Healthcare knowledge graph construction: A systematic review of the state-of-the-art, open issues, and opportunities

KGs are receiving a lot of attention from researchers who are involved in the drug development studies. The necessity to construct specific KG for the drug sector has several key motivating factors [46]; prescribing a particular drug to treat a certain disease might involve some non-medical factors including the demographics, insurance policy, drug availability, etc. Further, in some instances, healthcare professionals who are not qualified to prescribe drugs, might act upon an emergency, thereby initiating a treatment that in a normal situation has to be initiated by a specialist doctor. Such complications illustrate the need for an intelligent platform that can actively guide the search for the optimal drug to prescribe. In this context, Mann et al. [46] attempted to create such a platform that can assist in finding a valid treatment considering the known symptoms or identified disease. In this study, the authors integrated existing medical knowledge resources, thereby building a KG to benefit the entire domain. In the same line of research, Che et al. [47] proposed a method to integrate six knowledge bases into one coherent KG. The resulting KG is then embedded into Graph Convolutional Network with an Attention mechanism for Drug–Disease Interaction (DDI), which is used to predict and discover potential drugs capable of effectively treating COVID-19. The prediction of Drug-Target Interaction and Drug-Drug Interaction are important aspects of the development of new drugs. In another interesting study, Zhang et al. [48] constructed two designated KGs describing drugs captured from a biological dataset, namely Bio2RDF¹⁰. This is followed by developing a learning model based on graph representation (MHRW2Vec), whose output was fed to a neural network model (TextCNN-BiLSTM Attention Network (TBAN)). The ultimate objective was to predict potential interactions of various COVID-19 drugs. Ye et al. [49] developed KGE_NFM, an integrated framework comprising both a KG and a recommender system to predict DDI. The components of the KG were embedded in a low-dimensional space, after which a neural factorization machine was tasked to build the recommender system for drug target discovery. Drug discovery incorporating KG technology was also discussed in [50‐53].

Drug repurposing (a.k.a. reprofiling, redirecting, rediscovery, or repositioning) is an interesting domain that has come into focus recently. It aims to reuse existing drugs to treat emerging diseases (such as COVID-19) thereby reducing both drug development timelines and the associated costs [54]. Therefore, various studies attempted to provide intelligent solutions for the challenges inherent in drug repurposing. Regarding the use of KGs, there is a direction of research aimed at constructing KGs that can be used for drug repurposing. BenevolentAI’s proprietary KG [55] is amongst the most successful approaches in this research line. The BenevolentAI KG integrates an assortment of medical data obtained from structured and unstructured scientific repositories (including literature). It is queried by various algorithms to identify new relationships between entries, thereby suggesting new ways of treating diseases. In the same context, Wang et al., [56] proposed a framework called COVID-KG which aimed to construct a KG from multimodal data found in scientific literature into one actionable KG that can be used for drug repurposing. The proposed KG is built on an ontology of 77 entity subtypes and 58 event subtypes, as defined in the Comparative Toxicogenomic Database (CTD) (Davis et al., 2016), and entities are linked using Medical Subject Headings (MeSH) framework [57]. also proposed a multimodal drug repurposing KG for COVID-19 that was built with data harvested from scientific literature, and aimed to provide an overview of pathophysiology related to COVID-19. The construction of this graph was carried out manually using Biological Expression Language, and evaluated based on multiple case studies. Drug repurposing is further discussed in [58‐63].

ADRs refer to undesired reactions that occur after the use of a certain medical product [64]. ADRs carry significant risks to both patients and the hospital system [65], thus serious attention is required to tackle this issue and develop optimal technological solutions to mitigate it. The sophisticated structure of KGs presents an opportunity to define this problem conceptually and provides new ways to predict potential ADRs. Many studies were conducted in this direction, notably Bean et al. [66] benefited from access to two drug resources (namely DrugBank¹¹ and SIDER¹²) to build a KG that can predict ADRs. This KG contains four types of nodes and three types of edges, and is consolidated with a prediction model (similar to linear regression). Authors of [67] introduced a KG to represent drugs and ADRs, with data embedded using the Word2Vec model. On top of this model, logistic regression was used to predict whether a given drug causes any ADRs. Tumor-Biomarker Knowledge Graph (TBKM) [68] is another attempt to design a KG with four node classes (namely Tumor, Biomarker, Drug, and ADR) based on data from scientific biomedical literature. The aim of the KG is to discover ADRs of antitumor drugs as well as provide explanations why they occur. Predicting and discovering ADRs have been further reported in [69‐74]. Zhao et al. [75] designed their drug action mechanism KG after extracting information from 770,000 abstracts of medical papers. Despite the poor approach used to extract entities and their relationships, the paper managed to cover a large number of drugs and mechanisms of action. Table 1 illustrates a summary of currently proposed KG construction approaches for drug discovery, drug repurposing, and adverse drug reaction.

KGs have accelerated the pace of scientific discovery that aims to better understand diseases affecting the human body. For example, Zhang et al. [15] developed Health Knowledge Graph Builder (HKGB), which is a framework that can be used to construct a Health KG (HuadingKG) from multiple sources (namely EMRs, medical standards, and expert knowledge) to be used in the cardiovascular domain. To conceptualise subarachnoid haemorrhage stroke, the authors of [45] developed a comprehensive framework that allowed them to construct a KG from heterogeneous data automatically. In particular, the authors incorporated semantic analysis for entity and relation extraction, and implemented a knowledge prediction model based on the association rule and ensemble machine learning. KGHC [76] is a KG designed specifically for Hepatocellular Carcinoma. It brings together and connects entities captured from 5 different unstructured and structured data sources and extracted using information extraction techniques such as BioIE and SemRep. Yin et al. [77] constructed a KG for diagnosing and treating viral hepatitis B by adopting a top-down approach where a domain ontology was used to build the KG. The authors did not provide adequate details on the mechanism utilised to construct the KG or the evaluation metrics. Yet, they claimed that the designed KG benefits intelligent recommender systems that can be used to diagnose and treat viral hepatitis B. Another research direction identified the role of genes in human disease [78]. The authors built a convolutional neural network-based model on top of a biological KG to classify the genes highly correlated with cancer. While the construction of the KG itself was not adequately validated, the resultant embedding model was evaluated on downstream tasks. An attempt at preventing Myopia using KG technology was described in [79]. The authors developed a KG from various Chinese websites to provide intelligent Q&A services to users interested in Myopia prevention. However, the finalised KG lacks multimodal data that can be captured from medical databases and domain-relevant question answering systems. Conceptualising Stroke and its causes and effects is an extensively covered subject in the literature. For example, Yang et al. [80] constructed an integrated KG, named StrokeKG, that portrays various stroke-relevant relationships inferred from various medical datasets. Designing KGs to benefit the victims of stroke was also examined in [81, 82]. COVID19-related disease discovery using KGs was reported by Huang et al. [83]. Relying on a pipeline approach, the authors surveyed from relevant scientific papers related to COVID-19 and used the collected data to construct a KG that can identify diseases and drugs associated with COVID-19. The accuracy of the extracted knowledge was then verified using the time-slicing method. The use of KGs in the healthcare domain was discussed with a focus on disease identification and prediction in [84‐87], detecting the association between miRNA and disease in [88], chronic disease management in [89], and syndromes diagnosis in [90].

Yuan et al. [91] constructed a KG with minimal supervision to frame autism spectrum disorder diseases, using the articles obtained from the PubMed dataset. Entities were extracted using MinHash lookup/ UMLS [92] and formed into pairs which were then clustered using kmeans + + based on similarity between entities. Constructing KGs that can describe depression was undertaken by Huang et al. [93]. In particular, they attempted to generate a sub-graph that describes depression disorder, obtained by parsing data from a variety of major knowledge sources such as PubMed, Medical Guidelines, DrugBank, Unified Medical Language System (UMLS) etc. In the same line if research, Li et al. [94] proposed a UMLS-based semantic prediction program, known as SemRep, as well as SemMedDB to construct a KG for describing depression by using a bottom-up approach. Depression and its association with metabolism is also discussed in [95]. The authors developed MDepressionKG KG that integrates data about human microbial metabolism network, human diseases, microbes, etc., to offer semantic-based rational reasoning and establishing probable relations between depression and comorbid diseases. Although the authors furnish a useful online website to demonstrate utility of MDepressionKG, the knowledge inference mechanism is ineffective due to the incorporated traditional rules of logic. Furthermore, automatic extraction methods are required to enrich the functional diversity of the proposed depression KG. The conceptualisation of various mental disorders through graphs was also presented in [96‐99]. Table 2 shows a summary of KG construction approaches for diseases and disorders.

KG’s have been used with success for modelling both biological systems and pathologies, providing the means to understand this interplay between them. Several studies reported significant advances in this direction while incorporating KG technology. PharmKG [100] is a comprehensive KG built upon integrating six interrelated knowledge bases, with nodes representing genes, chemical compounds, and diseases. Entities of PharmKG are labeled with domain-specific information, keeping semantic and biomedical characteristics of the data. Percha et al. [101] compiled a basic KG known as the Global Network of Biomedical Relationships (GNBR) from biomedical literature. The process of populating GNBR with data was performed using PubTato (named entity annotator) tool, as well as Ensemble Biclustering for Classification (EBC) algorithm to annotate entities captured from Medline abstract. Wood et al. [102] developed RTX-KG2, an integrated KG that includes biomedical data captured from 70 biomedical knowledge bases. The aim of RTX-KG2 is to offer an open-source KG that can be used as a biomedical translational reasoning engine. Zhang et al. [103] reported the extraction of biomedical causality from the scientific literature. In their work, the authors constructed a biomedical knowledge graph to discover causal relationships in the biomedicine field. Authors of [82] developed a marine Chinese medicine KG using a top-down approach that takes guidance from a domain ontology. Developing an integrated KG to benefit the biomedical domain has also been discussed in [104]. The authors presented BioKG, a KG of drug-drug and drug-protein interactions data collected and compiled using modular software, namely BioDBLinker. This KG contains managed entities and relationships captured from at least five biomedical databases, such as UniProt, REACTOME, KEGG, DrugBank, SIDER, and Human Protein Atlas (HPA). He et al. [105] designed a KG for intestinal cells. First, the authors built an ontology as a conceptual model followed by extracting facts from the academic literature. Despite the problems with mechanisms used to construct the actual KG, the work presents an important attempt toward constructing KGs specifically to study the intestinal field, facilitating much easier observation of the processes of intestinal cytokines via various signalling channels. Constructing KGs to benefit generic biomedical domain was the subject of [106‐111].

KGs offer an excellent mechanism to conceptualise our understanding of microscopic organisms and their ecological traits. To this end, Joachimiak et al. [112] developed KG-Microbe, an integrated KG that contains prokaryotic data for phenotypic traits as well as supporting use cases in microbiology, biomedicine, and environmental science. Liu et al. [113] conceptualised gut microbiota using a semantically enriched KG, namely MiKG. MiKG integrates facts obtained from medical literature as well as other medical knowledge bases, thereby offering an interface for detection of possible connections between gut microbiota, neurotransmitters, and mental disorders. Authors of [114] developed a Microbe-Disease Knowledge Graph (MDKG) through an explorative study, thus identifying the associations between bacteria and diseases. MDKG is populated with entities and relations captured from textual content of Wikipedia as well as other semantic knowledge bases. Modelling Coronavirus using KG technology has recently attracted a lot of attention in the research community. For example, Zhang et al. [115] built a coronavirus KG by integrating entities captured from Analytical Graph and CORD-19 databases. The aim of the proposed KG is to provide a tool for the exploration of coronavirus on the entity level. Another attempt to help the biomedical research community comprehend the coronavirus using KGs is offered by Chen et al. [116]. The authors constructed a designated KG to discover any associated diseases, potentially effective drugs or treatments, and relevant genes and mutations. Using KG technology, modelling Coronavirus relevant information was also implemented and discussed in [57, 117, 118]. Further uses of KGs in microbiology are studied in [119]. Table 3 depicts a summary of KG construction approaches for the biomedical domain.

The ongoing efforts to leverage the proliferation of EMRs for multiple medical applications are well-documented in the scientific literature. Extracting valuable knowledge from such data silos has been made easier by KG technology. In this context, several studies attempted to construct medical KGs that can improve specific areas, for example, clinical decision support systems. One such attempt was undertaken by Li et al. [120], who followed a systematic approach consisting of eight steps to build a medical KG from EMRs obtained during the patients’ visits. The authors constructed a quadruplet-based medical KG incorporating an additional item (properties) which includes a set of characteristics to rank the embedded entities. The main objective of this study is to ensure the robustness of facts in the KG related to the medical domain. Evaluating the robustness of a constructed KG in healthcare is of utmost significance to ensure the quality of the inferred knowledge. In this context [121], presented a methodology to measure and evaluate the robustness of knowledge relating to diseases and symptoms, with data captured from existing health knowledge graphs as well as records of patient visits to the Beth Israel Deaconess Medical Center (BIDMC). Postiglione et al. [122] proposed an advanced entity recognition approach named PETER (Pattern-Exploiting Training for Named Entity Recognition), that integrates Pattern-Exploiting Training (PET) [123] to build an Italian-language KG for healthcare. EMRs represent a fertile source of information for healthcare KGs, hence their use for construction of KGs is becoming quite common, as exemplified in [124] and [125].

Incorporating health KGs into a QA system was discussed by Sahu et al. [126]. The authors proposed a system that can be used to search for various health-based KGs and obtain a set of healthcare-related response sub-graphs. The possibility of using medical KG’s to benefit QA applications was also discussed in [127]. Zhao et al. [128] made use of EMRs obtained from hospital patient records in Shanghai to build a medical knowledge graph based on the BILSTMCRF model. Here, a KG is used as a part of a QA system to provide support for establishing medical diagnosis. Xie et al. [129] attempted to create a KG for Traditional Chinese Medicine (TCM), yet the KG they ended up with is very limited in terms of entities and relationship; thus, the applicability and utility of the graph is questionable. Another Chinese medical KG was proposed by [130]. The authors developed this KG from various structured, semi-structured, and unstructured resources and built a QA system that was not adequately validated due to irrelevant results. Also, Huang et al. [131] designed a QA system based on a constructed KG, with the information in the graph used to identify the question’s intention. Deploying QA and Q&A systems based on medical and healthcare KGs is also a relevant topic in [132].

In healthcare, addressing the timing factor in KG creation is critical. Ma et al. [144] developed a temporal KG that is useful for studying episodic memory in cognitive tasks. The Integrated Conflict Early Warning System (ICEWS) dataset and the Global Database of Events, Language, and Tone were used to create this temporal KG (GDELT). Their work was unique in that it involved four substantial static KGs embedding data to four-dimensional temporal/episodic KGs, which set them apart from other efforts in this direction. Two new RESCAL generalisations were also proposed and considered. Another important effort that integrated plausible reasoning with fine-grained biomedical ontologies to tackle the data incompleteness problem was undertaken by Mohammadhassanzadeh et al. [42]. The authors proposed a Semantics-based Data analytics (SeDan) framework that performs an exploratory analysis of the KG using the OWL extension and query rewriting algorithm. The framework incorporates data from various knowledge bases, including the DrugBank, Disease Ontology, and the large-scale semantic MEDLINE database (SemMedDB). Rastogi et al. [145] framed their personal health KG as a combination of context, personalization, and integration with other knowledge bases. Their study indicated that the literature on personalised health-related KGs is incomplete and lacks a unified standard representation to adequately describe the designated domain. To provide an overview of effective medications, side effects, and target populations relevant to COVID-19, the authors of [146] proposed a KG-based framework to support COVID-19 clinical research. This framework benefited from Stanford’s Stanza toolbox to extract KG’s entities and relationships that can be fed into a visualisation module for querying information. The application of KGs in healthcare and medical domains was documented in other relevant tasks including epidemic contact tracing [147], food waste detection [109], drug similarity [148], clinical decision support systems [120], and medical recommender systems [149, 150]. Table 4 shows a summary of KG construction approaches used in various miscellaneous healthcare applications.