Public articles linked to the same research event.
medRxiv Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while 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), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while 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), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while 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), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while 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), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.