A Neuro-ontology for the neurological examination

There is no standard neurological examination and history [11‐13]. Depending on the examiner and patient, a neurological examination and history may take as little as three minutes or greater than 30 min to complete. Furthermore, there is no generally accepted format for recording the neurological examination and history. Some neurologists use an outline, some use a table, and others use a narrative. No agreed-upon terminology exists for recording the neurological examination and history. During an examination, a neurologist elicits symptoms such as weakness, slowness, memory loss, speech impairment, blurred vision, numbness, tingling, pain, or imbalance and abnormalities in the mental status, cranial nerves, motor system, sensory system, reflexes, coordination, and gait. Due to a lack of standard terminology, identical neurological abnormalities may be described variously. For example, a failure to abduct the eye may be variously recorded as a sixth nerve palsy, an abducens palsy, or an abducens nerve weakness. Similarly, an upgoing great toe upon stimulation of the plantar surface of the foot may be recorded variously as Babinski sign, upgoing toe, positive Babinski, or upgoing plantar response.

If neurological findings are to be converted into machine-readable codes, an ontology is needed [14]. At a minimum, an ontology for the neurological examination and history that supports big data applications should do the following:

Some ontologies have other useful characteristics. Many ontologies attach a definition to each concept. Other ontologies, SNOMED CT in particular, allow simple concepts to be combined to form more complex concepts. For example, the concept |ankle reflex| can be combined with |absent| and is equivalent to |absent ankle reflex|. This is known as post-coordination [15]. Concept ontologies that are organized hierarchically support the calculation of inter-concept distances [16‐24].

The UMLS Metathesaurus is not an ontology per se [25, 26]. The 2019 AB release is a curated compendium of 155 distinct terminology sources with 4,258,810 concepts. Each concept is linked to a unique machine-readable code and a specified name. For example, the concept aphasia has the CUI (concept unique identifier) C0003537. Each CUI is eight characters in length and starts with the letter C, followed by seven digits. The concept aphasia is contributed 58 times to UMLS from 58 different source terminologies. Each contribution of aphasia appears in the UMLS with the same CUI but a different atom unique identifier (AUI). The UMLS maintains concept synonyms as normalized concept names. The 2019AB release has 11,882,429 normalized concept names, each with a unique LUI. For example, the concept aphasia has 22 English language synonyms in UMLS, including |loss of words|, |losing words|, |loss of power of expression or comprehension|, |aphasic syndrome|, and |difficulty finding words|. The UMLS Metathesaurus also maintains files with concept definitions, files with relationships between concepts (child-parent, etc.), and files with ontology hierarchies (paths from each concept to the root concept).

We abstracted the neurological examinations from 419 published neurological case studies [27‐33]. Based on the neurological findings, we created a dataset of 2386 test-phrases for encoding as concepts by an ontology. Normal findings were ignored.

Test phrases were mapped to UMLS concepts using the UMLS browser. Except for a few concepts, we ignored laterality (e.g., right ataxia, left ataxia, and bilateral ataxia were all mapped to the UMLS concept ataxia |C0004134|. We mapped the 2386 test phrases to 601 concepts. Concepts recurred a mean of 3.9 ± 5.5 times (range 1 to 47 times) in the 419 test teaching cases. The 15 most common neurology concepts are shown in Table 1.

We reviewed the neurological history and examination as presented in three standard textbooks [11‐13] and identified 1100 findings, which were either signs or symptoms. Each of the 1100 concepts was entered into the Protégé ontology editor [34] and consisted of the following:

We used Protégé to organize the concepts into a mono-hierarchic ontology. The average number of children per concept was 3, and the maximum ontology depth was seven levels. The neuro-ontology is downloadable as a CSV or OWL file at the BioPortal of the National Center for Biomedical Ontology [35]. The neuro-ontology has five high-level branches to mirror the structure of the neurological examination (mental status finding, cranial nerve finding, motor finding, sensory finding, reflex finding) and four additional high-level branches (head finding, neck finding, skin finding, and neurological symptoms). Figure 1 demonstrates a partial expansion of NEO. Table 2 shows the distribution of concepts by the examination section. We had two considerations in building the concept hierarchy for NEO. First, findings related to each part of the neurological examination should be gathered under one consistent heading. Second, the concept hierarchy path distance between findings should reflect a neurologist’s thinking about which concepts are more similar while preserving subsumption. For example, a brisk biceps reflex could be subsumed under either biceps reflex or brisk reflex (see Fig. 2). We chose to subsume brisk biceps reflex under brisk reflex because a brisk biceps reflex is more similar to a brisk knee reflex than it is to an absent biceps reflex.

We identified five component ontologies of the UMLS Metathesaurus (Table 3) as a potential basis for a neurological examination ontology: SNOMED CT, HPO (Human Phenotype Ontology), MEDCIN, MeSH (Medical Subject Headings), and OMIM (Online Mendelian Inheritance in Man). We tested how well each of these ontologies provided concept coverage for the 601 neurology concepts identified above.

We calculated distances between concepts for the SNOMED CT and NEO ontologies by the method of Wu and Palmer [24]. Distances between concepts were normalized with a minimum of 0.0 (closest) and a maximum of 1.0 (most distant).

Dist (a, b) is the semantic distance between concept a and concept b, LCS is the lowest common subsumer in the ontology for both a and b, depth(a) is number of levels from the root concept to concept a, depth (b) is the number of levels from the root concept to concept b, and depth(LCS) is the number of levels from the root concept to the LCS.

One of the advantages of a small domain-specific ontology, such as NEO is that domain experts can more easily identify and correct subsumption errors. Some somewhat arbitrary subsumption decisions may influence how accurately the concept distance measures derived from a concept hierarchy align with expert opinion. Considerable effort has been devoted to finding the best distance metrics for concept hierarchies that give the closest results to expert opinion [19, 20, 22]. However, when the hierarchy itself is not aligned with expert opinion, the choice of distance metric may be less critical. For example, most neurologists would agree that an absent knee jerk is more similar to an absent biceps jerk than it is to a brisk knee jerk (Fig. 2). As a result, in NEO, we have subsumed absent ankle reflex, absent knee reflex, absent biceps reflex, and absent triceps reflex under absent reflex. SNOMED CT subsumes absent ankle reflex and brisk ankle reflex under ankle reflex finding and subsumes absent biceps reflex and brisk biceps reflex under biceps reflex finding. These are ontologically correct (Fig. 2) but yield distance measures that are not in accord with a neurologist’s opinion as to which concepts are more similar (Table 8). In a large ontology like SNOMED CT, some errors in subsumption are inevitable and need to be corrected over time based on input from domain experts [40, 41]. Two examples of likely errors in subsumption are noted in Figs. 3 and 4. Dysmetria is a form of ataxia and hence is correctly subsumed by a finding related to incoordination, but dysmetria is not a reflex finding (Fig. 3). Similarly, apraxia is a disorder of higher cortical functions akin to aphasia and agnosia and should not be subsumed under either incoordination or musculoskeletal disorders (Fig. 4). As a result, distance measures based on the SNOMED CT concept hierarchy show dysmetria as too close to absent reflexes and shows apraxia too close to ataxia (Table 8). One of the goals of SNOMED CT is to have more concepts fully defined, which is achieved by adding qualifiers to concepts to make them distinguishable from other concepts in the ontology. Fully defined concepts cannot be confused logically with any other concept in the ontology. However, adding additional subsumption relationships may yield anomalous distance measures between unrelated concepts. In Fig. 5, the subsumption of asterixis (a flapping tremor of the arm) under a finding of the upper limb yields an anomalous distance measure placing asterixis too close to hemiparesis (Table 8). Similarly, in Fig. 6, the subsumption of oral dyskinesia (adventitious spontaneous chewing and writhing movements of the lips and tongue) under oral cavity finding causes oral dyskinesia to be anomalously close to the unrelated concept of vocal cord paralysis (Table 8). Finally, in a small curated domain-specific ontology like NEO, related concepts can be grouped to provide more accurate concept distance measures. For example, by grouping manifestations of bradykinesia such as masked-facies and micrographia close to bradykinesia, distance measures calculated based on the concept hierarchy can reflect the relatedness of these concepts (Table 8).

One advantage of a small ontology like NEO is improved manageability. Both the NLM and SNOMED International recommend the use of subsets of UMLS and SNOMED for certain restricted applications [26, 43]. NEO has only 1100 concepts, and the downloadable CSV file has 1100 rows. In contrast, the UMLS 2019AB release has 4,258,810 concepts, the MRREL.RRF (relationships) file has 84,189,164 rows and the MRCONSO.RRF (concepts) file has 15,172,405 rows. Similarly, the 2019 SNOMED CT US Edition release has 357,533 concepts, the relationship snapshot file has 2,989,896 rows, and the concept snapshot file has 478,117 rows. Since each row in the NEO file has the child-parent relationship for each concept, hierarchies and subsumption tables can be generated directly from the primary file.

A major limitation of this study is that we did not attempt to combine concepts to generate missing or more complex concepts (post-coordination). Because we did not post-coordinate concepts, we were not able to grade muscle weakness when describing the motor examination. Muscle strength is usually graded on a scale of 0 to 5 (0 = ‘no movement’ to 5 = ‘full strength’). We elected to encode the clinical expression |weak quadriceps (4/5)| as |quadriceps weakness| C0577655 and ignore the degree of weakness. The sensory examination is notoriously difficult to describe in text, and in the era of the paper medical record, neurologists often used drawings to document the sensory examination. For this study, we did not use post-coordinated concepts to describe sensory findings but used instead granular concepts from MEDCIN to describe common sensory findings such as sensory loss in median nerve distribution, sensory loss in C8 dermatome distribution, a sensory level at the T2 level, etc. Similarly, we ignored quantitative visual acuity measurements such as 20/200 and encoded any visual acuity of less than normal as |reduced visual acuity| C0234632. Although UMLS has appropriate lateralized concepts for some findings (hemiparesis, rigidity, myoclonus, chorea, tremor, ptosis), there are not separate lateralized concepts for other common neurological findings (ataxia, hyperreflexia, spasticity, hyporeflexia). We did not develop a method to capture these lateralized findings when existing concepts were not available in UMLS and reverted to using the non-lateralized concept (e.g., ataxia for left ataxia).

Another limitation is that we have not validated the phrase abstraction methods or the phrase-to-concept mapping methods with other neurology experts or tested the methodology on de-identified medical records. More work is needed on whether a concise neurological examination ontology such as NEO is useful and acceptable to neurologists.