Echo: Visual Steering for AI-Assisted Cohort Definition and Comparison in Electronic Health Records

Abstract

Constructing clinical cohorts in Electronic Health Records (EHRs) requires translating clinical eligibility intent into executable database logic while preserving assumptions and downstream effects. Large language model (LLM)-based query generation can reduce SQL authoring barriers, but offers limited support for inspecting how criteria are grounded, refined, and compared across alternative cohort definitions. We present Echo, an early-stage visual analytics prototype for steering AI-assisted cohort construction in EHRs. Echo combines an editable Criteria Graph, criterion-level SQL review, and coordinated cohort comparison views. A case study with two domain experts suggests that Echo has the potential to support eligibility-logic inspection and sub-cohort derivation, while surfacing needs for codelist validation, value-based variables, temporal logic, and larger-scale evaluation.