Although the medication given to a patient during a hospital stay is fundamental for patient safety, quality assessment and economic aspects of health care services, in-patient drug prescriptions are still commonly done on paper in many institutions. In Europe and North America most of the pharmacies are computerized. This does not guarantee, however, that drug prescription information is always available in hospital information systems in a structured way. As long as the only computer-readable sources of medication information are clinical narratives such as dictated or typed discharge or outpatient letters, analysis of drug information depends on natural language processing (NLP). The performance of NLP critically depends on annotated clinical corpora. We are currently developing an annotation schema for mentions of drugs in discharge letters, of which ontological foundations are presented.
Manually annotated texts are key resources for information extraction systems. However, if the underlying annotation schema does not offer clear-cut criteria, a high agreement between human annotators is hard to achieve. The avoidance of ambiguity is therefore a prerequisite for precise human and machine annotations. We believe that definitions based on formal ontologies are helpful for the provision of precise annotation criteria.
The documentation of drugs given to a patient during a hospital stay is fundamental for patient safety, quality assessment and economic aspects of health care services. Nevertheless drug prescriptions are still commonly done on paper in many institutions. In Europe and North America most of the pharmacies are computerized, but this does not guarantee that prescription information is always available in hospital information systems in a structured format.
The primary purpose of clinical narratives in electronic health records is communication between clinicians. These texts are written to be naturally understandable, ideally across specialties. The mention of drugs in a clinical document is, however, not limited to the prescription section. Drug names can also be found as mentions of drug allergies, suspected triggers of side effects, drugs discussed but not prescribed, drugs suspended etc.
The mention of drugs names alone is not yet sufficient. It is often be supplemented by
strength, route, dose form and regimen, and it is embedded in a context, such as
affirmation, negation, or intentionality. In this paper, we focus only on text passages
that refer to drugs in a strict sense, not on the text that provides additional context,
although our future goal is to provide a full picture of drug mentions in clinical
documents. A fully-fledged ontology should also formalise drug and product families,
categorise names and their parts, introduce the notions of branded and packaged drugs
and, finally, provide formalisations of signature information, including drug strengths,
dose forms and dosing regimens, routes and temporal information. Finally such an
ontology should be compliant with existing and emerging standards, such as ISO
IDMP[
First analyses of clinical summary samples were done to shape the ontological
definitions, in parallel with the bottom-up formulation of a drug annotation guideline.
In our text samples, most textual references to drugs were found within the prescription
section of clinical summaries. It is expected that the patient’s GPs base their
medication decisions on these recommendations, and the assumption is reasonable that
the drugs mentioned in the prescription section were also administered during the
hospital stay and, possibly, even before hospitalisation. It is a known problem in our
setting (an Austrian university hospital) that drug therapies initiated and finished during
an inpatient stay are often not represented in computer-readable EHR documents.
We analyse drug mentions in clinical documents from both an ontology and language
engineering point of view. Formal ontologies state what is universally true in a domain,
using logical axioms. Upper-level ontologies guide the ontology engineering process
by supporting ontology engineers with a consistent framework, which can foster
semantic interoperability if used by different domain ontologies. Our drug model is
aligned with two existing ontologies, viz. the domain upper-level ontology BTL2 [
The basis of our analysis is a corpus of pseudonymised discharge summaries filtered by the diagnosis Malignant Melanoma (ICD-10:C43), counting 400 documents. The annotation workflow was defined as follows: mentions of drugs, signatures, etc., were read and annotated by a medical doctor. Subsequently, first 200 of these were re-read and checked for annotation errors by the same annotator. Afterwards, the contained annotations were compared to machine-generated annotations that were created using a drug dictionary. All conflicts of annotations were checked by the human annotator and mistakes detected during the crosscheck were manually corrected by the same person. The aim of the work was to identify semantic categories (classes), which upon annotation can be related to text tokens in the clinical narrative. The purpose is to fill a structured dataset in a way that all the relevant parameters of a prescription are represented. The following priorities, or guidelines, steered the bottom-up evolution of the annotation schema: • If a problem in the schema exists and requires a change of the model, the change that results in the lowest complexity is preferred. • Any change, especially if it increases complexity, must be assessed by its practical usefulness. • The goal is to provide a semantic structure that is simple, yet precise and potentially applicable of narratives from other sources.
Based on the annotation experience, we created a model of semantic annotation. It is
centred on Drug annotation, composed of three semantic categories: DrugProduct,
DrugSubstance and DrugFamily. Drug is defined as a “chemical substance used in the
treatment, cure, prevention, or diagnosis of disease or used to otherwise enhance
physical or mental well-being”[
In the following OWL DL [
Specific drug substances can be hierarchically ordered. For instance, Penicillin G is a
In our ontology approach, which does not allow for meta-categories, DrugFamily does not appear as a separate category, but just as a way to flag those nodes in the drug substance hierarchy that are not terminal. One way to conceive drug families is to express them as information object classes, with substance names as instances. These substance names then extend to ontology classes in the substance hierarchy. The mention of just drug categories would be incomplete in the prescription section, but it is rather common in other parts, e.g. in passages like “treated with low dose nonselective NSAIDs under PPI protection”, or "we kindly ask a specialist to optimize the cardiac medication".
Ontologically, DrugSubstance and DrugProduct are strictly different. In SNOMED CT, substance concepts extend to amounts of pure or mixed substances, whereas drug product concepts extend to countable entities like pills, which include active and inactive ingredients with a defined shape, strength and often even manufacturer. In our formalisation, following SNOMED CT, there is a logical implication between a drug product and one or more substances. If this information is complete, then a drug product always allows to infer its active ingredients but not the other way around: sct:Diclofenac(product) subclassOf
sct:hasActiveIngredient sct:Diclofenac(substance)
In the following, the term “drug” is used for branded products, active ingredients and drug families. The analysis showed where these are mentioned within the narratives. • • • • •
Diagnosis section. High impact drugs are mentioned together with the diagnoses. Typically, past and present cycles of chemotherapy are mentioned here, as well as the status of oral anticoagulation, etc.
Radiology section. Injected contrast agents are usually mentioned.
Lab result section. Occasionally, substances are mentioned in the lab results as entry points for therapeutic drug monitoring (TDM).
Prescription section. Commonly, most medications in a clinical summary are mentioned here in a semi-structured form, mostly as branded drugs with signature, occasionally also as recipes for drug compounding.
Across the document. Elaborations regarding medications to be changed, indications thereof, as well as medications like local anaesthetics that were directly applied are mentioned in several parts, mainly the last section of the document.
From an ontology point of view we now scrutinize the textual source, its denoting entities and especially the way of referencing to reality. Medical texts, as well as parts thereof are information objects in the sense of BTL2 (a subclass of ‘generically dependent continuant’ in BFO).
subclasses of btl2:Action:
btl2:hasAgent some Human and btl2:hasPatient some DrugProduct
As mentioned, the fact that there is a drug prescription issued at the end of a hospital stay implies that the drug was most probably administered during the stay. A drug prescription could therefore also be read as a (highly probable) drug administration.
OR • • • • • • • Let us assume e is a denoting entity within a text t, regarding prescription of drug D. The double meaning can then be expressed as follows: e Type (DenotingEntity and btl2:hasPart some CertainInformation and btl2:represents only MedicationPrescription and btl2:hasRealization only (MedicationAdministration and btl2:hasPatient some D) e Type (DenotingEntity and btl2:hasPart some ProbableInformation and btl2:represents only (MedicationAdministration and btl2:hasPatient some D)) We now turn briefly to the annotation process and the elaboration of the annotation schema. The annotation has to anticipate that the denoting entity does not always refer to a DrugProduct class. In this case the annotation is done with a subclass of DrugSubstance at the leaf node level as well as at a more generic (DrugFamily) level. The fact that drug mentions in clinical summaries are result from human writers or dictated and subsequently voice-recognized or human-typed, several challenges arise:
Nominal anaphora: a drug product is mentioned in some part of text; the same referent is mentioned by a more general expression in the following text Ellipsis: omission of information, such as strength, where it is already known, obvious or irrelevant Convenience mixtures of names of branded drugs and ingredients
Detailed, unstructured, free-text instructions
The core annotation schema provided in Table 1 poses several challenges for the completion of the drug prescription ontology. In the present abstract, we intentionally omit other parts of the schema that deliver specific temporal, logical and signature information to a given drug mention within a narrative. Finally, the representation of conditions that act as decision points must extend the scope to observables and clinical findings. The necessary formalizations would probably go beyond what can be represented by OWL-DL description logics. The same might be true with the interpretation of drug mentions in the context of therapeutic drug monitoring, which has been, so far, excluded from our annotation schema.
Boundary issues may arise like the therapy with submolecular particles (e.g. gamma knife, radiotherapy, and phototherapy). Its characterisation as drug therapy might be contentious. Similar considerations apply to functional foods and special diets as a category between drugs and foods.