To improve cancer prognosis and survival, it is crucial to gain a comprehensive view of potential risk factors (RFs) associated with cancer outcomes (e.g., stage of diagnosis, cancer survival). Guided by the National Institute on Minority Health and Health Disparities (NIMHD) Research Framework, cancer outcomes are influenced by RFs from multiple levels (e.g., individual, inter-personal) and multiple domains (e.g., biological, behavioral). Prior research on RFs of cancer survival, however, has primarily focused on RFs from the individual level (e.g., tumor characteristics) due to the lack of integrated datasets that contain multi-level, multi-domain RFs. It is important to pool RFs from heterogeneous data sources, so that we can examine as many RFs as possible in a multi-level integrative data analysis (IDA). However, RF selection and data integration are inconsistently performed and poorly documented in current cancer research, which threatens scientific reproducibility. Therefore, in this paper, we developed a preliminary reporting protocol for RF variable and data source selection based on our previous cancer survival research. Our protocol is informed by NIMHD framework that provides guidance and promotes structural thinking on identifying multi-level cancer RFs. Further, we propose an ontology-based approach to document RF variable and data source selection so that it is (1) explicitly modeled with a shared, controlled vocabulary, (2) understandable to humans and executable by computers, and (3) adaptive to changes when the process being refined.
In the United States (US), as the 2nd leading cause of death, cancer is responsible for 1 in every 4 deaths [1]. The lifetime probability of being diagnosed with cancer is 39.7% and 37.6% for men and women, respectively [2]. To improve prognosis and survival, the first and most crucial step is to gain a comprehensive view of potential risk factors (RFs) associated with various cancer outcomes such as the stage of diagnosis (the most important prognostic factor [3, 4]) and survival.
Recognized by the National Institute on Minority Health and Health Disparities
(NIMHD) Research Framework [5], individuals are embedded within the larger social
system and constrained by the physical environment they lived in. Within this
framework, cancer outcomes are influenced by RFs from multiple levels (i.e., individual,
interpersonal, community, and societal) and multiple domains (i.e., biological,
behavioral, physical/built environment, sociocultural environment, and healthcare system).
Prior research on RFs of cancer outcomes, however, has primarily focused on factors
from the individual level (e.g., tumor characteristics) due to the lack of integrated
datasets that contain multi-level, multi-domain RFs. Very few studies have explored
contextual-level RFs (e.g., access to health care services); and certainly no study has
comprehensively explored all possible RFs together. To do so, it is important to pool RFs
from heterogeneous data sources through data integration, so that we can examine as
many RFs as possible in a multi-level integrative data analysis (IDA).
However, RF selection and data integration are inconsistently performed and poorly
documented, threatening transparency and reproducibility. When reporting research, it
is critical to document the steps that were followed to select, integrate, and process data;
so that others can repeat the same steps and reproduce the findings. In this paper, based
on our previous experience with multi-level IDAs [6], we developed a preliminary
reporting protocol for RF variable and data source selection. Our protocol is informed
by the NIMHD framework that provides guidance on identifying multi-level RFs.
Further, we propose an ontology-based approach so that the selection process is (
In a previous study, we assessed the effect of data integration on predictive ability of cancer survival models [6] and created a semantic data integration (SDI) framework [7] to pool multilevel RFs from heterogenous data sources to support IDA. Table 1 lists the selected RFs and their data sources. Through the case study, a number of variable selection and data integration steps need to be clearly documented. For example, area-rurality status of an individual’s residency has different representations based on the choice of using either the rural-urban commuting area definition [8] (i.e., 10 levels from rural to metropolitan) or the NCHS urban-rural classification [9] (i.e., a hierarchal schema with 7 categories). It is important to document how rurality was defined as different representations of the same variable or concept have differential impacts on the predictive ability of the survival model. Further, a number of data assumptions were made as the different datasets were collected at different time periods and on different populations. For example, the Florida Cancer Data System data include cancer patients from 1996 to 2010, while the US Census data we used were from the general population in 2010. Thus, we made assumptions that the area-level characteristics derived from the Census data were applicable across different time periods. Without a clear documentation of such assumptions and choices, other researchers generally would not have a clear picture of these data integration nuances. Through this multi-level IDA case study, we realized that to ensure the transparency and reproducibility of our study, the documentation of our multi-level RF selection choices (e.g., individual smoking status vs county-level smoking rate), data source selection (e.g., individual-level data from FCDS and contextual-level data from US Census), integration (e.g., data integration strategies and use cases), and processing steps (e.g., the need to calculate body max index [BMI] using weight and height vs using a calculated BMI field that came with the raw data) in the study are the key elements. Through discussions with expert biostatisticians, data analysts and cancer outcomes researchers, we summarized the typical IDA process and developed a prototype reporting protocol for RF variable and data source selection. Further, we propose to inform the multi-level and multi-domain RF selection process with the NIMHD research framework, so that investigators can structurally and comprehensively identify relevant RFs and data sources in their IDA studies. 2.2
Scope. The scope of ODVDS was to standardize and document the selection, integration and processing steps of RF variables and data sources to support IDAs guided by the NIMHD research framework for cancer outcomes research.
Approach. Using the basic formal ontology (BFO) as the top-level ontology, the ODVDS was first developed with a top-down approach, where we started by identifying candidate entities (classes and relations) based on the reporting protocol for RF variable and data source selection. Following best practices, we reviewed existing widely accepted ontologies and reused their classes and relations identified using the NCBO BioPortal [10]. We also took a bottom-up process that started with creating the definition of the most specific classes and then subsequent grouped similar classes into more general concepts. The bottom-up approach helped us determine what new classes and relations are needed to fully represent the IDA process. 3 3.1
Informed by the NIMHD research framework, our preliminary reporting protocol
consists of two parts as shown in Fig 1(a), reporting (
In this work, we developed (