Автоматизоване формування простору ознак із неструктурованих медичних записів для задач виявлення статистичних аномалій

Abstract

The paper addresses the problem of automated structuring of text-based medical data for further statistical analysis and anomaly detection. Methods for Named Entity Recognition (symptom extraction) from unstructured patient complaints in Ukrainian were investigated. A synthetic text corpus based on the ICPC-2 classifier was generated for experimental analysis. A comparative analysis of three approaches was performed: rule-based, Large Language Model (LLM), and a hybrid method. It was determined that the hybrid approach provides the best balance between Recall and interpretability, which is critical for forming a reliable dataset for anomaly detection in medical statistics.