This paper proposes approaches to diagnosing disease stages using unstructured text data, using Duchenne muscular dystrophy as an example. The step-by-step approach involved extracting key features from primary anonymized medical records. Expert assessments of feature significance (from 0 to 10) were used to perform weighted voting, determining the disease stage. Using a Large Language Model (LLM), the quality of the responses was verified using an additional module that monitors the model's output. Data structured during the LLM process was also used for diagnosis based on a casebased reasoning method. Testing on real data from patients with Duchenne disease showed an accuracy of 82.5%, with the stage determined no more than one stage lower than the actual stage in 7.5% of cases and no more than one stage higher than the actual stage in 10% of cases using weighted voting. When using case-based reasoning, the system's accuracy was 80%; in 10% of cases, the stage was determined to be no more than one stage lower than the actual stage, and in 10% of cases, the stage was determined to be no more than one stage higher than the actual stage.
At the journal's website: http://www.jitcs.ru/index.php?option=com_content&view=article&id=949
Math-Net.RU: https://www.mathnet.ru/eng/itvs954
Kobrinskii B., Trunov N. Application of Large Language Models to Medical Diagnostic Tasks // Journal of Information Technologies and Computing Systems, 2026, No. 2, pp. 113–119.