TRV-2026-3020Version 1 · Published

Saved · see the current record

Why this version exists

Certified into the record

The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-3020
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-18T06:02:21.346321Z
status: published
lens: g_space
sector: health
headline: Commercially available multimodal LLMs for EEG-based prognostication after cardiac arrest
dek: Objectives: Machine-learning and deep-learning approaches have been successfully applied to EEG-based prognostication in comatose patients after cardiac arrest (CA), but typically rely on task-specific signal processing, feature extraction, and model development. Here, we investigated whether two commercially available multimodal large language models (LLMs) could predict neurological outcome directly from EEG images.
Methods: Thirty-three comatose patients on the first day after CA from a prospective cohort wer…
gain_title: In an exploratory study of 33 comatose cardiac arrest patients, ChatGPT 5.4 and Claude Opus 4.7 each reached 93.9% accuracy at predicting favorable neurological outcome from a single 10-second EEG image using few-shot prompting.
problem_title: (none)
trace_subject: (none)
gain_reading: In an exploratory study of 33 comatose cardiac arrest patients, ChatGPT 5.4 and Claude Opus 4.7 each reached 93.9% accuracy at predicting favorable neurological outcome from a single 10-second EEG image using few-shot prompting.
gain_evidence: The highest accuracy achieved by both models was 93.9% using few-shot prompting. | the AUC ranged from 0.86 to 0.99 for ChatGPT and from 0.74 to 1.00 for Claude
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers tested whether two commercially available chatbots, ChatGPT 5.4 and Claude Opus 4.7, could predict neurological outcome in comatose patients after cardiac arrest by looking at a single EEG image. The study used 33 patients from a prospective cohort, all recorded on the first day after arrest.

Across five prompting strategies, ChatGPT's AUC ranged from 0.86 to 0.99 and Claude's from 0.74 to 1.00. Both models peaked at 93.9% accuracy with few-shot prompting, where four example EEGs were supplied.

A previously published dedicated deep-learning model reached 100% on the same cohort, so the chatbots did not beat the specialised approach. The authors describe the discrimination as variable but potentially promising and say performance depended substantially on prompt design.
limitation: The study was exploratory and used only 33 patients, performance depended substantially on prompt design, and the authors say the findings require confirmation in a larger cohort.
tag: Evidence-backed gain
key_points: The study included 33 comatose patients from a prospective cohort, assessed on the first day after cardiac arrest. | Each patient was represented by one 10-second EEG segment in a longitudinal bipolar montage, rendered as a PNG image. | ChatGPT 5.4 and Claude Opus 4.7 were asked to estimate the probability of a favorable outcome. | Five prompting strategies were tested, including few-shot prompting with four example EEGs from outside the 33 patients. | AUC ranged from 0.86 to 0.99 for ChatGPT and from 0.74 to 1.00 for Claude. | Both models reached their best accuracy of 93.9% with few-shot prompting. | A previously published dedicated deep-learning model achieved 100% on the same cohort.
rundown: Machine-learning and deep-learning methods have been applied to EEG prognostication after cardiac arrest, but they typically need task-specific signal processing, feature extraction and model development. This study asked whether general-purpose commercial multimodal chatbots could skip that work and predict outcome directly from an image of the EEG.

Each of the 33 patients contributed one 10-second EEG segment plus a generic information block describing recording conditions. The chatbots estimated the probability of favorable outcome, and performance was scored with AUC and accuracy.

The spread in results was wide: Claude's AUC ranged from 0.74 to 1.00 depending on the prompt, and ChatGPT's from 0.86 to 0.99. The best accuracy for both models, 93.9%, came with few-shot prompting, while the dedicated deep-learning model scored 100% on the same cohort.
sources:
- peer_reviewed | Resuscitation | https://doi.org/10.1016/j.resuscitation.2026.111398 | 2026-11-17
prev: 0000000000000000000000000000000000000000000000000000000000000000
SHA-256 hash
ba473e53cfbee47809693dc94f5f8b668daf603b53e5f06b14397454a3c911fd
Previous version’s hash
0000000000000000000000000000000000000000000000000000000000000000
Check this record hasn’t been altered
How to check it yourself, without trusting this page

Download any version from /api/record/TRV-2026-3020 and hash its canonical field yourself, for example with shasum -a 256. You should get its content_hash. Each version’s text also ends with prev:and the hash of the version before it (version 1 points to 64 zeros). If anyone had changed a single character of any version after it was saved, those hashes wouldn’t line up.