We seek to leverage enhanced expressivity in OWL 1.1 via property chain axioms with right identities in order to organize and constrain anatomic concepts for use in clinical descriptions. Anatomic knowledge represented in SNOMED CT uses SEP triplets; we anticipate that property chains will allow a more parsimonious organization of anatomic concepts. However, these constructs may lead to unanticipated inference, especially when scaling to large numbers of concepts [1]. We used a bottom-up approach based on targeted use case questions to iteratively develop a “micro theory” that both identifies the sensible locations of fractures in long bones and also supports logic-based classification of fractures. Alternative representations of the statement “fractures occur in bone” were explored with the aim of creating rich clinical descriptors that support classification for inference and data mining. The process of creating this micro theory is discussed, where pragmatic decisions were made with an intention of both constraining data entry and enabling inferences within the scope of the use cases.
OWL and other forms of description logics have been
used extensively to model spatial relationships for
anatomical knowledge [
For our initial investigation, we focused on a single
use case limited to fractures of long bones. We
adopted an iterative bottom-up process to developing
a “micro-theory”—an axiomitization that yields
sensible and logically correct inference in a limited
domain. At each stage, we tested the incremental
theory against the use case scenario. We re-used
content from the Foundational Model of Anatomy
(FMA) [
Clinical Scenario and Use Case Questions Typically, a physician creates a clinical descriptor that is of sufficient granularity to support a management plan—the clinical descriptor is an index for the general management plan for a given pathology. Within the contemporary electronic health record, the clinical descriptor may be reused as data to drive point-of-care decision support, or as warehouse data to support reporting. For instance, if we need to report the number of patients who had a fracture of the proximal femur, we should include the number of patients who had a fracture of the femoral neck. In both cases, the original descriptor should support detailed classification schemes.
With respect to bone fractures, it is desirable to
describe fractures in detail with respect to the bone
features involved—the clinical detail drives the
management plan. The clinical detail may describe
either a fracture involving an anatomic landmark or a
functional region where all fractures act similarly. It
is equally important that the clinical descriptor not
admit any nonsensical description. While fractures
may involve bony landmarks, we generally do not
describe fractures of the periosteum—the bone
lining—or the bone marrow. While these are parts of
bones, they are not generally parts through which
fractures are described to occur. The GALEN project
used constraints called sanctions to specify the values
that could sensibly be applied to relations such as
has-location [
Humerus are covered by Articular Cartilage? 3. Is a fracture of the Femoral Neck also a fracture of the Proximal Femur (i.e., is a fracture through an anatomic feature a fracture of a functional region)? 4. Is a fracture of the Trochlea a fracture of the
Distal Epiphysis of the Humerus? 5. Is a fracture of the Trochlea an intra-articular fracture? 6. Is a fracture of the Trochlea an intra-articular fracture of the Distal Epiphysis of the Humerus?
We looked at the following two sources for creating the fracture ontology: (a) The SNOMED CT hierarchies spanning the femur and the humerus; and (b) The FMA hierarchies corresponding to the femur and the humerus. A brief description of the portion of these two knowledge sources is described below.
SEP-triplets are extensively employed in the anatomical part of SNOMED CT. For each SNOMED anatomical class representing one entire entity, called entity (or entire) class (E-class), there are two auxiliary classes, the structure class (S-class) and the part class (P-class). For example, in the femur † 1) The medial and lateral condyles of the Femur; 2) Trochlea and Capitellum; 3) Yes; 4) Yes; 5) Yes; 6) Yes. hierarchy, we ideally would have the following classes defined: StructureOfFemur EntireFemur Ն StructureOfFemur FemurPart Ն StructureOfFemur ՈӃ partOf.EntireFemur BoneStructureOfDistalFemur Ն FemurPart EntireDistalFemur Ն BoneStructureOfDistalFemur DistalFemurPart Ն BoneStructureOfDistalFemur Ո Ӄ partOf.EntireDistalFemur StructureOfDistalEpiphysisOfFemur Ն DistalFemurPart EntireDistalEpiphysisOfFemur
Ն StructureOfDistalEpiphysisOfFemur The E-class is instantiated by entire anatomical objects (such as the entire femur), and the P-class by the proper parts of the referred objects (such as the distal femur). The S-class, finally, is instantiated by instances that are either entire objects or their parts. This definition explains the is-a links from the Eclass and the P-class to the S-class, as well as the partOf link from the P-class to the E-class. The main idea underlying the SEP-triplet approach is to represent a part-whole relationship between two entity classes not by a part-of link between the Eclasses, but rather by an is-a link between the S-class of the “part” and the P-class of the “whole”. This is, however, sufficient to simulate transitivity of part-of through the inherently transitive relation is-a: EntireDistalEpiphysisOfFemur Ն StructureOfDistalEpiphysisOfFemur Ն DistalFemurPart Ն BoneStructureOfDistalFemur Ն FemurPart Ն partOf.EntireFemur This allows us to conclude that every Distal Epiphysis of the Femur is part of some Femur. Since characteristics are inherited along the is-a hierarchy, the SEP-triplet encoding also allows us to simulate inheritance of characteristics along the part-of hierarchy. In our example, by connecting a fracture via the findingSite property to the S-class, we can ensure that a fracture located in the Distal Epiphysis of the Femur is classified as a fracture located in the Femur. Another advantage of the SEP encoding is that one can suppress such inheritance along the partof hierarchy by connecting via findingSite to the Eclass.
There are, however, several problems with the
SEPtriplet encoding. First, from a formal ontological
point of view, it partially conflates the is-a hierarchy
with the part-of hierarchy, which may lead to
unintended consequences since the two relationships
are completely different by nature [
The FMA ontology defines a set of partonomic
relationships discussed in [
A constitutional part is defined as a primary partition of an anatomical structure into its compositionally distinct anatomical elements. In the context of the whole, an element is any relatively simple component of which a larger, more complex anatomical structure is compounded; i.e., the partition is compositional rather than spatial. For example, a stomach may be viewed as being partitioned into its wall and cavity. A regional part on the other hand is defined as a primary partition that spatially subdivides an anatomical structure into sets of diverse constitutional parts that share a given location within the whole; i.e., the partition is spatial rather than compositional. For example, a stomach may be viewed as being partitioned into its fundus, body and pyloric antrum to name a few of such parts. Constitutional parts are genetically determined, whereas regional parts are defined not only by genetically regulated developmental processes (e.g., lobe of lung, cortex of kidney, finger), but also by arbitrary landmarks or coordinates, such as used for demarcating the thoracic and abdominal parts of the aorta and the fundus of the stomach from adjacent parts of the corresponding wholes. A systemic part is defined as a secondary partition of an anatomic structure in accord with functional systems.
The distinction between regional parts determined by well defined genetically regulated processes and arbitrary landmarks and coordinates, is represented by associating the attributes anatomical or arbitrary with regional parts. Furthermore, these attributes provide the basis for the different views of regional partitions, as in the case of the liver, where its traditional partition into lobes based on arbitrary landmarks constitutes an arbitrary kind of regional view, while another partition based on the distribution of the tributaries of the hepatic veins or branches of the hepatic artery constitutes an anatomical regional view.
The FMA also supports topologic relationships
supporting connectedness and containment.
Connectedness describes whether structures are
continuous with, attached to, or synapsed with other
structures. Containment deals exclusively with the
containment of a material anatomic entity within an
anatomic space, e.g., Right lung -contained in- Right
half of thoracic cavity. Connectedness and
containment are orthogonal to regionality and
constitutionality and do not confer parthood [
We now present our approach to developing the long bone fracture ontology. We draw on the FMA as a primary source of anatomic content.
Regional vs. Constitutional Partitions As previously discussed, the FMA ontology draws a distinction between a regional partition and a constitutional partition. We reviewed this content to determine whether it was suitable for reuse within our ontology fragment. 1. The regional partition of long bones is exemplified by the following regional parts of the Femur (regPartOf): ProximalEpiphysisOfFemur ՆӃ regPartOf.Femur DiaphysisOfFemur ՆӃ regPartOf.Femur DistalEpiphysisOfFemur ՆӃ regPartOf.Femur FemoralNeckOfFemur ՆӃ regPartOf.ProximalEndOfFemur
Regional parts of the femur include true anatomic parts (epiphyses, diaphysis) as well as functional parts defined by fiat boundaries (proximal end of femur), illustrating the FMA’s anatomic and arbitrary types. 2. The constitutional partition of long bones is exemplified by the following constitutional parts of the Femur (constPartOf): BonyPartOfFemur 2I)HPX
ӃՆFRQVW3DU BoneOfFemur 2I%\)HPX
ӃՆFRQVW3DU PeriosterumOfFemur Ն MedullaryCavityOfFemur Ն
FӃRQVW3DU2I%\)HPX FӃRQVW3DU2I%\)HPX FӃRQVW3DU2I%\)HPX FӃRQVW3DU2I)HPX
FӃRQVW3DU2I)HPX VasculatureOfBonyPartOfFemur Ն ArticularCartilageOfDistalEpiphyisOfFemur Ն ArticularCartilageProximalEpiphysisOfFemur Ն VasculatureOfFemur 2I)HPX
ӃՆFRQVW3DU CavityOfFemur 2I)HPX
ӃՆFRQVW3DU The constitutional parts of the femur include the multiple tissue types that combine to form a long bone—the bone proper, the articular cartilage, etc. Note the bone proper also decomposes to include the bone material itself, the periosteum, and the medullary cavity.
The regional partition includes the structures where clinicians locate fractures and the relationships between these structures.
The constituents of long bone such as the periosteum, where fractures are not described to occur, are conveniently sequestered in the constitutional partition.
We adopted the relevant portions of the regional partition for use in our model. However, we adopted a simpler representation for the incorporation of articular cartilage into the model for this initial iteration.
We now present some high-level classes and object properties that characterize the entities in which we are interested.
Bone LongBone Ն Bone Femur Ն Humerus Ն
LongBone
LongBone ObjectProperty(regionalPartOf) reflexive(regionalPartOf) transitive(regionalPartOf) BoneRegion Ն QDO3W2I
UӃHJLR ObjectProperty(findingSite) domain(findingSite) = Disorder Bone
Disorder Fracture Ն Disorder ՈӃILQG ingSite.BoneRegion The class Bone is effectively the class BoneOrgan in the FMA. Within this initial iteration, we are neutral regarding the alignment of Bone with the
hierarchy consisting of aligning
with the is-a CavitatedOrgan,
Organ, AnatomicalStructure, MaterialAnatomicalEntity, and AnatomicalEntity, as we did not see an impact of this in the context of the application at hand. The property findingSite aligns with the SNOMED CT relationship which assigns locations to clinical conditions We declare the property regionalPartOf to be reflexive, thereby inducing Bone to be a BoneRegion. This has the important effect of unifying the treatment of entire long bones and bony landmarks with respect to findingSite—fractures may be declared to occur equally within the entire bone or at the landmark. We declare regionalPartOf to be transitive to support the interrelationships between discrete landmarks, larger regions of bone, and the entire bone.
We add the following subclasses of BoneRegion into the model: AnatomicBoneRegion Ն%RQH5JL FunctionalBoneRegion Ն%RQH5JL In clinical practice, pathology may be attributed to a true anatomic entity or a functional entity where unique pathologies behave similarly, are responsive to similar treatments, are aggregated for epidemiologic purposes, etc.
example, several unique fractures all aggregate to fractures of the proximal femur. As previously noted, the FMA incorporates true anatomic regions and functional regions.
Propagation of Locative Relationships A key functionality that is required to support the use case questions discussed earlier is the ability to propagate the location of a fracture from a given region to all the regions to which it has regionalPartOf relationships. For instance, if a fracture is located in the femoral neck, it is also located in the proximal metaphysis of the femur as the femoral neck is a regional part of the proximal metaphysis of the femur. This is represented using the following axiom: findingSite ӓ regionalPartOf Ն findingSite
It may be noted that the transfer of locative relationships is also propagated transitively due to the transitive nature of regionalPartOf. regionalPartOf ӓ regionalPartOf Ն regionalPartOf
In order to explore articular fractures—the fracture of a bone region covered by articular cartilage, we incorporated the following concepts : ArticularCartilage ObjectProperty(coveredBy) ArticularBoneRegionԘ%RQH5JLՈ ArticularFracture Ԙ)UDFWXH Ո
J&G%\$WLXOD FӃRYHU IӃLQGJ6WH
ArticularBoneRegion This representation provides a simple method to distinguish between articular and non-articular bone regions.
Using the initial ontology, we were able to create a series of detailed clinical descriptions which classified as expected. Some examples are discussed next.
Locative Transfer over Regional Parts rA facture of the Femoral Neck is classified as a fracture of the Proximal Femur. (Since findingSite ӓ regionalPartOf Ն findingSite)
Fx
A fracture of the Femoral Neck is classified as a fracture example and begin with the following reformulation of FemoralNeck.
FemoralNeckFx
findingSite. Ӄ findingSite.(Ӄ regionalPartOf.ProximalEndOfFemur) (Ӄ regionalPartOf.( regionalPartOf.Femur)) (Since ProximalEndOfFemur ՆӃ regionalPartOf.Femur) findingSite.(Ӄ regionalPartOf.Femur) (Since regionalPartOf ӓ regionalPartOfՆ regionalPartOf)
findingSite.Femur (Since findingSite ӓ regionalPartOf Ն findingSite) The proof above indicates that we can represent direct relationships between bones and bone features and infer regional partonomy relationships between
Extending the model to describe and classify articular fractures is also accommodated by the model and creates no additional complications.
The articular parts of the distal epiphysis of the humerus—trochlea and capitellum—are created as articular regions, while the non-articular parts—the medial and lateral epicondyle—are created as regular bone regions. The only caveat to this approach is that partially-covered regions are not considered articular regions; the distal epiphysis of the humerus is not considered an articular bone region by this criterion.
Fractures of the parts are created in the usual fashion by restricting the fracture finding site. Trochlear and capitellar fractures classify appropriately as articular fractures.
General fractures of the distal humeral epiphysis and articular fractures are then created in the same way.
humeral epiphysis are classified as subclasses of general fractures of the distal humeral epiphysis; trochlear and capitellar fractures are classified as further subclasses.
classify correctly as general fractures only. We did not specifically try to define non-articular fractures (fractures of parts not covered by articular cartilage).
We note one failure of the model in the subset of bones we examined.
The FMA contains a regional part of the humerus, ‘Nutrient Foramen of Humerus’, literally, the hole where the nutrient artery enters the humerus. Because fractures are usually not described through this feature, this constitutes a failure of the model to constrain the set of sensible locations of fractures.
In the FMA, the nutrient foramen is a subclass of
We can certainly remediate the
model to additionally restrict bone regions to subclasses of Material Anatomic Entity. However, the appearance of the nutrient foramen as an arbitrary bone region bears further discussion.
We have begun to explore the creation of a ‘microtheory’ for long bone fractures—an axiomitization that yields both sensible and logically correct inference. Using a set of framing axioms in combination with content from the FMA regional partition, we were able to describe and correctly classify a rich set of fractures, while maintaining a fairly parsimonious ontology. The resulting ontology is quite constrained as compared to an SEP triple approach.
Although this initial model successfully fulfills the use case, the breach raises significant questions. Because the FMA regional model admits arbitrary regions, there is no principled reason why an arbitrary region of a bone can sensibly be a fracture location. Our success seems to be an empirical finding— further analysis is necessary to see whether the model hold across all bones, or can extend to parenchymal organs such as the lung or the liver—organs made of the same ‘stuff’.
Our model may succeed because bones exemplify a
‘stuff/whole’ partonomy, where arbitrary regions are
compositionally homogenous. This does suggest that
if regional parthood could be compositionally
restricted, this would offer a more convincing model.
One obvious possibility for implementing this
restriction is to utilize the GALEN partitive attribute
“hasSolidDivision”, or perhaps a similar attribute
with even more specialized meaning [
This work provides initial insight into creating safe and effective inference over property chains for the purpose of creating and classifying clinical descriptions. Because of the tremendous change management implications of incorporating new idioms into terminologies such as SNOMED CT, it is important that we demonstrate that such idioms are safe, effective, and scale. Continuing work will investigate constraining the regional idiom with respect to homogenous compositionality, expanding our analysis to a larger set portion of the skeletal system, and examining the generalizability of the idiom to additional organ systems.