Unlike recent biomedical terminologies, the International Classification of Diseases (ICD) does not state any explicit associations between a given disease and the corresponding anatomical structure(s). As a consequence, clinical repositories coded with ICD cannot be searched by anatomical structure. The objective of this work is to find associations between diseases from ICD10 and anatomical structures. Toward this end, we investigated three approaches (symbolic, lexical, and statistical) which exploit various features of the Unified Medical Language System (UMLS). We evaluated these approaches according to i) the consistency of resulting anatomical concepts with the high-level anatomical concept(s) identified for the chapter in which the disease is listed; and ii) the validity of resulting anatomical concepts assessed manually. We show that the symbolic approach is both the most productive and the most accurate approach.
Biomedical terminologies developed recently often use formalisms based on
description logics for their representation
In contrast, although terminologies such as the International Classification of Diseases (ICD) organize diseases in hierarchies based for the most part on body systems, no explicit association between a given disease and the corresponding anatomical structure(s) is stated in ICD. Therefore, clinical data coded with ICD cannot be precisely searched by anatomical structure.
The objective of this work is to discover relations between ICD10 diseases and anatomical structures. Three approaches (symbolic, lexical and statistical) are investigated.
Terminological resources have been used for extracting or inferring relations
among concepts. The compositional nature of terms has been exploited for identifying
hierarchical and associative relations (e.g.,
2.1
The International Classification of Diseases is an international standard diagnostic
classification for epidemiology, health management, and clinical use
The Unified Medical Language System® (UMLS®)
Three distinct approaches are investigated in this study (symbolic, lexical, and statistical), exploiting various features of the UMLS. The symbolic approach uses the relations defined between ICD10 concepts and anatomical concepts in the source vocabularies included in the UMLS. The lexical approach attempts to identify anatomical concepts in the names of ICD10 concepts. Finally, the statistical approach relies on co-occurrence information. This investigation is restricted to the diseases from the 12 chapters of ICD10 organized around a given body system (Table 1). In this section, we present the three approaches, as well as the evaluation of the consistency and validity of the results. 3.1
In biomedical terminologies, the domain and range of defined relationships is constrained to specific semantic types as part of their definition. In contrast, the range of undefined relationships is unrestricted and must be constrained by users. Because of this difference, two strategies were investigated in the symbolic approach. 3.1.1
Starting from a given disease concept, we search for target concepts connected to the disease concept by defined relationships whose range is anatomy, including disease_has_associated_anatomic_site and finding_site_of. 3.1.2
When all the relationships of a given disease source concept are exploited, we constrain the range of these relationships to anatomy by requiring that target concepts be associated with the semantic group (SG) Anatomy, i.e. categorized by a semantic type (ST) belonging to the SG Anatomy.
As illustrated in Figure 1, within each strategy, relations to anatomical concepts are first searched directly from the disease concept (case a). Failing to find any, we start the search again from concepts in mapping relation to the source disease concept (case b). In practice, we allow the traversal of relations whose name contains the pattern “mapp” (e.g., mapped_from and mapped_to) before searching for related anatomical concepts. If none is found at this stage, we start the search for anatomical concepts again, but from the parent concept(s) of the source disease concept (case c). 3.2
This approach attempts to identify anatomical structures in the name of ICD10
concepts. It uses MetaMap, a linguistically-motivated named entity recognition
program specially developed for biomedical entities, which extracts Metathesaurus
concepts from text
In order to assess the consistency of the anatomical concepts identified for a given disease, we determine whether this anatomical structure is related to the body system corresponding to the chapter in which this disease is listed. One of the authors (OB) associated each chapter (or subdivision thereof) with one or two high-level anatomical concepts (e.g., Cardiovascular system (C0007226) for the chapter “Diseases of the circulatory system”). We computed the list of all descendants of each high-level anatomical concept. The association between a disease and an anatomical concept is deemed consistent if the anatomical concept is found among the descendants of the high-level anatomical concept(s) identified for the chapter in which the disease is listed (Figure 2).
ICD10 chapter D3
D4
High-level anatomical
concept A1
A2 D1
D2
A3
A4 4
4,391 ICD10 concepts amenable to our method were selected. The first three columns of Table 1 show the repartition of ICD10 concepts by chapter (or subdivision thereof). Detailed results for each approach are displayed Table 2. 4.1
Overall, 4,342 ICD10 concepts were associated with at least one anatomical concept. The resulting anatomical concepts are not systematically the same with both strategies.
For example, the ICD10 concept Idiopathic gout (C0149896) was associated with distinct anatomical concepts according to the strategy. As strategy A1 could not find any anatomical concept for Idiopathic gout through a defined relation, it first selected “mapped” concepts: Gout (C0018099) and Gouty nephropathy (C0391820), themselves related to Joints (C0022417), Connective and Soft Tissue (C1516798), and Kidney (C0022646) through defined relations from SNOMED CT. Conversely, strategy A2 selected one unique concept, Articular system (C0149896), through an undefined relationship (other related) with Idiopathic gout.
All ICD10 concepts successfully associated with anatomical concepts through the first strategy are included in those found through the second strategy. For 82 ICD10 concepts, only the second strategy was able to find related anatomical concepts. An example is the concept Lymphoedema, not elsewhere classified (C0494630) for which only the second strategy proposed Lymph nodes (C0024204).
Finally, it is worth noting that the two alternatives proposed when no anatomical concept could be found directly through the ICD10 concept itself (cases b and c) contributed to improve the results. For example, Embolism and thrombosis of iliac artery (C0155755) was associated with the anatomical concepts Structure of iliac artery (C0020887) and Arteries (C0003842) through its mapping relation to the concepts Thrombosis of iliac artery (C0235518) and Embolism and thrombosis of other specified artery (C0155754). On the other hand, the ICD10 concept Nonrheumatic aortic valve disorders (C0003502) was associated with the anatomical concepts Heart Valves (C0018826), Heart (C0018787), and Aortic valve structure (C0003501) through its parent concepts Heart valve disease (C0018824), Other heart disease C0178273, and Aortic valvular disorders (C1260873). 4.2
Overall, 2,987 ICD10 concepts exhibit at least one anatomical concept in their name. An example is the ICD10 concept Biliary Fistula (C0005417) associated with the two anatomical concepts Bile (Bile fluid) (C0005388) and Bile Duct (Bile duct structure) (C0005400) through the synonymous term Fistula of bile duct from the source vocabulary SNOMED CT. 4.3
Overall, only 783 ICD10 concepts co-occur with at least one anatomical concept. An example is Ainhum (C0001860) which is a painful constriction of the base of the fifth toe. This ICD10 concept co-occurs with the three following anatomical concepts: Fingers (C0016129), Hallux structure (C0018534), and Toes (C0040357). 4.4
As shown in Figure 3, the symbolic approach was able to find anatomical concepts for all but 49 of the 4,391 diseases from ICD10 (98.9%). The lexical approach found anatomical concepts for 2,987 disease concepts, including 23 of the 49 concepts for which the symbolic approach had failed. For example, the ICD10 concept Hereditary nephropathy, not elsewhere classified, minor glomerular abnormality (C0868873) has been associated with the anatomical concept Glomerular (Kidney Glomerulus) (C0022663) only through the lexical approach. No new anatomical concepts could be found through the statistical approach. Overall, only 26 ICD10 concepts could not be associated with any anatomical concepts through any of the approaches, including Defects of catalase and peroxidase (C0494349) and Other immunodeficiencies (C0494262). anatomical concepts associated were consistent and for 466 diseases, none of the concepts associated was consistent.
As an illustration, the concept Central pontine myelinolysis (C0206083) is listed in the chapter “Diseases of the nervous system”. Therefore, the associated anatomical concepts were expected to be found among the descendants of the corresponding highlevel anatomical concept, Nervous system structure (C0027763) (last column of Table 1). The consistency results are as follows: • Strategy A1: Two anatomical concepts were selected, namely Pontine structure (C0032639) and Neuraxis (C0927232), both consistent with Nervous system structure;
Strategy A2: In addition to the two concepts obtained with strategy A1, In Blood (C0005768) was also selected but is inconsistent with Nervous system structure;
Approach B: Only the (C0032639) is found (consistent);
Approach C: 26 anatomical concepts co-occur with Central pontine myelinolysis, of which 19 are consistent with Nervous system structure, including Basal Ganglia (C0004781) and Cerebellum (C0007765). Incompatible concepts include Tongue (C0040408) and Structure of pituitary fossa (C0036609).
anatomical concept Pontine structure 4.6
were valid and for 2 diseases, none of the concepts associated were valid. Finally, no anatomical structures was identified for 3 diseases.
Overall, the best results are obtained through the symbolic approach. Not only is this approach more productive (i.e., it associates more ICD10 concepts with anatomical structures), but it also exhibits superior consistency and validity (more concepts have a percentage of 100%). As expected, strategy A1 (based on defined relations) has better precision than strategy A2 (using all relations), but its recall is slightly lower. The lexical approach could not bring much improvement over the results of the symbolic approach. The statistical approach was disappointing as it did not find any anatomical structures not already found by the two other approaches.
The two evaluations in terms of consistency and validity sometimes lead to the same conclusions. In some cases, however, the evaluations yield different results. Indeed, the concept Hereditary haemorrhagic telangiectasia (C0039445) was associated with four anatomical concepts through strategy A2: i) Cardiovascular system (C0007226) and Vascular System (C0489903) which are consistent and valid; ii) In Blood (C0005768) which is inconsistent and invalid; iii) Microscopic skin vascular structure (C1283407) which is consistent but invalid. Surprisingly, consistency evaluation was sometimes less permissive than validity evaluation. For instance, Acute erythraemia and erythroleukaemia (C0001317) was associated, among other anatomical concepts, with Bone Marrow (C0005953) which was rightly deemed valid and was wrongly inconsistent. 5.2
Of the 12,320 ICD10 codes, only 4,265 (35%) are amenable to the kind of processing investigated in this study. For the UMLS concepts corresponding to the remaining ICD10 codes, no relation to anatomical concepts was represented in any vocabulary, and no anatomical entity could be identified in the names of these concepts. In practice, the methods proposed here could, at best, be used to complement other methods or expert knowledge.
The evaluation does not include every ICD10 chapter but it must be noted that chapters like “Mental and behavioural disorders” and “Certain infectious and parasitic diseases” are evaluated indirectly through the dagger-asterisk representation in ICD10. Indeed, some diseases are listed twice; in the chapter corresponding to their etiology (dagger) and in the chapter corresponding to their manifestation (asterisk). For example, the disease Alzheimer's disease (G30+) is listed primarily in chapter VI “Diseases of the nervous system”, while its manifestation Dementia in Alzheimer's disease (F00*) is found in chapter V “Mental and behavioural disorders”. Thus, a large number of diseases from chapters not investigated in this study have been considered indirectly through other chapters.
The recall of the lexical and statistical approaches is quite low. While it could likely be increased by exploiting mapping and parent relations as we did in the lexical approach, we believe that this extension would adversely affect the precision of our results.
Finally, depending on the application, the level of granularity of the anatomical structures associated with ICD10 diseases may or may not be appropriate. For example, as mentioned earlier, Idiopathic gout is appropriately associated with joints through the anatomical concepts Joints and Articular system. However, terminological knowledge from the UMLS fails to capture the fact that, in most cases of gout, joint inflammation affects the toes. Moreover, Idiopathic gout is also associated with Kidney, while kidney manifestations are much less frequent. 5.3
The coding system used by French pathologists supports the retrieval of specific anatomical structures from the code itself. However, ICD10 lacks this feature. Cancer registries drawing from repositories coded with ICD10 could be queried precisely with respect to anatomical structures if the relations between ICD10 codes and anatomical structures were asserted as we suggest here. 6
This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Library of Medicine (NLM). ICD10 Chapter / Subchapter label Malignant neoplasms of lip, oral cavity and pharynx Malignant neoplasms of digestive organs Malignant neoplasms of respiratory and intrathoracic organs Malignant neoplasms of bone and articular cartilage Melanoma and other malignant neoplasms of skin Malignant neoplasms of mesothelial and soft tissue Malignant neoplasm of breast Malignant neoplasms of female genital organs Malignant neoplasms of male genital organs Malignant neoplasms of urinary tract Mal. neoplasms of eye, brain and other parts of central nervous system Malignant neoplasms of thyroid and other endocrine glands Malignant neoplasms of lymphoid, haematopoietic and related tissue Diseases of blood and blood-forming organs...
Endocrine, nutritional and metabolic diseases Diseases of the nervous system Diseases of the eye and adnexa Diseases of the ear and mastoid process Diseases of the circulatory system Diseases of the respiratory system Diseases of the digestive system Diseases of the skin and subcutaneous tissue Diseases of the musculoskeletal system and connective tissue Diseases of the genitourinary system
Reference anatomical concepts Oral cavity, Pharyngeal structure Gastrointestinal system Respiratory System, Intrathoracic organ Skeletal system Skin Mesothelium, Soft tissue Breast Female genitalia Male Genital Organs Urinary system Eye, Neuraxis Endocrine system Hematopoietic System, Lymphoid organ structure Blood, Hematopoietic System Endocrine system Nervous system structure Eyes and eye appendages Ears and mastoid cells Cardiovascular system Respiratory System Gastrointestinal system Integumentary system Musculoskeletal System Genitourinary system