CohortX: Extracting Executable Cohort Definitions for Medical Imaging Research

Abstract

Medical imaging studies, clinical trials, datasets, and registries critically rely on well-defined patient cohorts. However, cohort selection criteria are typically described in unstructured free text within scholarly articles and trial publications, making them difficult to operationalize for large-scale medical image analysis and clinical translation. This disconnect between narrative cohort definitions and computable patient selection remains a significant barrier to reproducibility, fairness assessment, and efficient use of imaging and clinical data.The CohortX Challenge proposes a shared benchmark for extracting, structuring, and semantically representing cohort selection criteria from scholarly medical articles, with a focus on imaging-based and multimodal clinical studies. Participants will develop methods to identify inclusion and exclusion criteria, normalize clinical and imaging-related concepts using standard biomedical terminologies, and generate structured, machine-interpretable representations that capture logical, demographic, and temporal constraints. These representations enable downstream translation into executable cohort queries over imaging repositories and integrated clinical data (e.g., EHR-linked imaging cohorts).The challenge explicitly supports MICCAI 2026 priorities by promoting:* Integration of imaging and non-imaging data (clinical variables, diagnoses, procedures),* Clinically informed evaluation metrics focused on cohort correctness and usability,* Analysis of generalization and fairness across demographic and clinical subgroups,* Use of open, publicly available data and sustained benchmarking beyond the conference.By emphasizing computable cohort definitions rather than text-only extraction, CohortX advances translational medical data analysis and supports robust, reproducible, and fair evaluation of imaging-based AI methods.