Definition of qualifier classes used in clinlp

This page describes the definitions of qualifier classes for Dutch clinical text we use in clinlp.

Introduction

When processing clinical documents (e.g., doctor’s notes, discharge letters), detecting qualifiers (e.g., absent, historical, non-patient, increasing, decreasing, severe, light, etc.) follows the matching of concepts (e.g., diagnoses, symptoms, procedures). In clinlp we primarily use the term “qualifier”, although the terms “context”, “meta-annotation”, and “modifier” have also been used to denote the same concept.

A consensus on potential qualifier classes, along with clear definitions of the qualifiers they encompass, is needed to develop accurate algorithms for detecting them. Despite some shared intuitive understanding of recognizing, for instance, a negation in a sentence, there are numerous cases where intuition simply falls short. In practice this has impeded manual annotation of gold standards, used for training and evaluating algorithms. In turn the resulting annotations (with Kappa-Cohen as low as 0.5) lead to a difficult target for supervised machine learning models. A standardized classification of qualifiers, as proposed here, will hopefully advance both research and clinical implementation of NLP (Natural Language Processing) algorithms. This page is the result of some deliberation among various researchers and developers working with clinical text, but it is not necessarily definitive. We welcome feedback and suggestions for improvement.

For these definitions we took the Context Algorithm (Harkema, 2009) as a starting point, both because this is an influential paper, and because a corresponding Dutch corpus is available (Afzal, 2014). There are already some trained models available that can for a large part be re-used. We will here describe three qualifier classes: Presence, Temporality, and Experiencer, including some definitions, issues to resolve, and illustrative examples. These classes can be further de-aggregated at a later stage, and other classes may follow later as well. Note that choosing qualifiers is a trade-off between granularity and practicality. We aim for a balance that is useful for most clinical NLP tasks.

Qualifier classes are denoted by boldface, with the Qualifier (a mutually exclusive value a qualifier class can assume) formatted as inline code.

Presence

Absent / Negated	Uncertain	Present / Affirmed (default)
Concepts that are explicitly or implicitly described as absent in a patient	Whether the concept was absent or present is not asserted with very high certainty	Concepts that are explicitly or implicitly described as present in a patient

Assessing whether some concept was present or absent is one of the most important parts of a clinician’s job. Whether something is present or occurred in the real world is knowable in principle, but in the clinical world, such assertions are rarely made with complete certainty. This is already implied by the uncertainty at the core of the clinical reasoning process, but in clinical text the uncertainty is often made explicit by means of hedging. It’s therefore important to note that when we are extracting concepts from medical text, it’s very hard to make direct assertions about the real world, but we are limited to recognizing probability statements made by clinicians.

The presence class therefore captures whether a concept is present in three qualifiers. The Present and Absent qualifiers are used when the clinician assesses a concept as being present (or absent) with very high probability, extending beyond reasonable doubt. When neither presence nor absence is definitively asserted, the Uncertain qualifier applies. This qualifier therefore ranges from very unlikely to very likely.

The default qualifier for presence is Present. When the text does not indicate absence or uncertainty of a concept, we assume the writer intended to convey its presence.

To resolve

• In future work, the uncertain qualifier may be further split up, for instance, into a negative uncertain (i.e., unlikely) qualifier and a positive uncertain (i.e., likely) qualifier. Or perhaps an ‘uncertain uncertain’ qualifier in addition to those, for 50/50 cases.

• The exact threshold for absent and present should be further defined. What probability cutoff should be regarded as ‘beyond a reasonable doubt’? We can set this threshold at two standard deviations (<0.025, >0.975), but it would be even better to do some small empiric study with clinicians, to find out where each trigger term should go. Consider for example edge cases such as: very insignificant, very likely, subclinical, etc.

Examples

Example	Qualifier
Rechtszijdig fraai de middelste neusgang te visualiseren, vrij van poliepen.	Absent
Tractus circulatorius: geen pijn op de borst.	Absent
Een acuut coronair syndroom werd uitgesloten.	Absent
Werkdiagnose maagklachten bij diclofenac gebruik en weinig intake.	Uncertain
Waarschijnlijk hematurie bij reeds gepasseerde niersteen.	Uncertain
Dat er toen bradypacing is geweest valt niet uit te sluiten.	Uncertain
In juni 2023 longembolie waarvoor rivaroxaban met nu asymptomatische progressie.	Present
PTSS en recidiverende depressie in VG.	Present
Status na mild delier, heden wel slaperig.	Present

Temporality

Historical	Current (default)	Future
Concepts that were applicable at some point in history, but not in the last two weeks.	Concepts that were applicable in the last two weeks (potentially starting before that) up to and including the present moment.	Concepts that are potentially applicable in a future scenario.

The temporality class places concepts in a temporal framework, ranging from past to future, relative to the document date. The Historical and Current qualifiers distinguish between concepts that were applicable in the past, versus concepts that are applicable in the present. The exact cutoff between Historical and Current is problem-specific and therefore hard to definitively establish in a general sense. In a discharge summary, everything that happened before the admission period could be considered Historical, which can easily range up to months, while during a GP (General Practitioner) visit, events from a few days prior might be considered Historical. For the general case, we see no reason to divert from the threshold of two weeks in the original Context paper (Harkema et al., 2009). Note that the Current qualifier also applies when the concept is applicable in the last two weeks, but already started before that.

The Future qualifier is applicable when a concept is described in a future scenario, for instance when describing the risk of developing a condition at a later stage, or when describing a procedure that will take place later.

To resolve

• A way to dynamically define the threshold for Historical and Current, so that a cutoff can be established for each problem. In future work, we might map each concept to a timedelta (e.g., -1 year, -14 days, +5 days), but that does not fit the current qualifier framework very well. Also, it seems quite a hard problem.

Examples

Example	Qualifier
Zwanger, meerdere miskramen in de voorgeschiedenis.	Historical
Progressieve autonome functiestoornissen bij eerdere dermoidcyste.	Historical
Als tiener een osteotomiecorrectie beiderzijds gehad.	Historical
Echocardiografisch zagen wij geen aanwijzingen voor een hypertrofe cardiomyopathie.	Current
Al langer bestaande bloeddrukproblematiek.	Current
CT thorax: laesie rechter onderkwab bevestigd.	Current
Conservatieve maatregelen ter preventie van pulmonale infectie zijn herbesproken.	Future
Mocht hij koorts en/of tachycardie ontwikkelen, dan contact opnemen met dienstdoende arts.	Future
Wordt nu opgenomen middels IBS ter afwending van suïcide.	Future

Experiencer

Patient (default)	Family	Other
Concepts applicable to the patient related to the current document.	Concepts not applicable to the patient, but to someone with a genetic relationship to the patient.	Concepts not applicable to the patient, but to someone without a genetic relationship to the patient.

The experiencer qualifier distinguishes between concepts that apply to the Patient, to those that apply to Family members with a genetic relationship to the patient, and Other individuals with no genetic relationship to the patient (e.g. acquaintances). Clinical documents are typically obtained from electronic health records, where the relation between a document and a patient is explicit. Since a patient is a well separated entity, there is usually little ambiguity which class applies. If a concept applies to both the patient and another person, the patient label should be selected.

Examples

Example	Qualifier
Behandeling in WKZ ivm diabetes beeindigd.	Patient
Pte wil geen medicatie tegen parkinson ivm slechte ervaringen broer	Patient
X-enkel rechts: schuine fractuur laterale malleolus	Patient
Familieanamnese omvat: ADD/ADHD: broer	Family
Moederszijde: voor zover bekend geen kanker	Family
2. Covid positieve huisgenoot	Other

References

• Afzal, Z., Pons, E., Kang, N. et al. ContextD: an algorithm to identify contextual properties of medical terms in a Dutch clinical corpus. BMC Bioinformatics 15, 373 (2014). https://doi.org/10.1186/s12859-014-0373-3

• Harkema H, Dowling JN, Thornblade T, Chapman WW. ConText: an algorithm for determining negation, experiencer, and temporal status from clinical reports. J Biomed Inform. 2009 Oct;42(5):839-51. doi: 10.1016/j.jbi.2009.05.002. Epub 2009 May 10. PMID: 19435614; PMCID: PMC2757457.