COBE: A Conjunctive Ontology Browser and Explorer for Visualizing SNOMED CT Fragments

Abstract

Ontology search interfaces can benefit from the latest information retrieval advances. This paper introduces a Conjunctive Ontology Browser and Explorer (COBE) for searching and exploring SNOMED CT concepts and visualizing SNOMED CT fragments. COBE combines navigational exploration (NE) with direct lookup (DL) as two complementary modes for finding specific SNOMED CT concepts. The NE mode allows a user to interactively and incrementally narrow down (hence conjunctive) the search space by adding word stems, one at a time. Such word stems serve as attribute constraints, or “attributes” in Formal Concept Analysis, which allows the user to navigate to specific SNOMED CT concept clusters. The DL mode represents the common search mechanism by using a collection of key words, as well as concept identifiers. With respect to the DL mode, evaluation against manually created reference standard showed that COBE attains an example-based precision of 0.958, recall of 0.917, and F₁ measure of 0.875. With respect to the NE mode, COBE leverages 28,371 concepts in non-lattice fragments to construct the stem cloud. With merely 9.37% of the total SNOMED CT stem cloud, our navigational exploration mode covers 98.97% of the entire concept collection.

Introduction

Search and browsing interfaces are an integral part of ontological system dissemination. Although ontology search interfaces are distinct from general search interfaces because of the semantic and structural information already contained in the ontologies, they can still benefit from the latest developments in information retrieval. For example, conjunctive exploratory navigation interfaces CENI and SCENI [1, 2, 3] have been developed for exploring consumer health questions with health topics as dynamically search tags complementing keyword-based lookup. The conjunctive exploration mechanism allows users to quickly narrow down to the most relevant results in the most effective way.

In this paper, we introduce a conjunctive ontology browser and explorer (COBE) for searching SNOMED CT concepts and visualizing SNOMED CT fragments. The COBE search interface has three prominent features. First, it supports both direct lookup and navigational exploration to retrieve concepts. In the direct lookup mode, a user comes with specific terms or SNOMED CT identifier, enters into the search interface, and retrieves a list of relevant concepts. In the navigational exploration mode, a user may not have a targeted term, or cannot easily or effectively formulate descriptive lookup terms, and may rely on navigational features to browse and explore the concepts. COBE provides conjunctive search combining direct lookup with navigational exploration. Second, COBE meets the need of a concept retrieval interface to reach and visualize erroneous SNOMEC CT fragments for quality assurance. Third, in the navigational exploration mode, COBE utilizes a cloud of core stems mined from non-lattice fragments. In previous study [4], mining SNOMED CT non-lattice fragments have been shown to be an effective approach for detecting abnormal structures that are inconsistent with the principle that the subsumption relation (is-a) in an ontological system should conform to the lattice property. This principle is further enforced in more recent work [5], where non-lattice fragments have shown up to 40 times the change rate against background changes in the evolution of SNOMED CT.

The effectiveness of COBE for retrieving SNOMED CT concepts is evaluated in two ways. For the direct lookup mode, it is compared with the well-known IHTSDO SNOMED CT Browser [6] and NLM SNOMED CT Browser [7]. Evaluation against manually created reference standard showed that COBE attains an example-based precision of 0.958, recall of 0.917, and F₁ measure of 0.875. For the navigational exploration, we conduct an evaluation experiment on the search stems. The result shows that only using stems generated from 28,371 concepts which occurring in the upper bounds (uppermost level) of fragments, which is 9.37% of total, as navigational exploration menus covers 98.97% of all SNOMED CT concepts.

1 Background

1.1 SNOMED CT

SNOMED CT is the world’s largest and most comprehensive clinical healthcare terminology, developed and maintained by the International Health Terminology Standard Development Organization (IHTSDO). From a structural perspective, SNOMED CT can be seen as a series of large directed acyclic graphs, one for each of its 19 sub-hierarchies: Procedure, Physical force, Event, Staging and scales, Substance, Environment or geographical location, Situation with explicit context, Body structure, Observable entity, Pharmaceutical/biologic product, Physical object, Qualifier value, Special concept, Specimen, Social context, Clinical finding, Organism, Linkage concept, and Record artifact. Each concept comes with a unique SNOMED CT identifier, and concept descriptions associated with it. The version of SNOMED CT used in this study is dated September 1, 2014. It contains 302,902 active concepts, which are organized into 19 hierarchies linked by relationships such as is-a.

1.2 SNOMED CT Browsers

SNOMED CT Browsers are applications and tools used for viewing terminology content or hierarchy structure. Typically, a SNOMED CT browser can display descriptions and relationships of a concept by providing searching terms or identifier. The official IHTSDO SNOMED CT Browser [6] is an online, multilingual, multi-edition ontology browsing application. It enables browsing of International Edition of SNOMED CT, as well as various National Editions. User can enter terms or navigate in the taxonomy hierarchies to find target concept, with filters in semantic, refset, or language. The application can provide almost everything related to the concept, including descriptions, parents and children, attributes and diagram of structure. Another online browser, provided by U.S. National Library of Medicine (NLM), is NLM SNOMED CT Browser [7]. Different from the IHTSDO SNOMED CT Browser, searching and displaying SNOMED CT content, as included in UMLS Metathesaurus, is provided in NLM SNOMED CT Browser. It expands searches to include synonymous terms from over 100 Metathesaurus vocabulary sources. CSIRO provides a graph-based interactive interface for navigating SNOMED CT concepts, which is called Shrimp [8]. One can expand the hierarchy graph by clicking the concept, which provides an intuitive visualization of its relationships.

1.3 Conjunctive Navigational Exploration

There are two basic search modes in information retrieval. One is direct lookup (DL), and the other is navigational exploration (NE). Direct lookup is a basic search mode where a user knows precisely what to look for and comes up with input strings leading uniquely to the search target. On the other hand, in the navigational exploration mode, a user may not be able to easily and effectively formulate a descriptive search string, and must rely on navigational mechanisms such as menus to browse and explore the content in order to inform the user “what is there.”

Conjunctive Exploratory Navigation Interface (CENI [1]) is a recently introduced technique for effective access of online consumer health information using navigational exploration. CENI was achieved by assigning multiple topics for information items using semantic tagging [2, 3]. Crowd sourced comparative evaluation revealed that anonymous users from Amazon Mechanical Turk preferred 2 to 1 for CENI against other search mechanisms [1].

A defining feature of CENI is an interface which provides the user multiple paths to quickly narrow down to relevant contents. By selecting one topic at a time, in an incremental fashion, a user arrives at narrower and narrower content areas that are relevant to all the topics selected so far, conjunctively.

A novel feature of COBE is the NE mode, adapted from CENI. Instead of consumer health topics as menus in CENI, COBE uses word stems from an important sub-collection of SNOMED concepts – those appearing in non-lattice fragments as defined in [4]. Such word stems serve as attribute constraints, or “attributes” in Formal Concept Analysis, which allows the user to navigate to specific SNOMED CT concept clusters in multiple ways.

1.4 SNOMED CT Quality Assurance

It has been noted that the biomedical domain is developing tremendously. For example, SNOMED CT releases a new version almost every 6 months. The quality of an ontology is a key issue that determines its usability. Many auditing methods exist to ensure the quality of ontologies. They include but are not limited to lexical [9], structural [10], semantic [11, 12] and statistical-based [13]. A lattice structure is often one of the criterions of the well-formedness of ontologies [4]. A lattice is a structure that every two concepts in the ontology have no more than one minimal common ancestor or maximal common descendant. MaPLE [5] has found a large number of such non-lattice structure in SNOMED CT. Figure 1 is a non-lattice fragment from SNOMED CT. The double-circled concepts “Tissue specimen from breast” and “Tissue specimen from heart” share two minimal ancestors: “Tissue specimen” and “Specimen from trunk,” highlighted in pink, which makes them a non-lattice fragment. To make it lattice-conforming, one could add the concept “Tissue specimen from trunk” (dashed in Figure 1) [4, 5, 10].

Figure 1:

A non-lattice fragment in SNOMED CT [5].

2 Methods

Figure 2 depicts the overall architecture of the proposed COBE to search SNOMED CT concepts and facilitate the retrieval of non-lattice fragments in SNOMED CT. For the navigational exploration mode, SNOMED CT concepts are preprocessed (dotted rectangular box) to obtain a collection of core stems for tagging concepts. SNOMEC CT concepts are filtered by non-lattice fragments, and performed word tokenization and frequency ranking. As a result, a collection of core stems are mined and used for tagging concepts. In the navigational exploration (NE) mode of the COBE search interface, a user’s input (NE-input in Figure 2) is the selection of stem tags, based on which COBE performs NE-based conjunctive search and returns to the user a list of relevant concepts as well as numbers of related non-lattice fragments. In the direct lookup (DL) mode of the COBE search interface, a user’s input (DL-input in Figure 2) is a specified term formulated by the user. COBE performs term splitting, and DL-based conjunctive search to find a list of relevant concepts as well as numbers of related non-lattice fragments and returns them to the user.

Figure 2:

Overview of the Conjunctive Ontology Browser and Explorer COBE. NE: navigational exploration, DL: direct lookup.

2.1 Navigational Conjunctive Exploration

In the navigational exploration mode, a user may want to explore “what is there,” or may not be able to easily or effectively formulate a descriptive lookup term to find non-lattice fragments. In either case, the user may rely on navigational menus or facets to browse and explore. To address such need, we construct a cloud of informative terms serving as the navigational menus for the user to effectively explore SNOMED CT concepts and non-lattice fragments. To achieve this, we perform the following preprocessing steps:

• First, we extract all active concepts from the SNOMED CT distribution files. For each active concept, its identifier and descriptions (including fully specified name, preferred name, and synonyms) are further extracted and stored in a MySQL database.

• Second, we filter active concepts by non-lattice fragments, and only keep those concepts appearing in the uppermost level of non-lattice fragments. If a concept appears in the uppermost level of a non-lattice fragment, then the non-lattice fragment is called related to the concept.

• Third, for each of the remaining concepts, we remove punctuation marks and all stop-words such as “of” and “out” from its descriptions, which is then tokenized into individual stems. The resulting collection of tokenized singular words forms our COBE stem collection. Note that this is different from “word stems” in linguistics which refer to the prefix strings serving as the root of a word. When we mention stems in this paper, it is meant to be in the sense of COBE stem.

• Fourth, for each stem, we compute the frequency it appears in the remaining concepts to determine the order of the stems in the search menus.

After the above steps, the stems ranked by the precomputed frequencies serve as the navigational menus. Each stem is guaranteed to hit at least one fragment, which prevents the circumstance that retrieved concepts from COBE have no related non-lattice fragment.

Given a selection of input stems by a user, COBE performs conjunctive search to retrieve matching concepts, that is, only concepts matching all the selected stems are returned, ranked by the number of related non-lattice fragments in the descending order. COBE allows concepts being investigated and narrowed down by several times of navigation sequentially. For example, clicking navigation stem “heart” followed by clicking “attack” will narrow down the results to “Myocardial infarction (disorder).” COBE enables the combination of navigational exploration and direct lookup. For instance, clicking “attack” followed by typing “heart” in the direct search box will narrow down the results to “Myocardial infarction (disorder).”

2.2 Direct Lookup

The other way to navigate to non-lattice fragments is specifying a targeted term in the direct search box instead of formulating a term by the available navigation stems. In this direct lookup mode, a user knows what to look for, comes with specific term or exact SNOMED Identifier, and tries to retrieve a list of concepts. After a user types a specific term into the direct search box, COBE splits the term into stems and performs conjunctive search to retrieve concepts matching all these stems, and returns them by the number of related non-lattice fragments in the descending order.

2.3 Evaluation

We evaluate COBE in two ways. For the navigational exploration mode, we evaluate if the search stems obtained from the preprocessing step covers a wider range of concepts. To calculate the percentage of the coverage, we use the total number of concepts as the denominator, and the number of concepts that can be reached by at least one search stem as the numerator:

$$\text{coverage}=\frac{\text{number of concepts that can be reached by at least one search stem}}{\text{total number of concepts}}$$

For the direct lookup mode, we evaluate the search performance of the COBE interface by comparing it to other two SNOMED CT browsers: IHTSDO SNOMED CT Browser [6] and NLM SNOMED CT Browser [7]. We randomly select 24 concepts as search tasks (one lookup term per task) from the CORE Problem List Subset of SNOMED CT [14] to compare the three browsers. Since most SNOMED CT concepts have short length of descriptions, the terms chosen for the search tasks are 1-word, 2-word, 3-word or 4-word to make sure a certain number of concepts get returned by the browsers. 6 concepts (search tasks) are randomly selected for each word length. For each search task, an evaluator uses three browsers to retrieve a list of relevant concepts, respectively. A reference standard is created for each search task: the common concepts found by all three browsers are considered correct results and included in the reference standard; for the other concepts found but not shared by all three browsers, the evaluator manually reviews them and includes the relevant concepts into the reference standard only if the evaluator considers the result concepts correct. We use the example-based precision, recall and F₁ measure [2, 3, 15, 16] to evaluate the performances of the three browsers.

Let R be the reference standard consisting of m = 24 search tasks {(s_i, Y_i) | i = 1, …, m}, where Y_i is the set of all concepts included in the reference standard for the search task s_i. Let Z_i be the set of concepts retrieved from a search interface for s_i. The example-based precision (P), recall (R) and F₁ measure (F₁) are calculated as follows:

$$P=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{\left|Y_i\cap Z_i\right|}{\left|Z_i\right|}},R=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{\left|Y_i\cap Z_i\right|}{\left|Y_i\right|}},F_1=\frac{1}{m}{\displaystyle \sum _{i=1}^m\frac{2\left|Y_i\cap Z_i\right|}{\left|Z_i\right|+\left|Y_i\right|}}$$ (1)

3 Results

Among all the 302,902 active concepts, only 28,371 are in uppermost level of non-lattice fragments. We used 28,371 concepts to generate 12,623 search stems for navigational exploration. Table 1 displays the 10 most frequent stems appearing in the concepts related to non-lattice fragments.

Table 1:

Top-10 most frequent navigation search stems and their frequencies. Frequency: the number of concepts involved in non-lattice fragments a stem appears in.

Search Stem	Frequency
structure	2,640
finding	1,292
disorder	1,070
neoplasm	1,041
procedure	971
joint	755
artery	654
bone	641
system	629
entire	615

3.1 Conjunctive Ontology Browser to Explore SNOMED CT Fragments (COBE)

Figure 3 shows a sample screenshot after clicking “neck” and “head” from search stems for navigational exploration and typing “region” for direct search. The left column displays the list of navigation terms, while clicking the navigation menu bars of the left column, the chosen words are displayed inside the horizontal bar on the top of the right column, where the “Reset” button is used to start a new exploration by clearing the search words. The leftmost button of the right column shows the number of non-lattice fragments where the upper bounds contain this concept, where green button indicates existing non-lattice fragments and orange for no fragments found. The center area indicates the fully specified name and blue button for the SNOMED CT concept identifier. When hover the name COBE shows all the preferred terms and synonyms. Clicking either the green button or the concept name links to the page for visualizing relevant fragments. By default, all the concepts are displayed in the descending order of the number of fragments, if neither search term nor search keyword is specified.

Figure 3:

A screenshot of the conjunctive ontology browser explorer interface COBE.

3.2 Visualizing Non-lattice Fragments

As shown in the right column in Figure 3, COBE enables user to find some specified concepts listed in the concept displaying area. Once the concepts are displayed, a user can click the leftmost green or orange button to link to the index page of related fragments and finally direct users to the fragments visualization page.

Figure 4 demonstrates searching “neck head region” in both NE and DL modes to render a non-lattice fragment. The upper left part of Figure 4 demonstrates the navigational exploration mode of COBE, and upper right part for the direct lookup. Using each mode or combining using these two modes leads to the visualization part. The lower left part of Figure 4 is the fragment index page for the concept “Pain of head and neck (disorder).” The left column of the index page contains all fragments denoted by upper bounds concept identifiers, and right column displays the corresponding concept labels. After clicking the concept identifiers of a fragment, the browser directs to the page for the visualization of that fragment which is shown in the lower right part. The green nodes represent the upper bounds of the fragment, the yellow ones represent lower bounds and the gray ones represent intermediate nodes in between. The solid gray edges represent is-a relationship of two concepts. For example, “Chronic pain (finding)” and “Pain of head and neck region (finding)” are upper bounds of the fragment, while “Chronic neck pain (finding)” and “Chronic pain in face (finding)” are lower bounds of the fragment. The leftmost gray edge means “Chronic neck pain (finding)” is-a “Chronic pain (finding).”

Figure 4:

Searching “neck head region” in both NE and DL modes to render related non-lattice fragments.

3.3 Navigational Exploration Evaluation

Instead of using descriptions of all 302,902 concepts, we used a subset of 28,371 concepts related to fragments to generate the search stems for navigational exploration. Each of the 28,371 concepts must occur in the upper bounds of a fragment at least one time. 12,623 search stems were generated after punctuation marks removed, tokenized, and stop-words removed. To evaluate the search stems, we queried all search stems for each of the 302,902 concepts, and found that 299,789 concepts can be reached through at least one the search stems. This implies that using navigational exploration, at most 3,113 concepts might be missed. These 3,113 concepts must have no occurrence in the upper bounds of fragments. The coverage of the search stems to reach overall concepts is 98.97%(299, 789/302, 902). This indicates that our search stems formed by a small subset of concepts 9.37%(28371/302902) can cover most SNOMED CT concepts.

3.4 Direct Lookup Evaluation

To evaluate the search interface, we formulated 24 search tasks to compare the three SNOMED CT browsers’ performances: IHTSDO SNOMED CT Browser, NLM SNOMED CT Browser, and COBE interface. We randomly picked 24 concepts from the CORE Problem List Subset of SNOMED CT and used their descriptions as the searching key words for three browsers. To make sure certain number of results can be returned, we limited the search tasks to 1-word, 2-word, 3-word or 4-word concepts. Then the evaluator used three browsers to perform the search tasks. The reference standard was created for each search task: the common concepts found by all three browsers are considered correct results and included in the reference standard. For those concepts found but not shared by all three browsers, the evaluator manually reviewed them and decided whether to include them into the reference standard or not, according to the relevancy of the result and the search task. Only those concepts considered relevant by the evaluator were included into the reference standard.

Table 2 shows the search tasks, the number of concepts in reference standard, the number of concepts found by each browser and the number of concepts included in reference standard found by each search interface. Three interfaces yielded the same results for 17 of the 24 search tasks. The example-based precision (P), recall (R) and F₁ measure (F₁) were calculated based on the results and the reference standard using formula (1) in Section 2.3. Table 3 shows the overall example-based precision, recall, and F₁ for the results of 24 search tasks using NLM, IHTSDO and COBE. The result shows COBE carried out the best recall of 0.917; NLM achieved the best precision of 1.0; both COBE and NLM had the best F₁ measure of 0.875. This indicates that COBE is comparable to the state-of-the-art SNOMED CT browsers for looking up concepts.

Table 2:

Comparison of searching results of 3 SNOMED CT Browser. Y - the number of concepts in reference standard Y. Z_C - the number of concepts found by COBE. Z_C ∩ Y - the number of concepts found by COBE that included in reference standard. Z_I - the number of concepts found by IHTSDO browser. Z_I ∩ Y - the number of concepts found by IHTSDO browser that included in reference standard. Z_N - the number of concepts found by NLM browser. Z_N ∩ Y - the number of concepts found by NLM browser that included in reference standard.

Search Term	Y	ZC(ZC ∩ Y)	ZI(ZI ∩ Y)	ZN(ZN ∩ Y)
Panniculitis	39	39 (39)	39 (39)	39 (39)
Hepatoblastoma	2	2 (2)	2 (2)	2 (2)
Neurilemmoma	13	13 (13)	13 (13)	13 (13)
Oligomenorrhea	3	3 (3)	3 (3)	3 (3)
Microcephalus	5 (5)	5 (5)	5 (5)	5 (5)
Thalassemia	64	64 (64)	57 (57)	63 (63)
Hand pain	2	2 (2)	3 (2)	2 (2)
Mononucleosis syndrome	5	5 (5)	5 (5)	5 (5)
Deficiency anemias	2	2 (2)	2 (2)	2 (2)
Viral screening	4	4 (4)	4 (4)	4 (4)
Angina decubitus	1	1 (1)	1 (1)	1 (1)
Renal hypertension	17	17 (17)	16 (16)	16 (16)
Sleep terror disorder	1	1 (1)	1 (1)	1 (1)
Mantle cell lymphoma	4	4 (4)	4 (4)	4 (4)
Tricuspid incompetence, non-rheumatic	1	1 (1)	1 (1)	1 (1)
Acute frontal sinusitis	2	2 (2)	2 (2)	2 (2)
Renal tubular acidosis	11	11 (11)	11 (11)	11 (11)
Iatrogenic Cushing’s disease	1	1 (1)	1 (1)	1 (1)
Drusen of optic disc	4	1 (1)	1 (1)	4 (4)
Therapeutic drug monitoring assay	1	1 (1)	1 (1)	1 (1)
Malignant neoplasm of brain	14	14 (14)	13 (13)	12 (12)
Bronchopulmonary dysplasia of newborn	1	1 (1)	1 (1)	1 (1)
Antenatal ultrasound scan abnormal	1	2 (1)	2 (1)	1 (1)
Derangement of temporomandibular joint	2	1 (1)	1 (1)	2 (2)

Table 3:

The example-based precision, recall, and F₁ measure for the results of concept search using COBE, IHTSDO and NLM.

SNOMED CT Browser	Precision	Recall	F1
COBE	0.958	0.917	0.875
IHTSDO	0.958	0.792	0.75
NLM	1.0	0.875	0.875

4 Discussions

Our evaluation of the direct lookup mode of COBE is limited with respect to the reference standard. Since SNOMED CT contains 302,902 concepts, going through the entire concept list to find all relevant concepts of each search task to build the reference standard is infeasible for the evaluator. The other limitation is that we only had one evaluator. This also could possibly introduce bias to the reference standard creation. A third limitation is that only stem coverage was used to evaluate the navigational exploration mode of COBE. It would be helpful to conduct a comparative user evaluation for the navigational exploration mode. However, there are no similar mechanisms in other ontology browser for us to compare COBE with.

We also performed an experiment using the entire SNOMED CT to generate search stems and compare the performance with the search stems from concepts involved in non-lattice fragments. 83,139 search stems were generated from the entire 302,902 concepts. The number of stems generated by non-lattice fragments was 12,623, which is 15.18% of the entire stems. However, such 15.18% of the entire stems already covers 98.97% of SNOMED CT concepts.

In related work, existing ontology visualization toolkits such as KC-Viz [17] and ProtégéVOWL [18] are not designed for searching and visualizing non-lattice fragments. Therefore, COBE uniquely fulfills this need.

5 Conclusion

In this paper, we presented a conjunctive ontology browser COBE to search and explore SNOMED CT concepts and non-lattice fragments for ontology quality assurance. The direct lookup and navigational exploration of COBE allow multiple entry points for users to explore information of interest. The combination of searching and visualization mechanisms in COBE provides a novel way for structural auditing of SNOMED CT. Navigational exploration and direct lookup offer complementary modes for finding specific SNOMED CT concepts. With respect to the direct lookup mode, evaluation against manually created reference standard showed that COBE attains an example-based precision of 0.958, recall of 0.917, and F₁ measure of 0.875. With respect to the navigational exploration mode, COBE leverages 28,371 concepts in non-lattice fragments to construct the stem cloud. With merely 9.37% of the total SNOMED CT stem cloud, our navigational exploration mode covers 98.97% of the entire concept collection. COBE has been deployed as the resident search interface for linking, exploring, and rendering of SNOMED CT non-lattice fragments, which represent an active area of ontology curation work.