Overview

Bulfinch Building

CPC-Bench is a benchmark for evaluating the clinical reasoning capabilities of AI models using expert-curated cases licensed from The New England Journal of Medicine. The benchmark consists of 10 distinct tasks that test various aspects of clinical reasoning, from differential diagnosis to medical image interpretation.

7,102
Clinical Cases
100+
Years of CPCs
10
Benchmark Tasks
4,503
Unique Diagnoses

Benchmark Tasks

Click on any task box below for detailed information and examples.

Text-based Challenges
Differential Diagnosis (DDx)

Generate ranked list of potential diagnoses given clinical presentation.
Testing Plan

Recommend appropriate diagnostic tests and procedures.
Literature Search

Search for literature to support a medical claim.
Diagnostic TouchpointsPhysician Annotations

Diagnosis during key moments of clinical course.
Question Answering (QA)

Clinical concept understanding and factual knowledge.
Clinical ReasoningPhysician Annotations

Identify the confirmatory and disconfirmatory evidence for each diagnosis.
Information OmissionPhysician Annotations

Provide a differential diagnosis using only background information and initial presentation.
Multimodal Challenges
NEJM Image Challenge

Multiple-choice questions from NEJM Image Challenge.
Visual Image Interpretation (VQA)

Multiple-choice medical imaging questions constructed from the figures and captions in CPCs.
Visual Differential Diagnosis

Provide a differential diagnosis provided only the figures and tables from the initial presentation.

Leaderboard

Model Performance
Updated: Loading...
Rank Model DDx@1 DDx@10 DDx@1 (Omit Normal) Testing Plan Literature Search QA VQA Visual Differential Diagnosis Image Challenge Image Challenge Diagnosis Average

Performance Analysis

Text-based Tasks
Multimodal Tasks