Article Preview
TopIntroduction
Drug instructions contain a wealth of drug knowledge, which can be extracted for decision support in clinical diagnosis, prescriptions, and healthcare management. However, mining drug instruction knowledge is a challenging task. Textual descriptions in drug instructions usually take the form of long, complicated sentences, which are difficult to be processed by man or machine. Existing information extraction methods have been used for drug and disease entity recognition (Lin, & Xie, 2020; Sun et al., 2021; Zhu et al., 2021), drug-disease relationship extraction (Bose, et al., 2021; Fatehifar, & Karshenas, 2021; Mingliang, Jijun, & Fei, 2021), etc. Recently, pre-training models have become prominent in Natural Language Processing (NLP) tasks due to its generalization capability (Vaswani et al., 2017). However, many existing methods still consider the knowledge extraction task as a pipeline process that consists of Named Entity Recognition (NER) and Relation Extraction (RE) tasks. These pipeline methods are prone to error propagation since not all entities generated from NER are valid for RE (Pawar, Palshikar, & Bhattacharyya, 2017). In drug instructions, it is very common to have multiple entities and relations in a single sentence, without proper entity pair checking, the performance of fine-grained relation extraction will be significantly affected by the invalid entity pairs (Gao, Zhou, & Gu, 2021). Siriwon Taewijit, & Thanaruk Theeramunkong (2021) proposed an approach to learn meaningful patterns by hyperbolic embedding and then extract adverse drug reactions from electronic medical records. As a result, manual effort is required to prepare entity pairs. In other cases, joint methods are proposed to extract entities and relations in a single process (Wei et al., 2020), however, these methods struggle with many-to-many relations that involve overlapping entities (Tiktinsky et al., 2022).
Table 1. Sample drug instruction text with multiple entities and relations
Compound paracetamol should not be used together with chloramphenicol, barbiturates (such as phenobarbital), etc. |
To overcome the limitations in the current approaches, this paper proposes a novel pipeline drug information extraction framework using an entity pair calibration module based on pre-training models. It starts by automatically gathering the textual descriptions from public drug instruction data sources. Then, it uses deep learning models to identify four categories of entities: drugs, diseases, body parts, and symptoms. After the entity name recognition, it gathers the sentences with multiple relevant entities and applies the entity pair calibration (EPC) process. The goal of EPC is to distinguish between the main entity (the primary drug which the instruction is intended for, e.g., the first drug mention in Table 1) and the secondary entities (the other entities mentioned in the sentence of the instruction, potentially associated with the primary drug entity, e.g., the second, third and fourth drug mentions in Table 1). This step can reduce the noise for further drug relation extraction and facilitate accurate extraction of fine-grained relations. Finally, it extracts fine-grained relations from those entity pairs.
The main contribution of this paper is two-fold: