Auditing the multiply-related concepts within the UMLS

Abstract

Objective

This work focuses on multiply-related Unified Medical Language System (UMLS) concepts, that is, concepts associated through multiple relations. The relations involved in such situations are audited to determine whether they are provided by source vocabularies or result from the integration of these vocabularies within the UMLS.

Methods

We study the compatibility of the multiple relations which associate the concepts under investigation and try to explain the reason why they co-occur. Towards this end, we analyze the relations both at the concept and term levels. In addition, we randomly select 288 concepts associated through contradictory relations and manually analyze them.

Results

At the UMLS scale, only 0.7% of combinations of relations are contradictory, while homogeneous combinations are observed in one-third of situations. At the scale of source vocabularies, one-third do not contain more than one relation between the concepts under investigation. Among the remaining source vocabularies, seven of them mainly present multiple non-homogeneous relations between terms. Analysis at the term level also shows that only in a quarter of cases are the source vocabularies responsible for the presence of multiply-related concepts in the UMLS. These results are available at: http://www.isped.u-bordeaux2.fr/ArticleJAMIA/results_multiply_related_concepts.aspx.

Discussion

Manual analysis was useful to explain the conceptualization difference in relations between terms across source vocabularies. The exploitation of source relations was helpful for understanding why some source vocabularies describe multiple relations between a given pair of terms.

Introduction

Decades of natural language processing and artificial intelligence research on the methods for terminology acquisition and structuring¹ have resulted in an increasing number of available terminologies. These terminologies often describe complementary features of scientific and technical areas. Consequently, their integration can be useful for the comprehensive description and modeling of these areas. Moreover, the issues specific to integration may also be important in other contexts, such as maintenance, updating,² evolution,^{3 4} and transcoding or alignment⁵ of terminologies and ontologies. As an illustration, the Unified Medical Language System (UMLS)^{6 7} integrates over 170 biomedical terminologies. The result of this integration is particularly useful and widely employed in the biomedical area for various applications (eg, information retrieval and extraction, coding of discharge patient records, question/answering systems).

Although sometimes necessary, integration may however also cause conceptual and structural inconsistencies. At the scale of a single terminology, inconsistencies may be found, while the situation becomes more complex when terminologies are merged. Indeed, terminologies are designed and created with different objectives, and have different underlying principles. In order to detect and correct these potential limitations, researchers have proposed methods for auditing terminological resources. Such methods have been applied to WordNet⁸ for redundancy and consistency checking.^9–12 In order to apply this approach to the biomedical domain, we adopt the analysis grid proposed in a recent review of auditing methods,¹³ and distinguish three aspects:

1. Terms and concepts: this aspect focuses on term labeling,¹⁴ ambiguity and polysemy,^15–17 synonymy completeness,^18–22 coverage for a given subdomain or application,^{18 21 23–26} and modifier influence.²⁷

2. Semantic categorization: the consistency of the UMLS semantic categorization of concepts is checked according to the hierarchical relations associating these concepts.^{16 26 28–33}

3. Semantic relationships: the consistency of hierarchical relationships and their coverage have been widely studied,^{15 19 23 28 30 34–41} while other types of relationships have not been studied properly to date.

Our study concerns the first and third aspects. We propose analyzing the multiply-related UMLS concepts, that is, concepts which are associated through multiple relationships. This situation arises within the UMLS because, during the integration of source vocabularies, any information related to terms and relations is preserved. In our opinion, beyond the integration of synonymous terms and the increase in the lexical coverage provided by a single terminology, the generation of multiply-related concepts is another artifact of terminology integration. To our knowledge, this aspect has not yet been systematically investigated. The closest work,⁴² which focuses on such UMLS concepts, only performs manual categorization and review of some common situations. In our study, we propose automatic methods to audit all multiply-related UMLS concepts. Basically, we study the compatibility of the multiple relations which associate the concepts under investigation and seek to explain the reason why they co-occur. In addition, we randomly select 288 concepts associated through contradictory relations and manually analyze them. Our previous work⁴³ is strengthened with several detailed analyses of the data. In addition, new aspects have been added, such as clarification of the main terms used in the paper and investigation of incompatibilities between relations at the scale of source terminologies, especially by considering the term level.

Material

The UMLS is a terminological system whose main component, the Metathesaurus, integrates 173 source vocabularies and represents a huge graph composed of 2 669 792 concepts. A concept corresponds to a set of synonymous terms provided by different source vocabularies (in this paper, we use ‘source vocabularies’ and ‘source terminologies’ interchangeably). The concepts are organized within a very dense terminological network: 53 942 132 binary relations linking concepts are recorded in the investigated version of the UMLS (2012AA). Such a huge quantity of relations is due to the fact that, according to the UMLS building rules, all the relations existing in the source vocabularies have to be integrated within the UMLS, even when there are conflicting relations between two concepts.

There are 11 types of active UMLS relationships in the 2012AA version, which can be grouped into three general classes (table 1):

• Synonymy

• Hierarchical

• Associative.

Table 1.

UMLS relationships and the class to which they belong

Class of relationship	Abbreviation of the relationship	Meaning of the relationship
Synonymy	SY	Source asserted synonymy
Hierarchical	CHD	Has child relationship
	PAR	Has parent relationship
	RB	Has a broader relationship
	RN	Has a narrower relationship
	SIB	Has sibling relationship
Associative	AQ	Allowed qualifier
	QB	Can be qualified by
	RO	Has relationship other than synonymous, narrower or broader
	RL	Has similar or ‘alike’ relationship
	RQ	Related and possibly synonymous

Regarding the relationships at the scale of source vocabularies, the Metathesaurus records around 300 source relationships and assigns each of them to one of the 11 active UMLS relationships according to the source vocabularies documentation or their interpretation by the National Library of Medicine (NLM) team. For instance, the source relationship same_as is assigned to the SY relationship, inverse_isa to PAR, and has_component to RO.

Methods

For a given pair of concepts (C₁, C₂) associated through multiple relations, we study the compatibility between relationships at the UMLS scale and at the scale of source vocabularies. We then determine the reason for the presence of multiple relations between concept pairs. Before presentation of the methods, we define the main terms used and explain the material preparation.

Definitions

Relationship and relation: while relationship indicates a given type of relation (eg, synonymy relationship, hierarchical relationship), relation refers to every individual link between two given terms.

Multiply-related concepts: Two distinct UMLS concepts that are associated through at least two relationships, are multiply-related concepts. For example, Butyrolactone (C0178525) is multiply related to Lactones (C0022947) through PAR and RB.

Source relation(ship): A source relation(ship) corresponds to a relation(ship) which comes from source terminologies distinct from the 11 active UMLS relationships.

Symmetric relationships: Symmetric relationships correspond to relationships which can be read identically in both senses, that is, if a triplet (C₁, R, C₂) exists, then the triplet (C₂, R, C₁) is also present. Among the UMLS relationships, RL, RO, RQ, SIB and SY are symmetric.

Inverse relationships: When two concepts are linked to each other through reciprocal relationships from the same class, these relationships are characterized as inverse. The inverse relationships present in the UMLS are AQ/QB, PAR/CHD and RB/RN.

Material preparation and de-duplication

As mentioned previously, all (source) relations existing in the source vocabularies are integrated within the UMLS. Consequently, some relations may be redundant when described by distinct vocabularies. This results in identical triplets, of which we keep only one specimen. For example, the SY relation existing between Adrenocortical hyperfunction (C0001622) and Cushing's syndrome (C0010481) is defined both in MEDCIN and SNOMEDCT. Only one triplet (Adrenocortical hyperfunction, SY, Cushing's syndrome) is considered here.

Due to other building principles, binary relations are represented in both directions within the UMLS. In practice, if a source vocabulary describes that a given asymmetric/symmetric relation exists between C₁ and C₂, then its inverse/the same relation is also recorded between C₂ and C₁ when the source vocabulary is integrated within the UMLS. We de-duplicate such situations so that we do not analyze the same triplet twice. For instance, of the two triplets (Butyrolactone, RB, Lactones) and (Lactones, RN, Butyrolactone), only the first one is kept.

Finally, the labeled relationships mapped_to and mapped_from are defined as ‘one-to-one mappings between two vocabularies which are both present in the UMLS.’ We thus choose to ignore them because they are generated by the UMLS based on maps between source terminologies, and so are different from other relationships.

Compatibility of relationships associating multiply-related concepts

We study the compatibility of the relationships which associate multiply-related concepts and distinguish four situations (figure 1):

• Contradictory combinations: combinations which include inverse relationships (eg, CHD PAR, RB RL RN, AQ QB, PAR RN RO SIB). Although not natively inverse, the combinations involving PAR with RN and RB with CHD are also considered as contradictory

• Granularity difference: combinations which include SIB and/or SY combined with at least one of the following hierarchical relationships: PAR, CHD, RB, RN (eg, PAR SIB, PAR RO SIB, RB SY)

• Heterogeneous combinations: combinations involving relationships from distinct classes (ie, synonymy, hierarchical, associative) (eg, PAR RO, RQ SIB, PAR RB RO RQ)

• Homogeneous combinations: the two combinations PAR RB and CHD RN in addition to any combination of relationships within the associative class except for AQ QB because they are inverse (eg, RO RL, AQ RO RQ).

Figure 1.

The different categories of relation combinations between multiply-related concepts: (A) contradictory combinations, (B) combinations with granularity difference, (C) heterogeneous combinations, (D) homogeneous combinations.

The pairs are associated with only one of these categories in the order of their presentation. Thus, if a concept pair exhibits both contradictory and homogeneous combinations, then the pair is only counted in contradictory combinations.

The compatibility of relationships is investigated at the UMLS scale and at the scale of source terminologies. In addition, since contradictory combinations produce the most problematic situations, we manually analyze 10% of concept pairs randomly extracted from this set.

Reason for the presence of multiple relations between concepts

We want to determine, for each pair of concepts under investigation, if its combination of relations already exists in source vocabularies or if it results from their integration within the UMLS. Towards this end, for each pair of concepts, we first check if at least one source terminology describes its combination. If not, then the presence of multiple relations is due to the UMLS integration process. If the combination is observed in a given source vocabulary, it is necessary to check if it also exists at the term level. In practice, we analyze the terms which are clustered into multiply-related concepts and consider that the combination actually originates from the source vocabulary if the same pair of terms is multiply-related (figure 2).

Figure 2.

Analysis of multiple relations between UMLS concepts at the term level: (A) and (B) combinations generated during the UMLS integration process, (C) combinations already present in a source vocabulary.

Results

After de-duplication, the number of distinct concept pairs is 12 356 156 (involving 2 669 792 distinct concepts). Our study addresses the 439 087 concept pairs (involving 360 098 distinct concepts) which are associated through multiple relations, corresponding to 3.6% of the total number of concept pairs related within the UMLS (involving 13.5% of distinct concepts). The results presented in this section are available at: http://www.isped.u-bordeaux2.fr/ArticleJAMIA/results_multiply_related_concepts.aspx.

Compatibility of relationships associating multiply-related concepts

At the UMLS scale (table 2), combinations exhibiting contradictions and granularity differences represent 0.7% and 20.0% of the investigated concept pairs, respectively. Heterogeneous combinations are the most frequent (45.4%) and homogeneous combinations represent 33.9% of all investigated concept pairs. A total of 157 combinations are observed, and the 10 most common cover over 88.0% of the entire set. The three most frequent combinations correspond to different situations: homogeneous PAR RB (31.4%), heterogeneous PAR RO (26.4%) and granularity difference PAR SIB (11.6%). Conversely, 60 combinations (eg, CHD PAR RB RL RN RQ SIB, PAR RN SY, AQ RB, RB RL RO SY) occur less than 10 times. Although only 0.7% of concept pairs are related through contradictory combinations, this situation shows the highest number of combinations (59.9%).

Table 2.

Compatibility of relationships associating multiply-related concepts at the UMLS scale

	Contradictory combinations	Granularity difference	Heterogeneous combinations	Homogeneous combinations	Total
No. of multiply-related concept pairs	2880 (0.7%)	88038 (20.0%)	199272 (45.4%)	148 897 (33.9%)	439 087 (100.0%)
No. of combinations	94 (59.9%)	36 (22.9%)	23 (14.7%)	4 (2.5%)	157 (100.0%)
Most frequent combinations	CHD PAR (628) PAR RN (472) RB RN (159)	PAR SIB (50 785) PAR RB SIB (11 924) RB SIB (6951)	PAR RO (115 757) PAR RB RO (25 055) RB RO (18 281)	PAR RB (137 918) RO RQ (10 916) AQ RO (61)

At the scale of source vocabularies, the concepts investigated are provided by 66 source terminologies (table 3). We distinguish two categories:

• Twenty-two source vocabularies do not exhibit any combination of relationships (eg, HLREL, KCD5, MMSL, MTH, NIC, RXNORM). In other words, these source terminologies always contain only one relationship between two concepts.

• Forty-four source vocabularies describe multiple relations between the concepts under investigation (table 4). Only seven of them (all containing fewer than 70 pairs of multiply-related concepts) present a high percentage of contradictory combinations (eg, 100% for ICD10, CST and DSM4). Eleven source terminologies contain mainly concepts multiply related through relations with granularity difference (eg, over 97% for RCD, ICD9CM and ICD10AM), while 15 source vocabularies mainly exhibit heterogeneous combinations (eg, over 99% for LNC, NDFRT and NCI). The 11 remaining source terminologies exhibit a high percentage of concept pairs related through homogeneous combinations (eg, over 95% for AOD, CSP, PSY and NEU).

Table 3.

Acronym and name of source vocabularies providing multiply-related concepts

Vocabulary acronym	Vocabulary name
AIR	AI/RHEUM
ALT	Alternative Billing Concepts
AOD	Alcohol and Other Drug Thesaurus
AOT	Authorized Osteopathic Thesaurus
BI	Beth Israel Vocabulary
CCC	Clinical Care Classification
CCPSS	Canonical Clinical Problem Statement System
CCS	Clinical Classifications Software
CPM	Medical Entities Dictionary
CPT	Physicians’ Current Procedural Terminology
CSP	CRISP Thesaurus
CST	COSTART
DSM3R	DSM-III-R
DSM4	DSM-IV
FMA	Foundational Model of Anatomy Ontology
GO	Gene Ontology
HCPCS	Metathesaurus HCPCS Hierarchical Terms
HHC	Home Health Care Classification
HL7V2.5	HL7 Vocabulary V.2.5
HL7V3.0	HL7 Vocabulary V.3.0
ICD10	ICD10
ICD10AM	International Statistical Classification of Diseases and Related Health Problems
ICD10CM	International Classification of Diseases
ICD10PCS	ICD-10-PCS
ICD9CM	ICD-9-CM
ICF	International Classification of Functioning
ICNP	International Classification for Nursing Practice
ICPC	ICPC
ICPC2P	ICPC-2 PLUS
JABL	Online Congenital Multiple Anomaly/Mental Retardation Syndromes
KCD5	Korean Standard Classification of Disease V.5
LNC	LOINC 2.15
MDR	Medical Dictionary for Regulatory Activities Terminology (MedDRA)
MEDCIN	MEDCIN
MEDLINEPLUS	MedlinePlus Health Topics
MMSL	Multum MediSource Lexicon
MSH	Medical Subject Headings
MTH	UMLS Metathesaurus
MTHMST	Metathesaurus Version of Minimal Standard Terminology Digestive Endoscopy
MTHSPL	Metathesaurus FDA Structured Product Labels
NAN	NANDA Nursing Diagnoses: Definitions and Classification
NCBI	NCBI Taxonomy
NCI	NCI Thesaurus
NDFRT	National Drug File
NEU	Neuronames Brain Hierarchy
NIC	Nursing Interventions Classification (NIC)
NOC	Nursing Outcomes Classification
OMIM	Online Mendelian Inheritance in Man
OMS	Omaha System
PCDS	Patient Care Data Set
PDQ	Physician Data Query
PNDS	Perioperative Nursing Data Set
PPAC	Pharmacy Practice Activity Classification
PSY	Thesaurus of Psychological Index Terms
RAM	QMR clinically related terms from Randolph A. Miller
RCD	Read Thesaurus
RXNORM	RxNorm Vocabulary
SNM	SNOMED-2
SNMI	SNOMED International
SNOMEDCT	SNOMED Clinical Terms
ULT	UltraSTAR
UMD	UMDNS: Product Categories Thesaurus
USPMG	USP Model Guidelines
UWDA	University of Washington Digital Anatomist
VANDF	Veterans Health Administration National Drug File
WHO	WHOART

Table 4.

Compatibility of relationships associating multiply-related concepts at the scale of source vocabularies

Vocabulary acronym	Contradictory combinations	Granularity difference	Heterogeneous combinations	Homogeneous combinations	Total	Pairs uniquely related at the term level	Pairs multiply related at the term level	Number of combinations	Most frequent combination
UWDA	2 (0.0%)	2474 (2.8%)	6439 (7.4%)	78 568 (89.8%)	87 483	406 (0.5%)	87 077 (99.5%)	9	PAR RB (89.8%)
LNC	5 (0.0%)		64 948 (100.0%)		64 953	64 953 (100.0%)		2	PAR RO (100.0%)
RCD	241 (0.7%)	34 855 (99.3%)			35 096	34 920 (99.5%)	176 (0.5%)	4	PAR SIB (98.8%)
SNOMEDCT	242 (1.6%)	343 (2.3%)	14 021 (92.1%)	615 (4.0%)	15 221	15 036 (98.8%)	185 (1.2%)	29	PAR RO (86.7%)
MEDCIN	49 (0.3%)	209 (1.4%)	1919 (13.1%)	12 505 (85.2%)	14 682	14 678 (100.0%)	4 (0.0%)	16	PAR RB (85.1%)
AOD	12 (0.1%)	317 (2.3%)	299 (2.2%)	13 209 (95.5%)	13 837	345 (2.5%)	13 492 (97.5%)	27	PAR RB (95.1%)
CSP		82 (0.7%)	412 (3.5%)	11 245 (95.8%)	11 739	166 (1.4%)	11 573 (98.6%)	13	PAR RB (95.4%)
NDFRT	11 (0.1%)		9393 (99.9%)		9404	7874 (83.7%)	1530 (16.3%)	3	PAR RO (99.9%)
ICD10CM	76 (0.9%)	7550 (94.2%)	388 (4.8%)		8014	8014 (100.0%)		9	PAR SIB (93.7%)
ICD9CM	37 (0.6%)	5979 (97.3%)	129 (2.1%)		6145	6145 (100.0%)		10	PAR SIB (96.5%)
ICD10AM	73 (1.2%)	5930 (98.8%)	2 (0.0%)		6005	6005 (100.0%)		6	PAR SIB (98.2%)
PSY			19 (0.3%)	5548 (99.7%)	5567	27 (0.5%)	5540 (99.5%)	7	PAR RB (99.3%)
MDR	69 (1.6%)	4136 (93.4%)	221 (5.0%)		4426	4426 (100.0%)		11	PAR SIB (92.8%)
GO	15 (0.5%)	1224 (37.9%)	553 (17.1%)	1438 (44.5%)	3230	1301 (40.3%)	1929 (59.7%)	33	RB SIB (32.4%)
WHO		1018 (35.8%)	32 (1.1%)	1790 (63.0%)	2840	585 (20.6%)	2255 (79.4%)	11	PAR RB (63.0%)
NCI	8 (0.3%)	1 (0.0%)	2694 (99.7%)		2703	71 (2.6%)	2632 (97.4%)	5	PAR RO (99.4%)
FMA	1 (0.0%)	196 (9.3%)	1901 (90.6%)		2098	536 (25.5%)	1562 (74.5%)	5	RO SIB (79.2%)
MEDLINEPLUS		119 (6.1%)	1828 (93.8%)	1 (0.1%)	1948	373 (19.1%)	1575 (80.9%)	10	RO SIB (82.2%)
MSH	8 (0.4%)	760 (40.0%)	1109 (58.4%)	22 (1.2%)	1899	960 (50.6%)	939 (49.4%)	11	PAR SIB (38.1%)
NEU		1 (0.1%)		805 (99.9%)	806	1 (0.1%)	805 (99.9%)	2	PAR RB (99.9%)
UMD	11 (1.5%)	102 (13.8%)	599 (81.3%)	25 (3.4%)	737	234 (31.8%)	503 (68.2%)	17	RO SIB (54.8%)
VANDF			588 (100.0%)		588	588 (100.0%)		1	RB RO (100.0%)
SNMI	135 (29.5%)		322 (70.5%)		457	457 (100.0%)		3	PAR RO (70.2%)
OMS			376 (100.0%)		376		376 (100.0%)	1	PAR RO (100.0%)
AOT		6 (1.7%)	91 (25.6%)	258 (72.7%)	355	2 (0.6%)	353 (99.4%)	3	PAR RB (72.7%)
ICF		263 (100.0%)			263	263 (100.0%)		1	PAR SIB (100.0%)
CCS	1 (0.4%)	47 (19.0%)	200 (80.6%)		248	248 (100.0%)		8	PAR RQ (75.4%)
PDQ	4 (2.1%)	7 (3.7%)	177 (94.1%)		188	11 (5.9%)	177 (94.1%)	4	PAR RO (94.1%)
NOC	2 (2.1%)	93 (97.9%)			95	95 (100.0%)		2	PAR SIB (97.9%)
ICD10	61 (100.0%)				61	61 (100.0%)		1	CHD PAR (100.0%)
SNM	18 (62.1%)		11 (37.9%)		29	29 (100.0%)		2	CHD PAR (62.1%)
CST	21 (100.0%)				21	21 (100.0%)		1	CHD PAR (100.0%)
OMIM				20 (100.0%)	20	20 (100.0%)		1	RO RQ (100.0%)
CPM		14 (100.0%)			14	14 (100.0%)		1	PAR SIB (100.0%)
DSM4	11 (100.0%)				11	11 (100.0%)		1	CHD PAR (100.0%)
CPT	10 (100.0%)				10	10 (100.0%)		1	CHD PAR (100.0%)
HHC		8 (100.0%)			8	8 (100.0%)		1	PAR SIB (100.0%)
DSM3R	6 (100.0%)				6	6 (100.0%)		1	CHD PAR (100.0%)
ALT		6 (100.0%)			6	6 (100.0%)		1	PAR SIB (100.0%)
USPMG		4 (100.0%)			4	4 (100.0%)		1	PAR SIB (100.0%)
HL7V3.0			2 (100.0%)		2	2 (100.0%)		1	PAR RO (100.0%)
MTHMST			1 (100.0%)		1	1 (100.0%)		1	RB RO (100.0%)
BI				1 (100.0%)	1	1 (100.0%)		1	RO RQ (100.0%)
NCBI	1 (100.0%)				1	1 (100.0%)		1	CHD PAR (100.0%)

Except for the last two columns, the indicated numbers correspond to pairs of concepts which are associated through combinations of relations or through unique/multiple relations at the term level.

Source vocabularies highlighted in italics mainly present multiple relations between terms whose combination is homogeneous, while those highlighted in bold mainly present multiple relations between terms whose combination is not homogeneous.

Reason for the presence of multiple relations between concepts

For 179 963 concept pairs (41.0%), the combination of relations is not present in source terminologies and thus appears during their integration into the UMLS. For example, Skeletal system (C0037253) is related to Skeletal bone (C0262950) through CHD in LNC (represented by a PAR relationship from Skeletal system to Bones in figure 3A) and PAR in RCD. When investigating the term level for the remaining 259 124 concept pairs, we discover that for 150 109 concept pairs (34.2%), distinct term pairs are associated through unique relations. This means that the presence of multiple relations between the corresponding concepts actually results from the UMLS integration process, when the different terms were clustered into a unique concept. As an illustration, Ameboma of intestine (C0494031) is related to Amebiasis (C0002438) through PAR and SIB according to ICD10CM. However, the ICD10CM term belonging to the former concept, Ameboma of intestine, is associated with the ICD10CM terms Amebiasis and Amebiasis, unspecified (part of the second concept) through PAR and SIB, respectively (figure 3B). Consequently, ‘only’ 109 015 concept pairs (24.8%) are already multiply related in source vocabularies. For example, DNA, A-form (C0000702) is related to Nucleic acid conformation (C0028599) through PAR and RO because the corresponding MSH terms are associated through these multiple relations (figure 3C).

Figure 3.

Illustrations of multiple relations between UMLS concepts at the term level: (A) a contradictory combination generated during the UMLS integration process, (B) a combination exhibiting granularity difference generated during the UMLS integration process, (C) a heterogeneous combination already present in MSH.

At the scale of source terminologies, we further analyze the 44 source terminologies exhibiting multiple relations between the concepts under investigation. We indicate in the seventh and eighth columns of table 4 whether the relations are unique or multiple at the term level. The 14 source terminologies exhibiting mainly pairs multiply related at the term level are those which require further investigation because they associate a given pair of terms through multiple relations. Among them, two profiles can be distinguished:

• Seven source vocabularies mainly associate terms through homogeneous combinations of relations (eg, UWDA, AOD, CSP and PSY), PAR RB most of the time (highlighted in italics in table 4).

• The seven remaining source terminologies principally relate terms with non-homogeneous combinations (eg, GO, NCI, FMA and MEDLINEPLUS), such as RB SIB, PAR RO and RO SIB (highlighted in bold in table 4).

Detailed analysis of contradictory combinations

We first determine the reason for the presence of multiple relations between the 288 pairs of concepts under investigation (10% of 2880). For 182 of them, the combination of relations does not exist in the source vocabularies. For the 106 remaining pairs, only three of them are actually multiply related in source terminologies. For example, Common bile duct (C0009437) and Hepatopancreatic ampulla (C0042425) are related through CHD, PAR, RN and RO in UWDA. Thus, UMLS integration is responsible for 99.9% of these contradictory combinations.

Manual analysis of these 288 pairs indicates that contradictory combinations often exist because of inherent or terminology-induced semantic features of terms:

• The semantic value of compounds which coordinate terms (colorectal and colon/rectum), as previously observed in Bodenreider et al,²⁷ and of the coordination (and/or) in general may differ according to source vocabularies. For instance, in MDR, Esophageal stenosis (C0014866) is parent of Oesophageal stenosis and obstruction (C0851721), while the relation is inverted in MEDCIN. We assume this situation is due to the meaning given to the coordination: it may be used to create more general terms or to specify terms. The semantic meaning of the unspecified (NOS, NEC, etc) modifier can also vary according to source terminologies.

• The implicit nature of some modifiers may have an impact on semantic relations. For example, Total nephrectomy (C0176996) and Nephrectomy (C0027695) are associated through five relations (PAR, RN, RO, SIB, SY). The difference seems to be due to the total modifier, which may be implicit (SY is then proposed) or not (other relationships are then proposed).

• Functional and causal links between terms also present great variations when they are transformed into hierarchical relationships. For instance, Angioedema (C0002994) and Urticaria (C0042109), which are common manifestations of allergic reactions, are related through CHD, PAR, RB, RO, RQ and SIB.

• The ambiguity of some components of the terms can result in contradictory relationships across the source vocabularies. For example, Adrenal gland diseases (C0001621) and Dysfunction adrenal (C0549609) are associated through PAR, RN and RQ, in which dysfunction can mean (broadly) any disorder, or (narrowly) specific conditions in which endocrine function is either increased or decreased.

• The difficulty for reflecting the structure of chemical products sometimes results in contradictory relations representing the different compounds. For instance, the link between Clodronic acid (C0012081) and Clodronate (C0162357) attempts to translate the chemical derivation of products into hierarchical (PAR, RB, RN) or other (RO) relationships.

• Terms with descriptive labels or terms reflecting complex biomedical notions, coupled with the choice of source vocabularies to create all possible links between such terms, may lead to inconsistencies. In fact, this situation may involve and accentuate any other cause discussed above:

  • The meaning of terms may overlap, while each of them may have its own modifier(s), such as skin and other in the pair Skin diseases, bullous (C0085932) and Other bullous disorders (C0494828)
  • They may combine several other causes already mentioned.

Discussion

Findings and limitations

Our study has interesting findings. At the UMLS scale, contradictory combinations are infrequent and may result from the fact that the conceptualization of relations between terms can be very different across source terminologies. Explanations provided by manual analysis cast some light on this difference. Homogeneous combinations are observed in one-third of situations, indicating the presence of redundant relations between concepts. At the scale of source terminologies, compatibility analysis reveals that one-third of them always describe only one relation between a given pair of concepts. Among the 44 remaining source vocabularies, 14 of them use multiple relations to associate a unique pair of terms. When the combination of these multiple relations is homogeneous, this indicates distinct but coherent points of view for expressing a link between two terms. On the other hand, when the relations constitute non-homogeneous combinations, the situation is more troublesome because this indicates that a given source terminology describes multiple and potentially incompatible relations between two terms. Three of such source vocabularies are further investigated in the following section. Finally, analysis at the term level also shows that the source vocabularies are responsible for the presence of multiply-related concepts in the UMLS only in a quarter of cases.

Our study has several limitations. First, it concentrates on precise situations within the UMLS. Because these correspond to complex relation combinations, they represent a small percentage of the entire set of UMLS relations. Second, our study is limited to multiply-related concepts when they are distinct. It would be interesting to investigate multiple relationships existing within a unique concept. Finally, our analysis at the term level may have underestimated the responsibility of source vocabularies for the presence of multiple relations between the investigated concepts. In particular, MDR uses distinct terms (and codes) to represent a unique concept,⁴⁴ which may result in misinterpretations. As an illustration, Agranulocytosis (C0001824) and Neutropenia (C0027947) are related through PAR and SIB in the UMLS because the MDR term Agranulocytosis (MDR code: 10001507) is associated with Neutropenias (10029355) and Neutropenia (10029354) through PAR and SIB, respectively. Here, the actual reason why multiple relations appear between the corresponding concepts is because MDR does not cluster synonymous terms into a unique code.

Analysis of source vocabularies that give rise to non-homogenous combinations at the term level

As seen in the ‘Reason for the presence of multiple relations between concepts’ section, seven source terminologies principally relate terms with non-homogeneous combinations, which may be problematic. We study three of them in more detail by analyzing the combinations of source relationships.

In GO,⁴⁵ the most frequent combination is RB SIB, which is principally associated with the has_part none combination of source relationships. ‘None’ appears when a given relationship is not specified by a source vocabulary. Here, this is due to the presence of the relationship SIB, which is systematically added by the UMLS when two concepts have a common parent in a given source terminology. The relationship has_part is assigned to RB within the UMLS. When investigating examples of such combinations, we have observed redundant relations in GO, which then result in multiple relations within the UMLS. As an illustration from GO (figure 4), Intracellular canaliculus is represented as part_of Apical plasma membrane. In addition, Intracellular canaliculus is also defined as part_of Plasma membrane part, to which Apical plasma membrane is related through isa. The UMLS records a SIB relationship between Intracellular canaliculus (C0230646) and Apical plasma membrane (C1167182) because they are both associated with Plasma membrane part (C1820065) (although in the first case it is through a partitive relationship, whereas it is a subsumption one in the second case). As indicated in GO,⁴⁶ if A part_of B and B isa C, then A part_of C. Thus, the relation part_of existing between Intracellular canaliculus and Plasma membrane part should be removed from GO because it can be inferred. Without this redundant relation, no SIB relationship would be recorded in the UMLS and only a single relation would exist between the corresponding concepts.

Figure 4.

A redundant partitive relation (dotted line) described in GO.

In NCI,⁴⁷ the two major combinations of source relationships are inverse_isa parent_is_cdrh and inverse_isa is_biochemical_function_of_gene_product, both corresponding to the combination PAR RO. In any case, inverse_isa is assigned to PAR, while the second source relationship is assigned to RO. The relationship parent_is_cdrh is defined as a property ‘created to allow the source CDRH to assign a parent to each concept with the intent of creating a hierarchy that includes only terms in which they are the contributing source.’ According to this definition, the UMLS apparently misinterprets these relationships, which should preferably be assigned to PAR. Conversely, NCI surprisingly (and probably wrongly) combines inverse_isa with is_biochemical_function_of_gene_product for more than 1000 pairs of terms. An example is Interleukin-13 (C0214743) and Interleukins (C0021764), which are both gene products and should not be related through is_biochemical_function_of_gene_product.

In MEDLINEPLUS,⁴⁸ the only combination of source relationships is related_to none, corresponding to the combination RO SIB. This situation occurs because MEDLINEPLUS is not a real terminology. Indeed, it provides information about high level subject categories for consumer health information, which group together medical topics associated only through related_to relationships, although more precise relationships would sometimes be more appropriate. As an illustration, Spina Bifida (C0080178) and Neural Tube Defects (C0027794) are associated through related_to in MEDLINEPLUS (RO) and, because they belong to the same medical topic group Genetics/Birth Defects (C1456603), a SIB relationship is also recorded within the UMLS between these two concepts. Actually, Spina Bifida should be described as isa Neural Tube Defects (its MEDLINEPLUS definition begins with ‘It is a type of neural tube defect’).

Preventing contradictory relationships

It should be noted that the presence of multiple relationships within the UMLS is not problematic per se, especially because its objective is to preserve all the relations asserted in the source vocabularies. However, this situation becomes problematic when the relations between two concepts are contradictory. As shown in the sections ‘Reason for the presence of multiple relations between concepts’ and ‘Detailed analysis of contradictory combinations,’ incompatible relationships are predominantly generated during the UMLS integration process and we mentioned multiple reasons why this happens. To avoid such situations, a simple solution could be to create a new concept if symmetric hierarchical relationships co-occur. This could solve cases like that presented in figure 3B: with an additional concept containing Amebiasis, unspecified, there would no longer be a granularity difference between relationships. Nevertheless, this solution may cause a dramatic increase of the number of concepts within the UMLS and may not solve all the problematic situations generated during the integration process. Thus, a clear identification of the problem, similar to that proposed here, is required so that users are aware of such situations and consider them cautiously.

In some cases, however, some source vocabularies define contradictory combinations between terms. For example, SNOMEDCT relates Pleural membrane structure (C0032225) and Entire pleura (C1279036) through inverse_isa and part_of (recorded as PAR and RN within the UMLS, respectively), although this appears to be contradictory. We suggest that such relations should be removed from source terminologies.

Conclusion

Our study is different from previous proposals in several ways: (1) we focus on multiply-related concepts; (2) we propose automatic and manual analyses of synonymous, hierarchical and associative relations; (3) we investigate the relationships’ compatibility; (4) and we study these situations at the UMLS scale and at the scale of source vocabularies (concept and term levels). The source terminologies are actually responsible for the presence of multiply-related concepts in the UMLS in a quarter of cases. The manual analysis was useful for explaining the conceptualization difference in relations between terms across source vocabularies. Finally, the exploitation of source relationships was helpful for understanding why some source terminologies describe multiple relations between a given pair of terms.