [160]. Besides, a similar hybrid model comprising verb and noun phrases for the relationship between classification was proposed [161]. The feature set obtained via the proposed model was evaluated using a ranging algorithm and classified using a Support Vector Machine (SVM) and Naive Bayes (NB) classifier. More recent research utilized a deep neural network (DNN) based model to learn the latent semantic features to improve the generalization capabilities and overall accuracy in NED. A DNN-based approach was implemented to extract the entities to construct Health-KGs, which used the inverse document frequency-based filtering [144] while a similar idea was proposed in another research, but a Co-training algorithm was used. The idea was to train two different classifiers independently and use independent target instance distribution [162].

5.5.3. Entity Linkage

Entity Linkage (NEL), also known as entity resolution or record linkage, is a process in Health-KGs that involves identifying and linking related information about entities, such as patients, medical conditions, or treatments, across different data sources or within the same dataset [163,164]. The goal is to connect information that refers to the same real-world entity, even if the data is represented differently or contains variations [165].

Numerous works have been published to extract clinical concepts from reports. In this work, we particularly focus on the techniques and models that focus on extracting the Unified Medical Language System (UMLS) through NER and NEL. The state-of-the-art models like GloVe [166], BERT [39], ELMo [146], word2vec [167], and fastText [168] are popularly used to identify UMLS concepts. In most published literature, the BERT-based contextualized embedding approach has outperformed other techniques. The other techniques BlueBERT [169], BioBERT [115], SapBERT [170], UmlsBERT [171], NeighBERT [172] have utilized UMLS and their synonyms for EL. Although there has been significant success in the past, most of the published work suffers from numerous entity-linking errors attributed to unstructured text snippets and the presence of multiple relationships associated with entities [173]. Pretraining corpus, not architecture, is the dominant driver of clinical performance: models exposed to in-domain clinical notes (ClinicalBERT, BlueBERT) lead extraction tasks, while UMLS-aligned models (SapBERT, UmlsBERT, NeighBERT) lead entity linking. Table 6 summarizes these clinical and biomedical language models by pretraining corpus and primary use in Health-KGs.

Table 5. Representative LLM-based systems for Health-KG construction and reasoning. Grouped by primary role of the knowledge graph. Rows marked “(general)” are domain-agnostic paradigms adapted to the clinical setting.

System [Ref]	Year	Primary task	Role of KG & grounding mechanism	Data / benchmark	Reported result / finding
Agrawal et al. [174]	2022	Extraction	Schema-in-prompt zero/few-shot clinical information extraction	Clinical notes (i2b2/CASI)	Few-shot LLM IE competitive without task-specific training
SAC-KG [175]	2024	Construction	LLM Generator + rule-based Verifier + fine-tuned Pruner	Domain corpora	Automated domain-KG construction with staged error control
KARMA [176]	2025	Construction / enrichment	Nine collaborative agents; schema alignment + conflict resolution	1,200 PubMed articles	∼38k new entities; 83.1% LLM-verified correctness
Yang et al.  [177]	2025	Construction (EHR)	GPT-4 entity recognition into Neo4j; de-identification pre-step	Multicenter sepsis EHR (∼10.5k patients)	Prompt-engineered multicenter clinical KG for decision support
Think-on-Graph [178]	2023	Reasoning (general)	Interleaved LLM reasoning with explicit KG traversal	General KGs	Interpretable multi-hop inference chains
GraphRAG [179]	2024	Reasoning / QA (general)	Corpus indexed as KG; community-level summaries	General corpora	Global, sense-making queries beyond context window
Gao et al. [180]	2023	Diagnosis prediction	UMLS 2-hop subgraph injected into LLM	MIMIC-III	KG grounding improves diagnosis prediction
RAM-EHR [181]	2024	Clinical prediction	Retrieval augmentation over multi-source EHR	EHR predictive tasks	Retrieval augmentation improves predictive accuracy
MedRAG [182]	2025	Diagnosis / reasoning	KG-elicited reasoning over a four-tier diagnostic KG	DDXPlus + private CPDD	Lower misdiagnosis; proactive follow-up questions
MedGraphRAG [183]	2025	QA / reasoning	Triple Graph Construction + U-Retrieval; evidence linking	Medical QA benchmarks	Evidence-based, source-grounded responses
AMG-RAG [184]	2025	Question answering	Agentic auto-construction/updating of MKG + external retrieval	MedQA, MedMCQA	MedQA F1 ≈74.1% with an 8B backbone
HyKGE [185]	2024	QA / reasoning	Hypothesis-KG-enhanced retrieval	Medical QA	Improved accuracy and answer reliability
TC-RAG [186]	2024	Reasoning	Turing-complete/stateful RAG with a memory stack	Medical LLM systems	Controlled retrieval state for deeper multi-hop
PrimeKG [187]	2023	Resource (KG)	Openly available precision-medicine KG substrate	Drugs/diseases/ phenotypes/genes	Reusable multimodal biomedical KG for grounding

Table 5. Representative LLM-based systems for Health-KG construction and reasoning. Grouped by primary role of the knowledge graph. Rows marked “(general)” are domain-agnostic paradigms adapted to the clinical setting.

System [Ref]	Year	Primary task	Role of KG & grounding mechanism	Data / benchmark	Reported result / finding
Agrawal et al. [174]	2022	Extraction	Schema-in-prompt zero/few-shot clinical information extraction	Clinical notes (i2b2/CASI)	Few-shot LLM IE competitive without task-specific training
SAC-KG [175]	2024	Construction	LLM Generator + rule-based Verifier + fine-tuned Pruner	Domain corpora	Automated domain-KG construction with staged error control
KARMA [176]	2025	Construction / enrichment	Nine collaborative agents; schema alignment + conflict resolution	1,200 PubMed articles	∼38k new entities; 83.1% LLM-verified correctness
Yang et al.  [177]	2025	Construction (EHR)	GPT-4 entity recognition into Neo4j; de-identification pre-step	Multicenter sepsis EHR (∼10.5k patients)	Prompt-engineered multicenter clinical KG for decision support
Think-on-Graph [178]	2023	Reasoning (general)	Interleaved LLM reasoning with explicit KG traversal	General KGs	Interpretable multi-hop inference chains
GraphRAG [179]	2024	Reasoning / QA (general)	Corpus indexed as KG; community-level summaries	General corpora	Global, sense-making queries beyond context window
Gao et al. [180]	2023	Diagnosis prediction	UMLS 2-hop subgraph injected into LLM	MIMIC-III	KG grounding improves diagnosis prediction
RAM-EHR [181]	2024	Clinical prediction	Retrieval augmentation over multi-source EHR	EHR predictive tasks	Retrieval augmentation improves predictive accuracy
MedRAG [182]	2025	Diagnosis / reasoning	KG-elicited reasoning over a four-tier diagnostic KG	DDXPlus + private CPDD	Lower misdiagnosis; proactive follow-up questions
MedGraphRAG [183]	2025	QA / reasoning	Triple Graph Construction + U-Retrieval; evidence linking	Medical QA benchmarks	Evidence-based, source-grounded responses
AMG-RAG [184]	2025	Question answering	Agentic auto-construction/updating of MKG + external retrieval	MedQA, MedMCQA	MedQA F1 ≈74.1% with an 8B backbone
HyKGE [185]	2024	QA / reasoning	Hypothesis-KG-enhanced retrieval	Medical QA	Improved accuracy and answer reliability
TC-RAG [186]	2024	Reasoning	Turing-complete/stateful RAG with a memory stack	Medical LLM systems	Controlled retrieval state for deeper multi-hop
PrimeKG [187]	2023	Resource (KG)	Openly available precision-medicine KG substrate	Drugs/diseases/ phenotypes/genes	Reusable multimodal biomedical KG for grounding

6. Large Language Models for Health-KG Construction and Reasoning

The construction pipeline in Section IV has historically depended on task-specific extractors, translational or neural embedding models, and rule- or machine-learning-based linkers, each requiring substantial annotated data and domain engineering. The emergence of instruction-tuned large language models (LLMs) has reshaped every stage of this pipeline. In contrast to the encoder-only models discussed earlier, such as BERT [39] and its clinical variants, generative LLMs perform extraction, normalization, and multi-hop reasoning with few or no task-specific labels. The resulting relationship characterized as bidirectional: Health-KGs supply LLMs with verifiable, structured clinical knowledge and provenance, while LLMs lower the cost of constructing, completing, and querying Health-KGs [50]. This section organizes recent work along three threads: LLM-driven extraction for construction (VI-A), retrieval- and graph-augmented clinical reasoning that couples Health-KGs with LLMs (VI-B), and the use of LLMs as clinical knowledge sources together with the factuality challenge that constrains their autonomous use (VI-C). Figure 4 presents the end-to-end Health-KG construction-and-application pipeline, in which LLM- and RAG-based components augment extraction, linking, and reasoning while completed knowledge feeds back for continual update. Table 5 summarizes representative LLM-based systems for Health-KG construction and reasoning, organized by the role the graph plays and the grounding mechanism employed.

6.1. LLM-Driven Entity and Relation Extraction

The named-entity recognition, disambiguation, and linking methods reviewed in Section V frame extraction as a sequence of supervised classification tasks. Generative LLMs reformulate extraction as a conditional generation problem, in which entities, relations, and their normalized identifiers are produced directly from clinical text under natural-language or schema-constrained prompts. This reformulation substantially reduces the annotation burden that earlier supervised pipelines required, and in-context exemplars allow rapid adaptation across note types and institutions without retraining. The output can be constrained to controlled vocabularies, mapping extracted mentions to UMLS [22], SNOMED-CT [23], or ICD codes [24] during decoding, which integrates the recognition and linking stages that were previously separate. These gains come with characteristic failure modes: LLMs may hallucinate non-existent entities, assign inconsistent semantic types, or produce identifiers absent from the target ontology. Consequently, LLM-based extraction for Health-KG construction is most reliable when paired with ontology grounding and human-in-the-loop validation, trading the label efficiency of generative extraction against the precision and controllability of the supervised neural models.

Table 6. Clinical/biomedical language models for Health-KG extraction and linking.

Model	Pretraining Corpus	Primary Use in Health-KGs
SciBERT [112]	Scientific (Semantic Scholar)	Scientific-text NER / classification
BioBERT [115]	PubMed abstracts + PMC full text	Biomedical NER, RE, QA
ClinicalBERT [114]	MIMIC-III clinical notes (over BERT/BioBERT)	Clinical NER, readmission
BlueBERT [115]	PubMed + MIMIC-III	Biomedical + clinical benchmarks (BLUE)
SapBERT [170]	UMLS synonym pairs (self-alignment)	Entity linking / concept normalization
UmlsBERT [171]	+ UMLS Metathesaurus relations	Clinical NER, NLI
NeighBERT [172]	Relation-induced dense retrieval	Medical entity linking

Table 6. Clinical/biomedical language models for Health-KG extraction and linking.

Model	Pretraining Corpus	Primary Use in Health-KGs
SciBERT [112]	Scientific (Semantic Scholar)	Scientific-text NER / classification
BioBERT [115]	PubMed abstracts + PMC full text	Biomedical NER, RE, QA
ClinicalBERT [114]	MIMIC-III clinical notes (over BERT/BioBERT)	Clinical NER, readmission
BlueBERT [115]	PubMed + MIMIC-III	Biomedical + clinical benchmarks (BLUE)
SapBERT [170]	UMLS synonym pairs (self-alignment)	Entity linking / concept normalization
UmlsBERT [171]	+ UMLS Metathesaurus relations	Clinical NER, NLI
NeighBERT [172]	Relation-induced dense retrieval	Medical entity linking

6.2. Retrieval and Graph-Augmented Clinical Reasoning

The parametric LLMs cannot reliably access or update factual knowledge, a complementary line of work grounds their outputs in external structured knowledge. Retrieval-augmented generation (RAG) [188] retrieves relevant passages from EHR corpora, clinical guidelines, or literature and conditions generation on them, providing provenance and allowing knowledge to be updated without retraining. Graph-structured variants extend this idea to Health-KGs. GraphRAG [179] indexes a corpus as a knowledge graph and uses community-level summaries to answer global, sense-making queries that exceed an LLM’s context window, a capability well suited to longitudinal, multi-document patient records. Think-on-Graph [178] interleaves LLM reasoning with explicit traversal over a knowledge graph, yielding interpretable multi-hop inference chains rather than opaque parametric recall. These paradigms have been instantiated in clinical settings: retrieval augmentation has been applied directly to predictive tasks over electronic health records [181], and a medical knowledge graph has been injected into an LLM to improve diagnosis prediction on MIMIC-III [180]. The consistent finding across these systems is that coupling an LLM with a Health-KG improves factuality and interpretability and supplies the long-tail and temporally evolving knowledge that parametric models omit. The open problems are retrieval quality, schema alignment between the LLM’s outputs and the target KG, and the stability of multi-hop traversal as reasoning depth increases.

6.3. LLMs as Clinical Knowledge Sources and the Factuality Challenge

The final thread treats the LLM itself as a partial knowledge source [189] showed that large models encode considerable clinical knowledge and that comprehension and reasoning scale with model size and instruction tuning, while also demonstrating that performance remains below that of clinicians and that automated benchmarks understate safety-critical errors. This gap defines the principal barrier to using LLMs as autonomous constructors or curators of Health-KGs: hallucination, in which fluent but unsupported facts are generated, is difficult to detect with standard benchmarks [190], and an unsupported triple introduced into a clinical KG can propagate into downstream decision support. Knowledge-graph grounding, retrieval with provenance, and explicit verification are therefore not optional refinements but prerequisites for safe deployment, a point reinforced by surveys of RAG mitigation strategies [191]. These factuality requirements also intersect with the compliance [192,193] and privacy concerns of Section VII-D, since clinical text placed in LLM prompts constitutes protected health information, motivating on-premises, federated, or de-identified construction workflows. In aggregate, LLMs shift the bottleneck of Health-KG construction from extraction labor toward verification, grounding, and governance, and the absence of standardized clinical KG–LLM benchmarks and factuality guarantees remains the central open challenge that Section VIII develops further.

7. Application of Health-KGs

Health-KGs, an EHR-curated knowledge graph, is a powerful tool for representing complex real-world medical knowledge to assist various clinical decision support applications. In this contribution, several Health-KGs applications, i.e., drug development, clinical trial, diagnosis and predictions, treatment recommendation, question answering, and health policy compliance, will be covered. Table 7 summarizes representative Health-KG applications together with their methods, datasets, and evaluation metrics.

7.1. Drug Development & Clinical Trials

Health KGs have been popularly implemented in drug development and drug repurposing. In a recent ML-based approach [195], integrated KGs were used to predict drug-drug interactions utilizing the features from KEGG [205], DrugBank [25], and PharmGKB [26]. The default parameters provided in the open-source implementations of PBG9, SimpleIE12, TransE11, and CrossE10 were utilized for knowledge graph embedding. The complex embedding method based on the Convolutional-LSTM (C-LSTM) network was reported as the optimal performance model. In another work involving a similar objective, an end-to-end framework, KG-neural network (KGNN), was proposed to capture high-order structures and semantic relations [206]. The other work explored the discovery of candidate drugs for diseases of interest by creating Health-KGs from the therapeutic target database (TT), PubMed database, and UMLS. The authors proposed the LSTM-based model GrEDeL and compared it using Random Forest and Support Vector Machine [116], while in another literature to resolve similar objective, a logistic regression model called SemaTyP was trained to understand the semantic types of paths associated with known drug therapies within the Health-KGs [194]. The acquired model is applied to unveil drug therapies for novel diseases. SemaTyP model was compared with the two Random walk algorithm (RWA)- based methods [207,208]. In another study to predict DDI, authors utilized Logistic Regression (LR), Naive Bayes (NB), and Random Forest (RF) in conjunction with TransD, TransE, and RDF2Vec on the DrugBank knowledge graph, employing a 10-fold traditional cross-validation approach. The model was evaluated RDF2Vec on diverse drug knowledge graphs, including DrugBank, PharmGKB, and KEGG, to predict previously unknown drug-drug interactions [118].

To enhance drug discovery efficiency, scientists investigate the repurposing of ‘existing’ drugs [202]. In a recent study, KG-Predict was proposed and constructed using a stack of CompGCN layers. The Mouse Genome Informatics (MGI) database [209], Gene Ontology Annotation (GOA) database [210], and Genotype-Tissue Expression (GTEx) [211] were used. This approach involves mitigating risks associated with advancing to clinical phases earlier. Evaluating Adverse Drug Reactions (ADRs) in pharmacological analysis and clinical application is crucial. Consequently, during the stages of safety pharmacology analysis and phase III clinical trials, researchers often allocate substantial resources to assess ADRs by identifying side effects and Drug-Drug Interactions (DDIs).

7.2. Diagnosis Prediction & Treatment Recommendation

Health-KGs with EHR data from different sources have been widely used in health care recommendation systems for diagnosis prediction & treatment or drug recommendation. Diagnosis Prediction early is a significant research area in medical informatics because it can anticipate patient prognosis and mitigate preventable adverse events. In a recently published study [197,199], Health-KGs was built using the PLAGH dataset, the MIMIC-III dataset [20], and the MIMIC-IV dataset [21]. The Health-KGs were derived from patients’ current diseases or risk factors and a disease entity representing the predicted ailment. The diagnosis prediction and disease progression were achieved via a policy-based reinforcement learning module (RLM) [199]. Besides, the other study constructed sequential patient graphs and utilized the Graph Neural Network-Based Diagnosis Prediction (GNDP) for diagnosis prediction [197]. In a similar goal of diagnosis prediction task, knowledge of the medical graph (KAME) [196] an end-to-end, accurate, and robust recurrent neural networks (RNNs) with attention mechanism was implemented and large Language Models (LLM) for Diagnosis Prediction [180] and gated recurrent unit (GRU) was proposed [198]. The performance of the KAME & LLM model in comparison to the existing state-of-the-art improved significantly [196].

A recent study used a graph-based KG-Convolutional Neural Network (KGCNN) coupled with sentiment analysis (SA) extracted from the public reviews on drugs to develop a drug recommendation system. The Health-KGs developed captured user-item relatedness by mining its associated attributes effectively [203]. In another work, a thyroid treatment recommendation framework was proposed, which utilized TransD embedding and provided the treatment recommendation using a joint neural model (JNN). The output is ranked using a classifier trained on internal EMR data collected from Shanghai Hospital [200].

7.3. Question Answering

Question answering within the medical domain is crucial as a decision support tool. An EMR-based knowledge graph specific to the domain ensures the provision of intelligent and contextually appropriate answers to patient-specific queries, surpassing the capabilities of a conventional schema-based question-answer system [77]. The significance of comprehending the semantics and context of data retrieval for information retrieval, especially when seeking domain-specific information, cannot be overstated. In recent studies, knowledge graph-based question answering (KGQA) was proposed to extract named entities and relations from data. The related entities obtained were ranked using weighted path ranking on the knowledge graph [212]. In another study, the Multi-Modal Knowledge-aware Hierarchical Attention Network (MKHAN) was developed to efficiently harness a multi-modal knowledge graph (MKG) for transparent medical question answering. MKHAN utilized the underlying sequential dependencies within a path extracted from MKG to deduce the underlying rationale of question-answer interactions [201].

7.4. Health Policy Compliance

The significance of data in scientific experiments within healthcare is invaluable, particularly given the exponential growth in data generation. In the healthcare sector, the Health Insurance Portability and Accountability Act (HIPAA) is a pivotal privacy regulation dedicated to protecting patient data in the United States [193,213]. Comparable privacy regulations exist in other domains, exemplified by the General Data Protection Regulation (GDPR), which outlines companies’ restrictions on using customer information [192,214]. To ensure data privacy, data providers institute security protocols that precisely outline criteria for obtaining, analyzing, and sharing personal and confidential information during the integration process. A recent work proposed an integrated KG-based framework for protecting privacy data across healthcare providers. The framework prevented excess or unwanted access using Attribute Based Access Control (ABAC) [204]. In another study, the Trusted Compliance Enforcement Framework (TCEF) was developed using semantic web technologies where a trust Score is assigned to each participant involved in the data exchange process, incorporating ontologies alongside policy reasoners to guarantee that data access aligns with health regulations such as HIPAA [215,216]. In modern wearable fitness devices, data such as heart rate, glucose and stress levels, medication adherence, menstrual cycle, and body metabolic ratio are measured. Health data captured outside of hospital settings often operates beyond compliance regulations. A recent study suggested a KG-based framework (KGF) to scrutinize the privacy policies governing the data and information utilized by wearable devices [46].

8. Evaluation, Open Challenges and Future Direction

8.1. Evaluation and Data-Centric Challenges

Health-KG methods are evaluated with task-specific and largely incompatible protocols: link prediction reports MRR and Hits@k; extraction reports precision, recall, and F1; and downstream clinical tasks report AUROC and AUPRC. Because each study constructs its own graph from a different slice of MIMIC or a private cohort, results are rarely comparable across papers, and improvements cannot be attributed cleanly to representation, extraction, or integration. The absence of a standardized, openly released Health-KG benchmark with fixed splits, a shared schema, and harmonized evaluation tasks is itself a finding of this rwview and a prerequisite for measurable progress. A community benchmark built on credentialed but openly governed resources (e.g., a MIMIC-IV-derived graph with canonical link-prediction, NER, entity-linking, and diagnosis-prediction tracks) would let the field compare embeddings, extractors, and LLM-grounded systems on equal footing.

Beyond evaluation, the underlying data pose challenges that any such benchmark must confront. EHR data are heterogeneous, noisy, sparse, and irregularly sampled, and they mix structured codes with free text and images, which complicates the clean triple/quadruple abstraction on which most Health-KGs rely. Missingness is informative (a missing lab is not missing at random), and the same clinical concept appears under multiple codes and surface forms; addressing this requires imputation-aware knowledge-representation learning, multimodal fusion at the representation stage [65,66], and quadruple schemas that attach provenance and confidence to each fact so that downstream reasoning can discount low-confidence edges. Cross-institutional integration compounds the problem, since divergent terminologies, local code systems, and schema mismatch mean that equivalent concepts are encoded differently across sites and naive merging produces conflicting or duplicate facts. Ontology matching and schema alignment [128,133,134], combined with LLM-assisted concept normalization validated against UMLS/SNOMED-CT and multiview integration methods that tolerate partial overlap, offer a path forward, as demonstrated for cross-institutional COVID-19 harmonization [137].

Two further data-centric axes remain open. Temporally, clinical knowledge and patient state evolve, yet most Health-KGs are static snapshots: relations are valid only over intervals (a medication is current, then discontinued), and disease trajectories depend on order and timing, motivating temporal knowledge-graph completion [78], time-stamped quadruples, and continual/streaming construction that updates the graph as new encounters arrive rather than rebuilding from scratch. In terms of modality, most Health-KGs are text- and code-centric, omitting imaging and omics; aligning continuous signals (radiology, pathology, genomics) with discrete graph entities requires shared representation spaces and multimodal encoders that project image- and omics-derived features into the same embedding space as clinical concepts [67], enabling a single graph to support image-grounded reasoning.

8.2. Trustworthy, Privacy-Preserving, and Reproducible Construction

EHR-derived graphs contain protected health information, and LLM-based construction risks exposing it through prompts. HIPAA and GDPR [213,214] constrain data movement, while useful KGs require cross-site information; reconciling these pressures calls for federated knowledge-graph embedding, on-premises or de-identified LLM extraction, synthetic-data augmentation, and policy-aware access control such as the compliance frameworks discussed in Section VII-D [46,204,215].

Trust concerns extend from privacy to factuality as LLMs enter the construction loop, where factuality and reproducibility become first-order concerns. Hallucinated facts can silently corrupt a clinical KG [190], and few studies release code, graphs, or prompts, leaving reporting inconsistent. Mandatory provenance on every asserted fact, automated verification against trusted ontologies, and adoption of reporting standards are therefore needed so that construction and evaluation are reproducible.

9. Conclusions

This review examined the construction and application of Health Knowledge Graphs from electronic health records along three axes-representation, extraction, and integration which keeping the review organized the field through a unified taxonomy and an explicit positioning against prior reviews. We reviewed knowledge-representation learning and graph-embedding models, entity recognition, disambiguation, and linking, and the integration methods that reconcile heterogeneous clinical sources, together with the principal EHR datasets and clinical language models that underpin them. We then synthesized the emerging role of large language models in Health-KG construction and KG-grounded clinical reasoning, including retrieval- and graph-augmented generation and the factuality constraints that govern their safe use. Across drug development, diagnosis and treatment recommendation, question answering, and policy compliance, Health-KGs consistently improve the structure, provenance, and interpretability of clinical decision support. We argued that the central obstacles to measurable progress are the absence of a standardized, openly governed Health-KG benchmark and the intertwined demands of temporal modeling, multimodal fusion, privacy-preserving construction, and factuality verification, which we identify as priorities for future work.