2016 Theses Doctoral
Ontology-based Semantic Harmonization of HIV-associated Common Data Elements for Integration of Diverse HIV Research Datasets
Analysis of integrated, diverse, Human Immunodeficiency Virus (HIV)-associated datasets can increase knowledge and guide the development of novel and effective interventions for disease prevention and treatment by increasing breadth of variables and statistical power, particularly for sub-group analyses. This topic has been identified as a National Institutes of Health research priority, but few efforts have been made to integrate data across HIV studies. Our aims were to: 1) Characterize the semantic heterogeneity (SH) in the HIV research domain; 2) Identify HIV-associated common data elements (CDEs) in empirically generated and knowledge-based resources; 3) Create a formal representation of HIV-associated CDEs in the form of an HIV-associated Entities in Research Ontology (HERO); 4) Assess the feasibility of using HERO to semantically harmonize HIV research data. Our approach was guided by information/knowledge theory and the DIKW (Data Information Knowledge Wisdom) hierarchical model.
Our systematized review of the literature revealed that synergistic use of both ontologies and CDEs included integration, interoperability, data exchange, and data standardization. Moreover, methods and tools included use of experts for CDE identification, the Unified Medical Language System, natural language processing, Extensible Markup Language, Health Level 7, and ontology development tools (e.g., Protégé). Additionally, evaluation methods included expert assessment, quantification of mapping tasks between raters, assessment of interrater reliability, and comparison to established standards. We used these findings to inform our process for achieving the study aims.
For Aim 1, we analyzed eight disparate HIV-associated data dictionaries and developed a String Metric-assisted Assessment of Semantic Heterogeneity (SMASH) method, which aided identification of 127 (13%) homogeneous data element (DE) pairs and 1,048 (87%) semantically heterogeneous DE pairs. Most heterogeneous pairs (97%) were semantically-equivalent/syntactically-different, allowing us to determine that SH in the HIV research domain was high.
To achieve Aim 2, we used Clinicaltrials.gov, Google Search, and text mining in R to identify HIV-associated CDEs in HIV journal articles, HIV-associated datasets, AIDSinfo HIV/AIDS Glossary, AIDSinfo Drug Database, Logical Observation Identifiers Names and Codes (LOINC), Systematized Nomenclature of Medicine (SNOMED), and RxNORM (understood as prescription normalization). Two HIV experts then manually reviewed DEs from the journal articles and data dictionaries to confirm DE commonality and resolved semantic discrepancies through discussion. Ultimately, we identified 2,179 unique CDEs. Of all CDEs, data-driven approaches identified 2,055 (94%) (999 from the HIV/AIDS Glossary, 398 from the Drug Database, 91 from journal articles, and a total of 567 from LOINC, SNOMED, and RxNorm cumulatively). Expert-based approaches identified 124 (6%) unique CDEs from data dictionaries and confirmed the 91 CDEs from journal articles.
In Aim 3, we used the Protégé suite of ontology development tools and the 2,179 CDEs to develop the HERO. We modeled the ontology using the semantic structure of the Medical Entities Dictionary, available hierarchical information from the CDE knowledge resources, and expert knowledge. The ontology fulfilled most relevant criteria from Cimino’s desiderata and OntoClean ontology engineering principles, and it successfully answered eight competency questions.
Finally, for Aim 4, we assessed the feasibility of using HERO to semantically harmonize and integrate the data dictionaries from two diverse HIV-associated datasets. Two HIV experts involved in the development of HERO independently assessed each data dictionary. Of the 367 DEs in data dictionary 1 (D1), 181 (49.32%) were identified as CDEs and 186 (50.68%) were not CDEs, and of the 72 DEs in data dictionary 2 (D2), 37 (51.39%) were CDEs and 35 (48.61%) were not CDEs. The HIV experts then traversed HERO’s hierarchy to map CDEs from D1 and D2 to CDEs in HERO. Of the 181 CDEs in D1, 156 (86.19%) were found in HERO, and 25 (13.81%) were not. Similarly, of the 37 CDEs in D2 32 (86.48%) were found in HERO, and 5 (13.51%) were not. Interrater reliability for CDE identification as measured by Cohen’s Kappa was 0.900 for D1 and 0.892 for D2. Cohen’s Kappas for CDEs in D1 and D2 that were also identified in HERO were 0.885 and 0.688, respectively.
Subsequently, to demonstrate the integration of the two HIV-associated datasets, a sample of semantically harmonized CDEs in both datasets was categorically selected (e.g. administrative, demographic, and behavioral), and D2 sample size increases were calculated for race (e.g., White, African American/Black, Asian/Pacific Islander, Native American/Indian, and Hispanic/Latino) and for “intravenous drug use” from the integrated datasets. The average increase of D2 CDEs for six selected CDEs was 1,928%.
Despite the limitation of HERO developers also serving as evaluators, the contributions of the study to the fields of informatics and HIV research were substantial. Confirmatory contributions include: identification of effective CDE/ontology tools, and use of data-driven and expert-based methods. Novel contributions include: development of SMASH and HERO; and new contributions include documenting that SH is high in HIV-associated datasets, identifying 2,179 HIV-associated CDEs, creating two additional classifications of SH, and showing that using HERO for semantic harmonization of HIV-associated data dictionaries is feasible. Our future work will build upon this research by expanding the numbers and types of datasets, refining our methods and tools, and conducting an external evaluation.
- Brown_columbia_0054D_13617.pdf binary/octet-stream 5.39 MB Download File
More About This Work
- Academic Units
- Biomedical Informatics
- Thesis Advisors
- Bakken, Suzanne R.
- Ph.D., Columbia University
- Published Here
- October 11, 2016