TRV-2026-1361Certified recordPeer-reviewed

Can General-purpose Multimodal AI Reliably Measure Cobb Angles? A Head-to-head Diagnostic Agreement Study of GPT-5.1 and Gemini 3.1 Pro in Adolescent Idiopathic Scoliosis

General-purpose multimodal artificial intelligence (AI) models can interpret radiographs through natural-language prompts, but their quantitative agreement and anatomical traceability in adolescent idiopathic scoliosis remain uncertain. We compared GPT-5.1 and Gemini 3.1 Pro with a joint expert-consensus reference on 125 anonymized standing coronal whole-spine radiographs. Each model evaluated each image once in an independent chat session using the same standardized prompt; only the first response was retained.…

Health · Problem Space — documented harm · certified 2026-10-14 · v1 · article view · machine-readable

Current reading — problem

Neither GPT-5.1 nor Gemini 3.1 Pro agreed closely enough with expert consensus on case-level Cobb angles and end-vertebra identification to measure scoliosis curves autonomously.

What this doesn’t fix

Results are relative to expert consensus and do not establish absolute measurement accuracy or model repeatability, since each image was assessed once per model in a single session.

Evidence

Reader signal

How should this claim be treated?

Cite this record

Truvace Impact Record TRV-2026-1361, v1: “Can General-purpose Multimodal AI Reliably Measure Cobb Angles? A Head-to-head Diagnostic Agreement Study of GPT-5.1 and Gemini 3.1 Pro in Adolescent Idiopathic Scoliosis.” Truvace, 2026-10-14. /record/TRV-2026-1361 (accessed at citation time). sha256 d6bd4756bae89db4…

Calibration history

Every change to this record since certification, in the open. None yet — the reading has held since it entered the record.

  1. Certifiedv1d6bd4756bae8…

    Certified into the record

Verify this record
How to verify without trusting this page

Fetch the canonical text of any version from /api/record/TRV-2026-1361 and hash it yourself — for example shasum -a 256 on the saved canonical field. The result must equal content_hash, and each version’s text ends with prev:followed by the prior version’s hash (version 1 chains to 64 zeros). If a single character of any version had been altered since certification, the chain would not reproduce.