Chemicals and drugs have made a great contribution to human health care. At the same time, they are rarely free from occasional adverse drug reactions (ADRs) [1], which are defined by the World Health Organization (WHO) as any noxious, unintended, and undesired effects of a drug that occur at doses used for the prevention, diagnosis, and treatment of a disorder. A significant number of ADRs occur each year. ADRs are the sixth leading cause of death worldwide, and the fourth primary cause of death in the United States and Canada, behind cardiovascular disease, malignant neoplasm, and stroke [2]. Although the actual incidence of ADRs is difficult to access precisely, it is known that ADRs have a considerable impact upon both health care and pharmaceutical manufacturers.

Package inserts (sometimes called patient information leaflets) are the primary official papers which accompany most prescribed drugs and over-the-counter medications. Although different countries have diverse requirements for the obligatory contents, the package inserts serve at least two main purposes. They contain informative details regarding the generic names of drugs, active ingredients, indication for use, instructions for use, special warnings, contraindications, and statistical values from clinical trials, including the percentage of people who had side effects, the types of side effects, and additional precautions. Furthermore, the package insert is an easy reference for physicians when prescribing medications, and can help them avoid prescribing drugs that may be contraindicated. The inserts also serve as an easy reference for patients. However, the informative package inserts are generally semistructured and cannot be understood easily by machines. As the number of newly licensed drugs increases, the demand for automatic technology for semantic integration and linkages, as well as bioinformatic analysis of the information (including ADRs) enclosed in package inserts, has become an urgent issue both in biomedical research and the pharmaceutical industry.

In information science, an ontology is the formal, explicit specification of a shared conceptualization of a domain [3]. Generally, biomedical ontologies not only represent the essential properties of biomedical entities and their correlations to other biomedical concepts, but also provide a standardized vocabulary and formalized knowledge source for the biomedical community. Hundreds of biomedical ontologies have been elaborately built to support scientific discovery and the analysis of biomedical data [4-8]. Moreover, systematically evaluated ontology is one of the two fundamental sources of background knowledge for artificial intelligence algorithms in biomedicine (the other is the knowledge graph).

In this paper, we propose an ontological organization of traceable ADR information extracted from licensed package inserts, which aims to provide machine-understandable knowledge for bioinformatics analysis, semantic retrieval, and intelligent clinical applications (Figure 1). This entails the following: (1) Present a generic ADR ontology model from two dimensions (and nine subdimensions) covering the essential ADR information and medication instructions. (2) Customize a Python natural language processing (NLP) method to automatically retrieve the identified information enclosed in package inserts. (3) Collect the approved package inserts of quinolone drugs and build a specialized ontology for algorithm verification and validation. (4) Bioinformatically analyze the adverse reactions caused by quinolones based on the obtained ADR ontology, and discuss potential applications including semantic retrieval and a clinical decision-making system.

We built the ADR ontology using two dimensions: ADR information and drug-related medication instructions, which were further separated into nine subdimensions. The former covers the following: (1) adverse drug reactions, presenting diverse types of ADRs; (2) the occurrence of ADRs, describing the frequencies of ADRs after the administration of a drug in a population; (3) the severity of ADRs, describing a general measure of the subsequent risks of potential ADRs; (4) populations affected by specific ADRs, noting the individual human patients associated with an adverse reaction after the administration of a drug; and (5) the mitigation methods of ADRs, referring to any measure that shortens the duration of an adverse reaction or reduces its severity. The drug-related medication instructions contain generic information about rational drug use: (6) drug names, describing different pharmaceutical agents; (7) dosage forms of drugs, collecting the complete form of the pharmaceutical preparation used to administer the prescribed dose of medication; (8) administration routes of drugs, consisting of the various ways of administering a drug to a patient to allow the chemical to be absorbed into the blood and delivered to the target tissue; and (9) contraindications of drugs, predefining a condition or factor associated with a recipient that makes the use of a specific drug improper or inadvisable. These nine branches determine the fundamental concepts and classes of our ontology. Figure 2 demonstrates the nine classes of proposed ADR ontology as well as the eight semantic relations among them. The generic model will help to generate specialized ADR ontologies on request as illustrated in the next section.

The biomedical terms denoting nine classes within the ADR ontology are much more comprehensive than shown in Figure 1. With reference to controlled biomedical vocabularies and well-developed pharmacopeia, we integrated many synonymous terms (both in English and Chinese) into various concepts and organized them in a hierarchical way. Generally, the WHO Adverse Reactions Terminology (WHOART) is reused to generate the ADR class and its hierarchy in four levels: system/organ classes (SOC), high level terms (HLT), preferred terms (PT), and included terms (IT). Among them, SOC and HLT consist of broad grouping terms, while PT represents more specific adverse reactions and IT are entry terms (synonyms) for PT. The large number of PT and IT as well as their synonyms from the Medical Dictionary for Regulatory Activities (MedDRA) will become the essential vocabulary to recognize the named entities of ADRs in package inserts. Furthermore, the Council for International Organization of Medical Science (CIMOS) has recommended five terms derived from different percentages to classify the occurrences of ADRs: Very Rare (<0.01%), Rare (≥0.01%, <0.1%), Uncommon (≥0.1%, <1%), Common (≥1%, <10%), and Very Common (≥10%). This forms the basis for the classification of ADR frequencies in our work. The Common Terminology Criteria for Adverse Events (CTCAE) is a standard classification and severity grading scale for adverse events in clinical trials and oncology settings. We adopted CTCAE for labeling the severities of ADRs with five levels: Grade 1 (mild), Grade 2 (moderate), Grade 3 (severe but not immediately life-threatening), Grade 4 (life-threatening) and Grade 5 (death caused by ADRs). The other classes that are closely related to the pharmaceutical agents and their medication instructions will be acquired and represented according to the classic pharmacopeia for ease of use.

In addition, two kinds of data properties will be incorporated into the ADR ontology, including a general description of the drugs (such as the active ingredient, injection excipients, drug specifications, antibacterial mechanism, interaction drugs); the key codes and definitions from referenced vocabularies containing ARecNO (the PT code) and the SOC code from WHOART, the MedDRA code, the NCIt code; and definitions from CTCAE. Furthermore, the origin of the package inserts will be recorded in the ontology’s annotation property for ontology data identification and traceability.

From the viewpoint of ontology construction, the ADR information and medication instructions extracted from package inserts will become the instances of object properties within the ADR ontology. Due to the large amount of data enclosed in package inserts, manual extraction would be a labor-intensive process. We developed a Python NLP-based algorithm to automatically retrieve ADRs and drug-related information, which consists of two steps: named entity recognition (NER) and relation extraction (RE).

Briefly, NER will recognize a string of text as an entity (eg, an adverse reaction) that is already defined in our ADR ontology. The Jieba word segmentation model implemented in Python is adopted to segment the words enclosed in package inserts, while the names of classes and their synonyms from the proposed ontology will function as a domain vocabulary to improve performance. RE is a process that determines whether two entities have a specific relationship (eg, the “has_ADR” causality between a particular drug and an adverse reaction). Since formal package inserts have already been separated into several titled sections, these titles will be used to implement the RE task. Specifically, the titles “ADVERSE REACTIONS” and “WARNINGS AND PRECAUTIONS” are converted into semantic relations about ADR information, while the medication instructions for a particular drug would be extracted from the sections titled “DOSAGE AND ADMINISTRATION,” “CONTRAINDICATIONS,” and “USE IN SPECIFIC POPULATIONS.”

The information automatically retrieved from package inserts will be passed to downstream biocuration for data identification. The major criteria of the manual review process emphasize the following two points: (1) whether the medication instructions and information about ADRs caused by a specific drug was accurate and complete, without missing data or mistakes; and (2) the frequency and severity of a drug-ADR triple must be explicit; vague descriptions were not recorded. Eventually, the human biocurated semantic information is used to build the ADR ontology.

The Ontology Web Language (OWL) is a widely used programing language for defining and instantiating web-based ontologies. It provides a machine-understandable schema to describe classes and their semantic relations in a specific domain. In this work, the OWL application programming interface, a Java interface and implementation for OWL, is used to build the ADR ontology automatically based on content data primarily obtained from package inserts and well-established vocabularies (eg, WHOART, MedDRA, CTCAE). Moreover, to ensure high-quality results, the ADR ontology will be evaluated and validated by checking the clarity, coherence, extendibility, minimal encoding bias, and minimal ontological commitment, which are the fundamental principles of building a domain ontology [31].

Quinolone drugs have become commonly used antibacterial agents due to their strong and broad-spectrum antibacterial activity, as well as their rapid and complete absorption in humans. As quinolone usage increases, the risk of ADRs increases proportionally. According to the Annual Report of National ADR Monitoring in China, the number of adverse reactions and events from quinolones has continuously been the second-highest among antibacterial drugs. Therefore, the safe administration of quinolones is a serious matter that requires more attention.

After half a century of development, quinolones have evolved from the first generation to the fourth generation. We collected 165 specific drug names of quinolones from the China Pharmaceutical Reference and the National Essential Drugs List in China, and further organized them based on their generation. The electronic package inserts of these drugs were then downloaded from the National Medical Products Administration and Yaozhi drug database in China. Although these package inserts were written in Chinese, the Chinese and English synonyms of the class names defined in the proposed ADR ontology will enable information extraction based on a customized Python program. Finally, an ADR ontology for quinolones was automatically built and evaluated for further analysis.

Currently, the specialized ADR ontology covers a total of 2879 classes grouped into nine categories: ADRs and their occurrence, severity, population, and mitigation methods, as well as 165 drugs and their dosages, administration routes, and contraindications (Figure 3A). These classes were deeply divided into subclasses on the basis of biomedical concepts and arrayed hierarchically from most general to most specific in up to 4 levels, with the abovementioned nine categories at the top (Level 1) of the ontology hierarchy (Figure 3B). Furthermore, the obtained ADR ontology also includes eight types and 15,711 nonredundant semantic triples (Figure 4A) extracted from package inserts, where the causalities between ADRs and quinolones (with the type label “has_ADR”) account for a large proportion (n=7725, 49.17%). Among the 7725 identified drug-ADR triples, 4043 (52.34%) and 716 (9.27%) explicitly stated the frequency and severity, respectively. The statistics of drug-ADR triples by frequency and severity suggest that most ADRs caused by quinolones (n=4037, 52.26%) occur at a low frequency (<10%), and none of them are life-threatening (Figure 4B).

Figure 5 demonstrates how the proposed ontology organized the ADR information enclosed in the package insert, using the Protégé OWL editor for ontology visualization. Briefly, levofloxacin hydrochloride tablets have been reported to account for different types of ADRs (including dizziness, headache, and insomnia) that have already been defined in our ontology. For each adverse reaction, it is likely that the frequency, severity, and mitigation method are informed by the patient population. Therefore, we used four semantic relation types (“has_occurrence,” “has_severity,” “has_mitigation method,” and “has_population”) to link information associated with a particular adverse reaction caused by a specific drug (the lower right section of Figure 5). The “AND” clause is adopted to combine multiple semantic triples into a composite knowledge unit. It is worth mentioning that we extracted ADR information from the explicitly stated content in package inserts; thus, some of the four types linked to a given drug-ADR triple may not be obtained due to a missed or implicit description. To enable knowledge tracing, the package insert citation was recorded as a referenced annotation property (the upper right section of Figure 5). In addition, the Internationalized Resource Identifier (IRI) [32], which is the unique identifier for ontology sharing and reuse around the world, was customized as “ADR+six numbers” within the proposed ADR ontology. For instance, the IRI for the levofloxacin hydrochloride tablets used in the example is ADR000572.

We compared our work with four other ontologies (described in the Introduction) in terms of the semantic relations defined in the proposed ADR ontology (Table 1), since they are the essential knowledge unit for the wide utility of an ontology. The proposed ontology represents a complete set of information pertaining to ADRs containing drug-ADR triples and the associated frequency, severity, and mitigation methods. Moreover, the explicit medication instructions (eg, the dosage form, route, and contraindication) extracted from the package insert of a particular drug support physicians in guiding patients in the safe, effective, and rational use of drugs, as an adverse reaction can occur when using the prescribed dosage.

The ontology of ADRs caused by quinolones consists of 7725 drug-ADR triples retrieved from 165 package inserts. After duplicate removal, 331 ADRs were identified to be caused by quinolones. Table 2 lists the major ADRs from two dimensions: the quinolone generation and active ingredient. Since topical drugs have fewer ADRs than oral and intravenous ones due to their administration route, we concentrated on the comparison of quinolone drugs that are only administered by the oral and intravenous routes. After bioinformatic analysis, there were three important points that could be summarized as follows: (1) Levofloxacin (third generation) and gatifloxacin (fourth generation) induce a significant number of ADRs (n=139, 43.57% and n=122, 38.24%, respectively), and many drugs are made from these ingredients (n=21, 17.21% and n=20, 16.39%, respectively). (2) The highest numbers of ADRs are caused by ciprofloxacin (n=221, 69.28%) and enoxacin (n=156, 48.90%), respectively. (3) There are relatively few ADRs associated with nalidixic acid (n=21, 6.56%) and pipemidic acid (n=11, 3.45%), which implies that they are comparatively safe in terms of known adverse effects. In future work, we will investigate the ADR differences related to drug dosages.

The 331 ADRs could be further classified into 18 categories that we defined in the ontology (Table 3). Most organs are involved in ADRs. Skin reactions (eg, itching, skin rash) are the most common reactions and are linked to 141 drugs (85.45%), followed by nervous system reactions (eg, headache, hypertonia), which are associated with 127 agents (76.97%). Moreover, the least frequently reported adverse reactions are those of genital organs and the application site. This may be due to the difficulty in detecting genital organ diseases and the minor impact of injection site reactions.

Table 4 lists the top 10 ADRs caused by quinolones and shows that itching and skin rash are listed in nearly 80% of currently licensed quinolone package inserts in China. Although alanine transaminase (ALT) elevation and phlebitis frequently occur after the administration of 3 drugs made of pazufloxacin mesylate, other adverse reactions caused by quinolones are usually infrequent (<10%). All of the reported ADRs have a severity of Grade 1 (mid) to Grade 3 (severe but not immediately life-threatening), and stopping medication is generally recommended as the mitigation method after the detection of an adverse reaction. Additionally, quinolone drugs are administered to patients at various dosages and through different routes (Table 2). Finally, analysis of contraindication data found that the top 5 specific populations are those who are allergic to quinolones, pregnant women, teenagers, infants, and patients with central nervous system diseases.

The ADR ontology proposed in this study has two major applications. The first is the semantic retrieval system that can use the ADR knowledge to integrate various external sources of information. Since the primary ADR terms and their key codes were reused from WHOART and MedDRA, it should be straightforward to integrate our ontology with these controlled vocabularies, as well as other medical terminologies (eg, Logical Observation Identifiers Names and Codes [LOINC], Systematized Nomenclature of Medicine-Clinical Terms [SNOMED CT]), for semantic knowledge retrieval by providing formally represented ADR information. Another major application is intelligent clinical decision-making support. The proposed ADR ontology provides machine-understandable knowledge, which could be used by artificial intelligence algorithms in biomedicine. To pursue clinical and therapeutic approaches to optimal disease management and rational drug use, it is useful for a physician treating a specific disorder to know all the identified adverse reactions induced by the drugs prescribed for patients with that condition. Manually reading the package inserts to find the ADR information and medication instructions is laborious and time-consuming, as the number of newly approved drugs and reported ADRs increases every year. Conversely, our ontology will aid in the development of an intelligent clinical decision-making system, which would positively affect the drug prescribing patterns of physicians and potentially have a significant socioeconomic impact.

We have shown that the ADR ontology can be used to formally represent and organize the ADR information and medication instructions enclosed in official drug packages. In addition, it can provide machine-understandable knowledge for bioinformatic analysis. By collecting 165 package inserts of quinolone drugs, a specialized ADR ontology was built to classify various identified ADRs caused by quinolones. Systematic analysis of the obtained ontology data improves the deep recognition of drug-specific ADRs, making it possible to intelligently guide safe drug use and benefit human health.

The proposed ADR ontology can be generalized to organize the ADR information from other channels, not just package inserts. Electronic health records, in which the ADR data are frequently reported, will be acquired for the enrichment of the ADR ontology in the near future.