Short report from Lithuanian conference on AI Technologies in Medicine
retained text
Full info: [https://lithuania.ai/en/events/konferencija-dirbtinio-intelekto-technologijos-medicinoje-tyrimai-ir-diagnostika](https://lithuania.ai/en/events/konferencija-dirbtinio-intelekto-technologijos-medicinoje-tyrimai-ir-diagnostika) Things I noticed among other things presented (almost every presentation was interesting): * experimental decision models for the medical Q&A (provided medical abstract, questions and choices) were presented in welcome speech by the head of [AIML.lt](http://AIML.lt) lab: [https://huggingface.co/AIML-VU/laya-multilingual-lt-med](https://huggingface.co/AIML-VU/laya-multilingual-lt-med) and [https://huggingface.co/AIML-VU/Qwen3.5-4B-LT-med](https://huggingface.co/AIML-VU/Qwen3.5-4B-LT-med) * Work featuring an MCP server used to speed up genetics specialists work: "Interpretable language model assistant for evaluating evidence of structural chromosomal rearrangement breakpoints in rare disease diagnostics" by Vytautas Rimas, Erinija Pranckevičienė, Eglė Preikšaitienė, Živilė Maldžienė * Using local models for bulk translation of medical text datasets into Lithuanian with correct Lithuanian terms: "Strategies for translating medical training data into Lithuanian" by Augustė Raišytė, Linas Petkevičius, Dalius Matuzevičius There were more applications of local models. Some models mentioned: Qwen3.5-4B, Qwen3.6-35B-A3B, Qwen3.5-122B-A10B, medgemma-27B. Models were sometimes outdated, but this is lots of times because of long duration of academic work (when you have a lot of work to do other than checking for the newest released model) or the lack of new medical domain-specific models. The proportion of classical ML and neural networks was even higher, one example: "Diagnostic Reliability of Multi-Label Chest X-Ray AI outside the Training Hospital: Accuracy, Calibration and Fairness across Seven Architectures" by Paulius Bundza and Justas Trinkūnas Regulation representatives where present, and their topics where, simplified, funding+application quality; the latter was focused on patient data safety and technology use safety, e.g. regular incidents analysis. **The most important thing why I posted this: researchers in medicine use local models.** This is one of the cases where local models win (medical professionals are absolutely strict about patient data staying private, and local models are the only way).
document fetched 2026-10-21 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xaxffx/short_report_from_lithuanian_conference_on_ai