ClinEpiDB: an open-access clinical epidemiology database resource encouraging online exploration of complex studies

Ethical statement

The ClinEpiDB platform has received approval from the University of Pennsylvania under IRB#7, Protocol #828806. All studies included in ClinEpiDB have undergone ethical approval at applicable institutions prior to data collection (ClinEpiDB is generally not involved in this process). Data providers also obtain approval from their institutions to have their data hosted on ClinEpiDB. Community engagement programs have not yet been undertaken to assess study participants’ attitudes towards data sharing via the ClinEpiDB platform.

Implementation

ClinEpiDB integrates studies conducted by various primary research groups and can accommodate a variety of study designs including observational studies (surveillance, cross-sectional, longitudinal cohort, and case-control) and randomized control trials. Researchers supply flat data files along with data dictionaries, data collection forms, and protocols to help contextualize the data. Variables within the data set may contain categorical, continuous, discrete, or free text data.

Once the data are received, a series of files are constructed according to a standard operating procedure to process the variables, map them to ontology terms, and map coded categorical values to the descriptive terms displayed on the website (Figure 1). In rare cases where personally identifying variables – such as participant names and addresses – are included in the files received, those variables are removed before files are stored on a private file system only accessible to the data loading team to ensure participant confidentiality. All systems used to store and process the data are Federal Information Security Modernization Act (FISMA) compliant and undergo regular security review. Variables used solely for data cleaning purposes are also excluded. During initial processing, some standard checks are performed to ensure data are relatively clean, including ensuring that unique identifiers match across files, dropping variables with no data, identifying variables with unexpected values (i.e. character values for an expected numeric variable), and identifying free-text variables that might benefit from standardization and/or translation. The study team is asked for clarification, revision, and input as needed. Creating ISA-based (Investigation, Study, Assay) files for loading often requires merging data files, and data conflicts can also be identified and fixed at that point.

Figure 1. Pipeline for processing studies.

(1) The ClinEpiDB team generates an “allVariables” file from the raw data files, data dictionaries, and data collection forms that contains all variables collected as part of the study and indicates whether each variable will be displayed on the website or not. This file is used to make (2) a “valueMap” file that maps coded categorical values to descriptive terms to be displayed on the website and (3) a “variableMap” file that maps variables to existing ontology terms and labels for display on the website. (4) The “variableMap” file is further processed by the ontology team and new ontology terms are created as needed. (5) All files are passed to the data loading team to pre-process the data, shift dates based on a random number algorithm, and create ISA files to load into the GUS4 database. (6) Once files are loaded, the data appear on an access-restricted website. Any additional searches required by a study are designed and implemented.

To deal with the challenges of integrating distinct studies with highly heterogeneous data while providing user-friendly mechanisms to identify similar variables, we employ an ontology-based approach to generate a unified semantic framework as described in Zheng et al. (2016). Wherever possible, variables are mapped to existing terms drawn from Open Biological and Biomedical Ontologies (OBO) Foundry, which supports interoperable ontologies (Smith et al., 2007) through adhering to published principles. Necessary ontology terms are imported into the OBO Foundry registered VEuPathDB ontology, which is used as a single resource for all terms used in VEuPathDB resources. New ontology terms with VEuPathDB Internationalized Resource Identifiers (IRIs) are created as required. The use of ontologies to represent variables from different studies guides how data are loaded into relational databases and supports presentation of variables on the website to facilitate searching and analysis.

Once the data, ontology, and value mapping files are prepared, the data undergo processing to obfuscate dates to protect participant confidentiality. All dates for a given participant are consistently shifted forward or backward by 0–7 days according to a random number algorithm. While this obfuscation may introduce noise into longitudinal analysis—two events occurring on the same day for different participants will appear to have occurred 0–2 weeks apart—dates will still cluster within epidemiologically relevant timeframes for most analyses. All data are then transformed into an ISA-based format (Sansone et al., 2012) and loaded into a relational database based on the Genomics Unified Schema, version 4 (GUS4) (Davidson et al., 2001) running in an Oracle database management system (DBMS). Build database servers are located at the University of Pennsylvania and production instances are mirrored at the Universities of Pennsylvania and Georgia for redundancy purposes and to ensure uptime. All servers are housed in FISMA-compliant computational facilities, and industry standard backups of all data are performed.

Searches for each study are made available to users in an intuitive user interface (the “Search Wizard”), driven by a series of SQL queries against the GUS4 database (code available on GitHub, see Software availability). These searches vary depending on study design and record types. For example, in the longitudinal PRISM study, users can specifically retain or exclude observations occurring within a specified time relative to another observation through the “Related Observations” step in the Search Wizard (e.g. identify children diagnosed with febrile malaria at least twice within a six-month period). In the GEMS case-control study, users can compare cases to matching controls and choose whether to return data from the selected participants, matching cases/controls, or selected participants plus their matching cases/controls. Implementation of the strategies web development kit (WDK) (Fischer et al., 2011) allows users to construct even more complex queries using logical operators (union, intersection, subtraction) and to save and share search strategies.

Exploration applications for additional data visualization are created with Shiny, an open-source R package for building interactive web applications (RStudio Inc, 2019). The applications are hosted on the website via the Shiny Server Open Source software. SQL queries against the Oracle database identify all variables in the study and their format, which informs which variables appear as options to plot, how to build a custom dichotomous variable (i.e. hemoglobin <=10 mg/dL vs. hemoglobin >10 mg/dL), and how the data are plotted within the applications.

Studies are reviewed by ClinEpiDB staff for quality control and made accessible to primary data providers using a protected internal website to ensure data accuracy and query functionality. Data are only scheduled for public release following data provider approval. Updates to the database are released every two months and can include new studies, features, and/or software updates.

Operation

ClinEpiDB can be accessed via any web browser at https://clinepidb.org. User support is available via the “Contact Us” link and tutorials are accessible via the “Community” drop-down menu at the top of each web page.

Study pages. Clicking a study name on a card on the ClinEpiDB homepage (Figure 2A) or under the “Search a Study” drop-down menu brings up the study page, which provides a description of that particular study’s goals and objectives, methodology, investigators, and links to associated publications. This page also provides links to data collection forms and data dictionaries, which detail variable definitions, allowed values, skip patterns, etc. For studies that require permission to download data, the study page also incorporates a table listing individuals who have been granted access to the data and their brief stated purpose of use.

Figure 2. Using the Search Wizard to explore variables.

(A) Clicking a card study name opens a study page. (B) Clicking on a card search icon initiates a search. (C) The Search Wizard categorizes the variables into discrete steps. The grey buttons let users move between steps. (D) The variable tree contains all variables within that step of the Search Wizard. To subset the data, users can open a variable from this tree. (E) The “Find a variable” search bar searches for variables based on variable names and values across all Search Wizard steps. (F) Continuous data are displayed as a histogram and can be constrained by typing the exact range of values or clicking and dragging the mouse across the range of interest. (G) Clicking cards underneath “Explore Example Searches” opens up examples of searches conducted using the datasets indicated. These searches can be edited. (H) Clicking cards underneath “Explore Visualization Tools” opens up examples of how the exploration applications can be used.

Search strategies. ClinEpiDB permits users to execute searches on epidemiological data sets. Depending on available study data, up to four search types are currently supported that identify Households, Participants, Observations, or Entomology Collections of interest (Table 2). For example, surveillance studies with just a single observation per participant offer only a participant search (e.g. “How many participants presented with both a fever and cough”). Studies with multiple participants from the same household will have household searches as well (e.g. “Which households contained children with asymptomatic parasitemia”). Longitudinal studies permit observation-level searches (e.g. “Identify all observations of children with malaria from houses with unscreened windows”). Within each search, users can subset the data based on any of the variables available through the “Search Wizard”, explore associations between variables, and return an interactive table of selected data.

Table 2. Search types currently available in ClinEpiDB.

Search type	Default steps available via the Search Wizard	Results Table format	Results Table variables
Household	Household Participant Observation	One row per household observation (multiple rows per household if household data was collected longitudinally)	Household-level variables relating to geographic location, dwelling characteristics, socioeconomic status, etc.
Participant	Household Participant Observation	One row per participant	Participant-level variables relating to demographics, enrollment, data summaries, etc. May also include upstream (household-level) variables.
Observation	Household Participant Observation	One row per observation (multiple rows per participant if data was collected longitudinally)	Observation-level variables relating to anthropometry, symptoms, laboratory test results, treatment, etc. May also include upstream (household- and participant-level) variables.
Entomology	Household Entomology collection	One row per entomology collection (multiple rows per household if collections were done in multiple rooms or longitudinally)	Entomology variables relating to mosquito counts and species. May also include upstream (household-level) variables.

From the home page, clicking a search icon on a study card initiates a search (Figure 2B). The “Search Wizard” at the top of the page (Figure 2C) categorizes the data, providing a step-wise approach to selecting data. On the left-hand side of the page, the variable tree presents all variables within that step of the Search Wizard (Figure 2D), while the search bar at the top of the page allows users to search for variables across all Search Wizard steps (Figure 2E). To subset data, users click on a variable of interest (e.g. “Temperature (C)”) and specify desired values. Continuous data are displayed as a histogram and can be selected by typing in a specific range or by clicking and dragging the cursor across the range of interest (Figure 2F). Categorical data (e.g. “Malaria diagnosis and parasite status”) are displayed in a table and can be selected via the adjacent check boxes (Figure 3A).

Figure 3. Adding, editing, and removing filters.

Categorical data are displayed in a table and (A) can be selected via check boxes next to the values. (B) The “Remaining” column indicates the data remaining given all other data selections (including selections in upstream steps), while the (C) Observations column indicates the total counts for all data. (D) For both continuous and categorical variables, data that meet the filter criteria (“remain”) are shown in red on the distribution graph while data that do not meet the filter criteria are shown in grey. (E) Clicking the green filter icon brings up a box that lists all applied filters. (F) Users can click the blue link to edit a filter or the “x” to remove it. (G) The blue button takes the user to the results page.

As data are selected, the data available for other variables in that Search Wizard step and any downstream steps are dynamically updated so the user can visualize the impact of their selection(s) on other variables. The “Remaining” column in the variable tables indicates the data remaining given all upstream filters (Figure 3B), while the column to the immediate right indicates the total counts (Figure 3C). For both continuous and categorical variables, data meeting upstream selection criteria are shown in red on the distribution graph while data excluded by the selection criteria are shown in grey (Figure 3D). Selections can be reviewed, edited, or removed by clicking the green filter icon (Figure 3E) and then clicking the blue link to edit selections for a variable or the “X” to remove it (Figure 3F). Combined with data visualization through bar charts and histograms, this ability to conveniently add, edit, and remove filters makes it possible to rapidly assess the structure of the data and potential associations between variables of interest.

New users wanting to get a sense of what types of searches are possible can choose to view and edit publicly available searches under the “Explore Example Searches” section of the homepage (Figure 2G).

Results page and exploration apps. Data selected as described above are displayed on the Results Page (Figure 4) when the user clicks the blue button at the right-hand terminus of the Search Wizard (Figure 3G). The selected data are displayed as a table (Figure 4) and may be passed to a suite of web applications for additional visualization and analysis. Variables available as columns are based on the type of search performed (Table 2). Histogram icons in the column headers allow users to assess the distribution of the subset of data for that variable (Figure 4A); links in the top right corner allow users to add additional variable columns (Figure 4B) or download the selected data (Figure 4C).

Figure 4. Using the Results Page.

(A) Clicking a histogram icon opens a pop up showing the distribution of data for that variable. (B) The “Add columns” button allows users to change which variables are shown in the table. (C) The “Download” link directs users to a page where they can choose which variables to download. The data subset is based on the selections applied in the Search Wizard. (D) The “Analyze Results” tab leads to a suite of applications for further data visualization.

The “Analyze Results” tab (Figure 4D) leads to a suite of web applications. Three applications are currently available in ClinEpiDB: Distributions, Contingency Tables, and Data Summaries (also accessible under “Explore Visualization Tools” on the homepage; Figure 2H). The Distributions application shows the distribution of any variable in the data set and allows stratification based on other variables. The Contingency Table application generates a 2 × 2 contingency table for two selected variables and calculates a p-value, odds ratio, and relative risk, enabling assessment of associations (note that these statistics should be interpreted with caution as they do not control for confounding or other forms of bias). The Data Summaries application plots a variable of interest over time for longitudinal studies or two variables of interest against each other for non-longitudinal studies. For each app, users can toggle between tabs to define plot parameters, view summary statistics, display a plot grid or individual plots, and obtain help. Drop-down menus allow users to specify which variables to graph and whether to stratify data based on additional variables. Updating parameters automatically regenerates all statistics and plots. By default, the applications utilize the entire study data set, but users may choose to examine data selected in the Search Wizard by stratifying based on search results. The plots may be downloaded, but we encourage users to use the applications primarily for initial data exploration and to make their own graphs for presentations and publications after careful execution of a statistically robust, thorough data analysis, considering any potential biases and statistical assumptions.

Data downloads. Data in ClinEpiDB may be downloaded in two ways. Clicking the “Download” link on the Results Page (Figure 4C) allows users to customize downloads, specifying which variables to retrieve based on the search type (see Table 2). All other variables can be downloaded, and data can be linked across files via observation, participant, and household IDs. Users may specify .txt or .csv formats, both of which can be consumed by most modern data analysis tools. Users can also download the entire data set via the “Download Data” link on the homepage study card and the study page. An ontology term association file links variables to their original study labels so users can reference study data collection forms and data dictionaries to learn more about each variable. Variables are also mapped to ontology terms via Internationalized Resource Identifiers (IRIs) which are included in each column header of the download file (Ong et al., 2017). Following OBO Foundry principles, the terms are reused or requested from existing ontologies when possible but placeholder terms are also created as needed. Once defined, the terms are made public in the VEuPathDB application ontology along with imported terms from other ontologies and are searchable on Ontobee.

Accessibility of datasets

ClinEpiDB is committed to making epidemiologic data sets accessible to global research and biomedical communities while protecting the rights of study participants and data providers. Prior to viewing the website for the first time, users are required to agree to a Data Access and Use Policy outlining expectations regarding data use, protection of participant privacy, and acknowledgement of data providers and ClinEpiDB.

Some studies require data access restrictions at the data provider’s discretion. There are five access levels data providers can choose from that differ in their requirements for users to view aggregate versus disaggregate data (Table 3). Aggregate data are accessible in the Search Wizard and through the exploration applications, while disaggregate data can be found on the results page, individual record pages, and in the download files. Except for studies classified as private, which require approval to see any data, users can see variables and aggregate data for all studies. Note that on occasion, variables that are potential indirect identifiers are loaded into the database but not displayed on the website. These variables are available for download to individuals with access approval.

When a user reaches a restricted section of the website, they are automatically prompted to either log in with a ClinEpiDB account or log in and submit a data access request, depending on the access restrictions. The data access request form requires the purpose for which the requested data will be used, whether the requester has been in contact with the study team, hypotheses and/or research questions, analysis plan, and planned dissemination of results. The request is then sent to data providers for approval. Users are contacted within a few days with any conflicts that are identified or with notification of approval. Once approved, they may view and download that study’s data at any point by logging into their ClinEpiDB account. To ensure transparency and promote collaboration within the wider scientific community, the requestor’s name, organization, request date, and indicated purpose appear publicly on the corresponding study page once approved.