Real-world data medical knowledge graph: construction and applications

Highlights

• This paper proposes a systematic approach to construct medical KG from EMRs.

• The constructed KG contains 9 entity types, totally 22,508 entities and 579,094 quadruplets.

• Quadruplet structure is proposed to represent a fact in the real-world KG.

• The proposed ranking function PSR achieves the best performance under all relations.

• The obtained KG is successfully applied to several practical applications.

Abstract

Objective

Medical knowledge graph (KG) is attracting attention from both academic and healthcare industry due to its power in intelligent healthcare applications. In this paper, we introduce a systematic approach to build medical KG from electronic medical records (EMRs) with evaluation by both technical experiments and end to end application examples.

Materials and Methods

The original data set contains 16,217,270 de-identified clinical visit data of 3,767,198 patients. The KG construction procedure includes 8 steps, which are data preparation, entity recognition, entity normalization, relation extraction, property calculation, graph cleaning, related-entity ranking, and graph embedding respectively. We propose a novel quadruplet structure to represent medical knowledge instead of the classical triplet in KG. A novel related-entity ranking function considering probability, specificity and reliability (PSR) is proposed. Besides, probabilistic translation on hyperplanes (PrTransH) algorithm is used to learn graph embedding for the generated KG.

Results

A medical KG with 9 entity types including disease, symptom, etc. was established, which contains 22,508 entities and 579,094 quadruplets. Compared with term frequency - inverse document frequency (TF/IDF) method, the normalized discounted cumulative gain (NDCG@10) increased from 0.799 to 0.906 with the proposed ranking function. The embedding representation for all entities and relations were learned, which are proven to be effective using disease clustering.

Conclusion

The established systematic procedure can efficiently construct a high-quality medical KG from large-scale EMRs. The proposed ranking function PSR achieves the best performance under all relations, and the disease clustering result validates the efficacy of the learned embedding vector as entity’s semantic representation. Moreover, the obtained KG finds many successful applications due to its statistics-based quadruplet.

where $N_{co}^{\text{m}\text{i}\text{n}}$ is a minimum co-occurrence number and R is the basic reliability value. The reliability value can measure how reliable is the relationship between S_i and O_ij. The reason for the definition is the higher value of N_co(S_i, O_ij), the relationship is more reliable. However, the reliability values of the two relationships should not have a big difference if both of their co-occurrence numbers are very big. In our study, we finally set $N_{co}^{\text{m}\text{i}\text{n}}$ = 10 and R = 1 after some experiments. For instance, if co-occurrence numbers of three relationships are 1, 100 and 10000, their reliability values are 1, 2.96 and 5 respectively.

Introduction

Knowledge graph (KG) has received a lot of attention in recent years. In 2012, Google applied the KG in search engines; since then, the knowledge graph has been used in many application fields [1]. In the medical domain, the knowledge graph is the fundamental component for artificial intelligence (AI) aided medical systems, such as clinical decision support systems (CDSSs) for diagnosis and treatment [[2], [3], [4], [5]], self-diagnosis utilities to assist patient evaluating health condition based on symptoms [6,7].

The KG is a graph-based knowledge representation and organization method, which uses a set of subject-predicate-object triplets to represent the various entities and their relationships in a domain. Each triplet is called as a fact as well. In KG, nodes represent entities and edges represents relationships between entities. For example, ‘Parkinson's Disease’ and ‘Tremor’ are concrete entities of type Disease and Symptom in the medical domain. Given that ‘disease_related_symptom’ is a relationship between disease and symptom entities, ‘Parkinson's Disease disease_related_symptom Tremor’ is a triplet to represent that ‘Tremor’ is a related symptom with ‘Parkinson's Disease’.

Most previous works tried to construct the KG from medical articles, some of them are constructing manually and others are automatically. However, manually constructing KG requires tremendous clinical expert time and effort. For example, it was reported that about fifteen person-years are required to build the Internist-1/QMR knowledge base [8]. Automatically constructing KG from articles is a challenging work as the materials are almost unstructured, which is difficult to understand by computer.

In recent years, thanks to the rapid progress of big data and natural language processing (NLP) technologies, automatically mining knowledge from electronic medical records (EMRs) becomes a promising research trend [[9], [10], [11], [12], [13], [14], [15], [16],20,21],. Learning medical KG from EMRs is less labor-consuming and more feasible than learning from articles. More importantly, the statistical properties of real-world data based KG make it easier to use.

A lot of papers introduces EMR processing algorithms, including named entity recognition (NER) [[9], [10], [11], [12], [13]], entity normalization [[14], [15], [16]], relation extraction/ranking [26] and graph embedding [20,21]. However, there still lacks an efficient and systematic procedure to build medical KG from EMR data end to end. This paper aims to establish a systematic procedure to construct the medical KG from large-scale EMRs. The study is performed on a big-data platform of a 3A-class hospital in China and the constructed KG contains 9 entity types, totally 22,508 entities. Based on the data, we propose a new quadruplet structure to represent a KG fact, in contrast to the classical triplet structure, and build a total of 579,094 quadruplets. PrTransH [21] is used to train embedding vector for each entity and relation. To evaluate the effectiveness, the obtained KG is applied to several practical applications including CDSS, information retrieval and knowledge transfer with neural networks. At the end of this paper, conclusion of this paper and the prospect of the future work are drawn.