VariCat: An interactive variable categorization program to streamline integration of archival datasets (Preprint)

BACKGROUND Rigorous clinical research on conditions such as older-age bipolar disorder (OABD) patients is imperative to guide prevention and treatment efforts. Large-scale prospective studies are challenging and resource-intensive; combining existing, smaller datasets allows for powerful analyses that are globally representative. The Global Aging & Geriatric Experiments in Bipolar Disorder (GAGE-BD) consortium, which integrates and harmonizes archival data from sites across the world, is one such effort. This requires development of approaches to aid in classifying the domains covered by variables contained in contributed datasets and to set the stage for optimal and efficient harmonization of variables across instruments and data structures. OBJECTIVE To develop a domain-specific ontology for systematically classifying variables in older-age bipolar disorder (OABD) research. To design and implement an computational tool (VariCat) to standardize variable categorization and capture metadata crucial for integration and harmonization across heterogeneous datasets. To demonstrate the capability of this framework via a case study in enabling scalable data integration and harmonization of archival clinical datasets across multiple study cohorts and sites. METHODS Based upon our systematic review of measures used in 53 studies of OABD, and upon data dictionaries and numeric data from 19 cohorts comprising over 1400 individuals with and without bipolar disorder (BD), we developed an ontology of data comprising 20 domains, 102 subdomains, and 129 sub-subdomains for classifying OABD-relevant variables. Subsequently, we created a computer-based interactive program, VariCat, implemented in both Python and REDCap, for data analysts to classify each contributed variable according to its domains/subdomains and to input harmonization-relevant characteristics. RESULTS VariCat provides a way to systematically classify and harmonize heterogeneous variables across datasets by generating a structured metadata repository. Across 37 studies, this approach identified domain coverage (e.g., cognitive variables present in 19 studies) and enabled harmonization of 212 variables into an integrated dataset spanning 20 sites. In a single-site application (UCSD Inflammaging dataset), VariCat categorized all variables and revealed domain distributions (e.g., physical health, illness severity, and cognition as the most represented domains). Across datasets, the framework identified commonly used instruments (e.g., Trail Making Test Part B) and enabled systematic identification of data gaps and harmonization opportunities. CONCLUSIONS We developed and employed a bipolar disorder-specific ontology based on existing datasets and literature and incorporated this into a variable categorization program available in two different formats. This work will help better characterize OABD and also benefit the growing number of international consortia that study psychiatric disorders using archival and/or prospective data. CLINICALTRIAL NA