CrossCheck-BHC

How to rate

Each item is a statement from an AI-written brief hospital course that the VeriFact clinicians labelled Not Supported by the patient's record. Rate how clinically severe the error would be if the statement appeared in a discharge summary used by the next clinician caring for the patient.

• 0 no clinical relevance · 1 minor, unlikely to affect care · 2 moderate, could change treatment, monitoring or follow-up · 3 severe, potential for patient harm
• You may disagree that the statement is wrong: answer No if the record supports it and Unsure if you cannot decide. Still rate severity as if it were not supported.
• The VeriFact excerpts are notes retrieved by the published VeriFact system, a reading aid that may be incomplete. The Full record holds every reference note; notes from other stays are marked “outside index admission”.
• There are 95 statements, roughly 6–7 hours in total. You can pause at any time: every answer is saved automatically, and you can sign in again later and continue where you stopped. Time per item is recorded automatically.
• The comments of the VeriFact clinicians (if there are any) appear after you have answered the agreement question, so please form your own judgement first.
• No model outputs are shown. Please rate independently and do not discuss items with the study team until all ratings are returned. A short set of about 10 items will be presented again in roughly two weeks.
• All clinical text comes from MIMIC-III (PhysioNet). Use it only under your own credentialed access and data use agreement; do not copy it elsewhere.

Shortcuts: 0–3 severity · Y/N/U agreement · ←/→ previous/next · / search the record