Abstract
INTRODUCTION: Pilot projects ("pilots") are important for testing hypotheses in advance of investing more funds for full research studies. For some programs, such as Clinical and Translational Science Awards (CTSAs) supported by the National Center for Translational Sciences, pilots also make up a significant proportion of the research projects conducted with direct CTSA support. Unfortunately, administrative data on pilots are not typically captured in accessible databases. Though data on pilots are included in Research Performance Progress Reports, it is often difficult to extract, especially for large programs like the CTSAs where more than 600 pilots may be reported across all awardees annually. Data extraction challenges preclude analyses that could provide valuable information about pilots to researchers and administrators. METHODS: To address those challenges, we describe a script that partially automates extraction of pilot data from CTSA research progress reports. After extraction of the pilot data, we use an established machine learning (ML) model to determine the scientific content of pilots for subsequent analysis. Analysis of ML-assigned scientific categories reveals the scientific diversity of the CTSA pilot portfolio and relationships among individual pilots and institutions. RESULTS: The CTSA pilots are widely distributed across a number of scientific areas. Content analysis identifies similar projects and the degree of overlap for scientific interests among hubs. CONCLUSION: Our results demonstrate that pilot data remain challenging to extract but can provide useful information for communicating with stakeholders, administering pilot portfolios, and facilitating collaboration among researchers and hubs.