An ontology-guided semantic data integration framework to support integrative data analysis of cancer survival

As the second leading cause of death, cancer is responsible for one in every four deaths in the United States [1]. In 2017, there were approximately 1.68 million new cancer cases and 600 thousand cancer deaths in the US [2] estimated by the American Cancer Society. When first diagnosed with cancer, patients ask about their prognosis, whether their cancer is relatively easy or more difficult to treat, and the likelihood of survival. There is a huge variation in survival between cancer types, stages, age groups, races/ethnicities, genders, and many other factors. For example, among some of the frequently diagnosed cancers, including lung, colorectal, breast, and prostate cancers, the 5-year overall survival rates are 18.3, 64.9, 89.7, and 98.3%, respectively [3]. These rates are much worse when the cancer is metastasized, which are 4.5, 13.9, 26.9, and 29.8% for the same types of cancers, respectively [3].

To improve cancer survival rates and prognosis, one of the first steps is to improve our understanding of contributory factors associated with cancer survival. Priori research such as the National Institute on Minority Health and Health Disparities (NIMHD) Research Framework [4] and the social-ecological model [5] recognizes that individuals are embedded within the larger social system and constrained by the physical environment they lived in. Thus, the determinants of individuals’ health span across different domains of influence (i.e., biological, behavioral, physical/built environment, sociocultural environment, and healthcare system) as well as different levels of influence (i.e., individual, interpersonal, community, and societal). Within these frameworks, cancer survival is influenced by multiple factors from multiple levels and multiple domains. At the individual level, cancer survival is influenced by not only cancer stage of diagnosis and treatment, demographics, and financial status, but also risky health behaviors such as smoking, alcohol drinking, and physical inactivity. For example, cigarette smoking is by far the most important risk factor for lung cancer; 80% of lung cancer deaths in the US were caused by smoking [2]. Beyond individual-level factors, occupational or environmental exposure to secondhand smoke, air pollution, radiation, and some organic chemicals are also significant risk factors. Further, at the contextual level, cancer survival is influenced by public policies that influence health care delivery which could impact patients’ travel distance to the treatment facility [6].

Prior epidemiologic research on cancer survival in the US, however, has primarily focused on contributory factors from the individual level due to limited data availability. Very few studies have explored contextual factors, and certainly no study has explored all possible factors together. Most of these analyses used data from a single source, such as data from a hospital (e.g., electronic health records, EHRs), a cancer registry (e.g., the Surveillance, Epidemiology, and End Results, SEER registry) or administrative claims systems (e.g., data from Centers for Medicare and Medicaid Services, CMS) [7–10]. SEER is an extremely popular data source for studying cancer survival [8–10]. However, it is important to pool heterogeneous data sets with variables beyond the individual level for integrative data analysis (IDA) that simultaneously examine as many cancer survival predictors as possible (i.e. top down approach to the model building) so that confounding effects and interactions among predictors can be fully understood. For example, the linked SEER-Medicare data give us a more complete picture of cancer patients beyond their cancer status with other clinical characteristics such as comorbidity as well as their healthcare utilization patterns [11–14]. Nonetheless, the ability to integrate risk factors of more domains and levels from other data sources such as socioeconomic status of the community from US Census data and community smoking rate from the Behavioral Risk Factor Surveillance System (BRFSS) will further advance our understanding of the determinants of cancer survival.

Nevertheless, researchers are faced with key challenges when integrating data from different sources. Data integration is a daunting task because data from different sources can be heterogeneous in syntax (e.g., file formats, access protocols), schema (e.g., data structures), and semantics (e.g., meanings or interpretations). The effort required to connect different sources is substantial due to lack of clear definitions (i.e., data semantics) of variables, measures, and constructs. Many traditional data integration techniques have been used on large scale in biomedical research [15–17], such as rule-based links (i.e., link variables from different data sources directly base on the names and definitions), data warehouses (i.e., create a new system to store a copy of the data from difference data sources, and manage the data separately from the original data systems) and ad-hoc query optimizers (i.e., re-phrasing a user’s query into multiple subqueries according to the structures of individual distributed databases) and federated middleware frameworks (i.e., link multiple applications and user interfaces to multiple data sources, act as the overarching facade across multiple applications). However, all these traditional methods did not consider the semantic knowledge, which intend to integrate information based on the meaning of the data elements. For example, how to distinguish synonyms, homonyms and related terms (e.g., different representations of the same disease using different coding standards) across different data sources. Therefore, adopting a semantic data integration approach, we propose to generate a universal conceptual representation of “information” to bridge the data heterogeneities across different sources. The “information” includes not only data elements but also their relationships, via “ontologies”. An ontology is a computational representation of a domain of knowledge based upon a controlled, standardized vocabulary for describing entities and the semantic relationships between them [18–21]. The use of ontologies can facilitate data integration in many ways, including metadata representation, automatic data verification, global conceptualization, support for high-level semantic queries, and extend beyond traditional approaches of using common data elements (CDEs) and common data models (CDMs) [22–24], especially in the biomedical domain [15, 25].

Marenco et al. developed a Query Integrator System (QIS) to address robust data integration from heterogeneous data sources in the biosciences in 2004 [26]. An ontology server was used in QIS to map data sources’ metadata to the concepts in standard vocabularies [26]. Cheung et al. developed a prototype web application called YeastHub based on a Resource Description Framework (RDF) database to support the integration of different types of yeast genome data in different sources in 2005 [27]. Lam et al. used the Web Ontology Language (OWL) to integrate two heterogeneous neuroscience databases [28] in 2005. In a follow-up study, Lam et al. designed AlzPharm that used RDF and its extension vocabulary, RDF Schema (RDFS), to facilitate both data representation and integration [29]. Smith et al. built the LinkHub system leveraging Semantic Web technologies (i.e., RDF and RDF queries) to facilitate cross-database queries and information retrieval in proteomics in 2007 [30]. In 2008, Shironoshita et al. introduced a query formulation method to execute semantic queries across multiple data services in the cancer Biomedical Informatics Grid (caBIG), named Semantic caBIG Data Integration (semCDI). Mercadé et al. developed an ontology-based application called Orymold for dynamic gene expression data annotation, integration and exploration in 2009. Based on the QIS [26], Luis et al. designed an automated approach for integrating federated databases using ontological metadata mappings in 2009 [31]. Chisham et al. created the Comparative Data Analysis Ontology (CDAO) and developed the CDAO-Store system to support data integration for phylogenetic analysis in 2011 [32]. Kama et al. built a Data Definition Ontology (DDO) using the D2RQ (i.e., a platform to provide RDF-based access over relational databases) for accessing heterogeneous clinical data sources [33]. Pang et al. developed BiobankedConnect to speed up the process of integrating comparable data from different biobanks to get a pooled data using ontological and lexical indexing in 2014 [34]. Ethier et al. designed the Clinical Data Integration Model (CDIM) based on the Basic Formal Ontology (BFO) [35] to support biomedical data integration in 2015 [36]. Mate et al. proposed an ontology-based approach to organize and describe the medical concepts of both source and target systems in order to integrate the data across different clinical and research systems [37]. Livingston et al. created an integrated knowledge base of biomedical data from multiple sources, called KaBOB, based on Open Biomedical Ontologies [38]. In 2016, Liang et al. proposed an ontology-oriented approach to represent the relations between genes, drugs, phenotypes, symptoms, and diseases from multiple information sources in aiding the analysis of psychiatric drug repurposing [39]. Similar to our approach, Kock-schoppenhauer et al. used the ontology-based data access (OBDA) model and the Ontop framework to access relational clinical databases with SPARQL queries [40]. However, most of these existing semantic data integration systems and frameworks have focused on 1) the harmonization and alignments of data elements using semantic resources; 2) creating tailored ad hoc resources for specific use cases that may not be generalizable; and 3) the integration of data from similar data sources (e.g., data from different electronic health record systems) and addressing the syntactic (i.e., data formats) and schematic (i.e., data models) heterogeneity. Very few studies have fully leveraged the reasoning ability provided by ontologically structured data. And none of the studies has used ontologies as a knowledge representation tool to document the data integration process.

This paper describes a case study of semantic data integration linking five data sets that cover both individual and contextual level factors for the purpose of assessing the association of predictors of interest with cancer survival. The main contribution of our work is that we applied an ontology-based data integration framework to integrate both individual and contextual level factors to facilitate integrative data analysis (i.e., pool heterogeneous data sets). The use of ontologies can facilitate data integration in many ways and extend beyond traditional data integration approaches. Unlike existing ontology-driven data integration methods, our study focused on encoding the different data integration scenarios explicitly using a formal and computational model with a shared vocabulary—the Ontology for Cancer Research Variables (OCRV). Our goal is not only to make the data integration process easier, but also to facilitate documentation and communication of the data integration processes between scientists. This is significant for research rigor, transparency, reproducibility as well as data reusability.

Our overall goal is to facilitate the data integration needs of a theory-driven multi-level IDA of cancer survival informed by the NHMID Research Framework and the socio-ecological model with an ontology-based semantic data integration approach.

Overview of a semantic data integration pipeline

Our semantic data integration workflow is based on an ontology-based data access (OBDA) framework demonstrated in Fig. 1. The first step of semantic data integration is to construct synthesized, integrated descriptions (i.e., a global ontology) of the information coming from multiple sources. An ontology—the Ontology for Cancer Research Variables (OCRV)—in our case, is a metadata representation of the data elements and their semantic relationships in a both human- and machine-understandable structure. After building the ontology, an OBDA model was created—using semantic mapping axioms—to link the source data elements to the entities in OCRV. Given a global view of available data from different sources, a user can pose data (integration) requests for the selected variables (and desired representations) against our data integration pipeline, which converts the requests to a set of SPARQL queries. Based on the semantic mapping axioms defined in the OBDA model, Ontop’s Quest—a SPARQL query engine—can translate a SPARQL query over the ontology into a union of sub-queries over the data sources. The integration of the sub-query results constitutes the answer to the semantic query. The sub-queries are subject to the structure of source schemas, and often expressed in the native query languages of the sources (e.g., Structured Query Language, SQL commonly used for relational databases).

We used the Ontop platform that provides an infrastructure for querying relational databases through ontologies [55]. Within Ontop, entities (i.e., classes and properties) in the ontology are mapped to the data elements (and the relationships between the data elements) in the databases and presented as virtual Resources Description Framework (RDF) graphs. Subsequently, we can then query the virtual RDF graphs via SPQRAL queries.

Constructing an ontology for Cancer research variables (OCRV)

Scope

The OCRV was built for integrating and unifying multi-level predictors of cancer survival across heterogeneous data sources. The OCRV covered a broad range of individual- and contextual-level factors, as shown in Table 1. We used the Basic Formal Ontology (BFO) [35] as the upper-level ontology, and imported the NCI Thesaurus (NCIt) and the Time Event Ontology (TEO) as the foundation for creating the OCRV.

Approach

Using the BFO as the overarching organization, the OCRV was developed with both top-down and bottom-up strategies to catalog relevant entities. A top-down approach was followed to identify candidate entities based on the predictors (i.e., identified through a comprehensive literature review of existing predictors related to cancer survival guided by the NIMHD Research Framework and the social-ecological model) used in the cancer survival analysis. In addition, a review of existing widely accepted ontologies (using the NCBO BioPortal [56]) was conducted to find the relevant entities that can be reused in OCRV. As shown in Table 1, many of the entities that we needed are captured in the NCIt. A bottom-up approach was then used to characterize the entities that have been identified in the top-down approach. We examined the data sources that contained these risk factors, especially the metadata (i.e., the table structures of the relational database). We inspected how these risk factors were expressed in the relational source databases as well as the relationships among the raw data elements, and determined what additional entities and relations were needed to fully represent these risk factors in OCRV.

More importantly, we considered the specific data integration use cases for supporting the multi-level IDA of cancer survival. For example, in our cancer survival analysis, we only needed to model ‘marital status’ as ‘single’ vs. ‘married’, while the raw data has a more fine-grained categorization (e.g., ‘widowed’ and ‘divorced’). Thus, we constructed our ontology to support the data analysis needs by declaring ‘widowed’ and ‘divorced’ as subclasses of ‘single’.

Implementation

We used Protégé 5 [57] to construct the ontology. We worked collaboratively, finalizing the list of entities based on each one’s domain expertise (e.g., oncology, cancer prevention and outcomes, and cancer population science). An example of the raw data records annotated with the OCRV is shown in Fig. 2.

The goal of our data integration tasks was to link predictors from different data sources to generate a single pooled dataset for cancer survival analysis. Thus, we created a data integration pipeline using the Ontop OWL Java application programming interface (API) [58] to translate user requests into SPARQL queries and to organize the query results into an analytic format. The OWL API [59] is a reference implementation for interacting with OWL ontologies. The Ontop OWL API extended the OWLReasoner interface in the OWL API to support SPARQL query answering against relational databases. As required by our data analysis models, the final results were organized into a data table (i.e., a matrix), where each row represented a patient’s cancer diagnosis record (as one patient can have multiple cancer diagnoses in the FCDS data), and each column represented a cancer risk factor.

Our experience in building an ontology-based data integration approach for linking heterogeneous datasets for multi-level cancer survival analysis has demonstrated the feasibility of using semantic data integration to resolve semantic, syntactic, and schematic heterogeneities across different data sources.

The ability to encode data processing and data integration knowledge in the ontology

As discussed in details above, we leveraged the OWL restrictions to properly define the ‘ocrv:BRFSS current smoker’ in OCRV and used object properties to enable grouping of raw categorical variables into different subgroups based on the data analysis needs. These two examples illustrated the ability to reason with ontologically structured data to fulfill data processing and data integration needs. The same logics can indeed be achieved with having a human explicitly write out the specific queries in SQL. And, ultimately in our system, the SPARQL queries were translated into low-level SQL queries. However, the key difference is that in our case the computer was able to reason with the knowledge encoded in the ontology and generate the proper SQL queries rather than having a human to reason with the data and manually construct the SQL queries, which often time is error-prone and labor-intensive. However, not all data manipulation processes can be easily encoded in the ontology. Many of the data transformation procedures were achieved through built-in functions in the SPARQL queries. For example, the process of extracting the year of diagnosis from date of diagnosis had to be realized in SPARQL as shown in Fig. 5.

Another example of needing computing capability in addition to the logics encoded in the ontology is to calculate the average smoking rate for each Florida county using the raw data in BRFSS. The calculation is straightforward: for each county, the average smoking rate equals to the sum of the BRFSS final weights for all current smokers divided by the sum of the final weights for all BRFSS population in the county. As shown in Fig. 3, we can identify who is BRFSS defined current smoker. A county code and a BRFSS final weight were also attached to each BRFSS respondent as shown in Fig. 10. With these information, we can easily calculate the average smoking rate for each county. Nevertheless, these computations were programed in the data integration pipeline Java code.

Data quality and consistency checks of the source data using the ontology

Another advantage of using ontology as a data model is to conduct automated data quality and consistency checks of the source data, because the dependencies among and constraints of the data elements are explicitly modeled in the ontology. For example, FCDS has two related variables: ‘age at diagnosis’ and ‘date of diagnosis’. As shown in Fig. 11, we explicitly modeled the relationship and constraints between these two entities, as the ‘ncit:age at diagnosis’ can be calculated from the ‘ocrv:date of diagnosis’ and ‘ncit:birth year’. We can then compare the raw ‘age of diagnosis’ and the computed age of diagnosis to exam data consistency. Further, we also used OWL restrictions to enforce the required formats and ranges of the data elements. Equipped with automated data consistency checks, data analysts shall follow best practices in dealing with data quality issues [62].

We have presented an ontology-based semantic data integration approach for multi-level integrative data analysis of cancer survival. Our approach solves key data integration challenges: (1) not only provide clear definitions of the data elements using standardized common vocabulary, but also explicitly expressed the relationships among the variables, (2) be able to clearly document data processing and integration procedures, (3) leverage ontologically structured data to infer and derive the required data elements and formats automatically, (4) provide a convenient mechanism for documenting data quality and consistency checks, and most importantly (5) allows scientists to clearly document and communicate their data manipulation processes, which is significant for research rigor, transparency, reproducibility as well as data reusability.