{
  "abstract": "ObjectiveTo evaluate task-dependent selection of local and cloud models for structured extraction from multilingual non-English stroke discharge summaries, distinguishing entity detection, record assembly, and raw-document processing.\n\nMethodsThis retrospective system evaluation compared multilingual encoders, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5. Primary test cohorts comprised 332 section cases, 149 medication cases with 2,475 reference records after identity-based exclusions, and 191 laboratory cases. Outcomes were section-span F1, drug-name F1, normalized seven-field record recovery, test-name F1, and laboratory quintuple F1. Paired human comparisons used a common second-annotator reference on 50 cases per task. Saved cascade outputs were compared with curated-section controls using document-bootstrap intervals.\n\nResultsSection F1 was 0.919 for the encoder, 0.926 for GPT-5.5, and 0.932 for Qwen. GPT-5.5 led medication detection (0.966 versus 0.940 for Qwen), whereas Qwen led normalized medication recovery (0.381 versus 0.311; difference 0.070, 95% CI 0.026-0.114) and laboratory quintuple F1 (0.892 versus 0.822). With matched inference stacks, medication recovery declined from 0.377 on curated inputs to 0.204 for single-window and 0.246 for all-block cascades; quintuple F1 on the original laboratory cohort declined from 0.898 to 0.811. Corpus-wide extraction processed 193,101 summaries in 13.3 H100 GPU-hours.\n\nConclusionLocal models achieved strong detection performance without frontier-scale inference in this setting. Fine-tuned local LLMs were advantageous for record assembly under the evaluated recipes; annotation conventions, normalization, and upstream section extraction remained important constraints.",
  "abstract_shown": true,
  "affiliations": [
    "Nazarbayev University School of Medicine"
  ],
  "article_type": "PUBLISHAHEADOFPRINT",
  "arxiv": null,
  "author_total": 6,
  "authors": [
    {
      "name": "T. Aimyshev"
    },
    {
      "name": "I. Arupzhanov"
    },
    {
      "name": "G. Zhakhina"
    },
    {
      "name": "A. Rakhmankulova"
    },
    {
      "name": "S. Smagul"
    },
    {
      "affs": [
        0
      ],
      "name": "A. Gaipov"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/5xbaeg6c6sftm.bib",
  "cited_by": 0,
  "doi": "10.64898/2026.09.03.26362149",
  "fulltext_url": "https://www.medrxiv.org/content/10.64898/2026.09.03.26362149v1.full",
  "id": "5xbaeg6c6sftm",
  "issns": [],
  "landing_url": "https://www.medrxiv.org/content/10.64898/2026.09.03.26362149v1",
  "language": null,
  "licence": "CC-BY-NC-ND",
  "licence_label": "CC BY-NC-ND",
  "licence_source": "article",
  "licence_url": "https://creativecommons.org/licenses/by-nc-nd/4.0/",
  "metadata_licence": null,
  "native_subjects": [
    [
      "medrxiv",
      "health informatics"
    ],
    [
      "openalex",
      "1702|Machine Learning in Healthcare"
    ]
  ],
  "openalex": "W7211948664",
  "pdf_url": "https://www.medrxiv.org/content/10.64898/2026.09.03.26362149v1.full.pdf",
  "pmcid": null,
  "pmid": null,
  "published": "2026-09-03",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/5xbaeg6c6sftm.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:17:53Z",
      "native_id": "10.64898/2026.09.03.26362149",
      "source": "medrxiv",
      "url": "https://www.medrxiv.org/content/10.64898/2026.09.03.26362149v2"
    }
  ],
  "subjects": [
    {
      "field": "medicine",
      "origin": "medrxiv:health informatics",
      "subfield": "epidemiology-public-health"
    }
  ],
  "title": "Structured extraction from Russian/Kazakh discharge summaries: three model families at the human-agreement ceiling",
  "title_text": "Structured extraction from Russian/Kazakh discharge summaries: three model families at the human-agreement ceiling",
  "type": "preprint",
  "updated_at": "2026-10-30T12:18:31Z",
  "url": "https://research.draconitic.com/paper/5xbaeg6c6sftm",
  "venue": "medRxiv",
  "versions": []
}