Maintenance Notice

Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Wednesday, July 01, 2020 at 8:00 PM to 10:00 PM EST. We apologize in advance for any inconvenience this may cause you.

Who will be affected?

Currently submitted to: JMIR Formative Research

Date Submitted: Feb 4, 2026
Open Peer Review Period: Feb 8, 2026 - Apr 5, 2026
(currently open for review)

Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.

A Comparative Benchmark of 19 Large Language Models for Structured Data Extraction from Neurosurgical Clinical Records

  • Martin Černý; 
  • Martin Májovský; 
  • Sanju Lama; 
  • Guo Edward; 
  • Hana Hallak; 
  • Rahul Singh; 
  • Homer Riva-Cambrin; 
  • Petr Šustek; 
  • Lucie Široká; 
  • Ivana Tullett; 
  • Katarina Horčičáková; 
  • Samuel Hužava; 
  • Kateřina Sajfrídová; 
  • Klára El-Haj Ahmadová; 
  • Garnette Roy Sutherland; 
  • David Netuka

ABSTRACT

Background:

Large language models (LLMs) are increasingly used to extract information from electronic health records (EHRs). Given the rapid pace of LLM development, robust scenario-specific benchmarks are essential to evaluate clinical usefulness and support safe deployment.

Objective:

To compare contemporary LLMs on structured data extraction from real neurosurgical EHRs written in the Czech language.

Methods:

In a prospective single-center cohort, 172 hospitalized patients provided informed consent for use of anonymized EHRs. For each patient, predefined records were collected and concatenated. Ground truth for 35 data points was established by dual extraction with consensus. A standardized prompt requesting JSON output was submitted to 19 LLMs. Primary outcome was overall accuracy; secondary outcomes were category-level accuracy and proportion of complete machine-readable outputs.

Results:

6,264 documents were collected (median 33 per patient). Ground truth was established with 92.6% initial inter-rater agreement before consensus seeking. Several models produced complete JSON outputs for 100% of cases (Claude 4.1 Opus, Grok 4, Gemini 2.5 Flash); GPT-4.1 (DeepSearch) and GPT-5 completed 99.4%. Highest accuracy was achieved by GPT-4.1 (87.6%), followed by GPT-4.5 (85.6%), Claude 4.1 (84.8%), and Grok 4 (84.2%). Accuracy declined by data type: binary (up to 95%), numeric (~89%), multiple-choice (~75%), and short text (~78%).

Conclusions:

Currently available LLMs can reliably extract structured clinical information from full, non-English EHRs, while older or smaller models show major limitations. A hybrid workflow—automated extraction with targeted validation—appears practical for research use.


 Citation

Please cite as:

Černý M, Májovský M, Lama S, Edward G, Hallak H, Singh R, Riva-Cambrin H, Šustek P, Široká L, Tullett I, Horčičáková K, Hužava S, Sajfrídová K, El-Haj Ahmadová K, Sutherland GR, Netuka D

A Comparative Benchmark of 19 Large Language Models for Structured Data Extraction from Neurosurgical Clinical Records

JMIR Preprints. 04/02/2026:92871

DOI: 10.2196/preprints.92871

URL: https://preprints.jmir.org/preprint/92871

Download PDF


Request queued. Please wait while the file is being generated. It may take some time.

© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.