MEDOTOR: AI assistant for otolaryngologist based on reasoning

Project leader: Sergey Koltcov

Authors: Sergey Koltcov, Anton  Surkov, Marina Budkovaya, Svetlana Rebrova, Patimat Dakhadaeva, Sergey Levin

Copyright holder: National Research University Higher School of Economics, HSE – Saint Petersburg. St. Petersburg Research Institute of Ear, Throat, Nose, and Speech

Software program: "Medotor: An AI assistant for an otolaryngologist based on a reasoning LLM"

A software prototype of an intelligent assistant for an otolaryngologist, implementing local large language model (LLM) technology in combination with retrieval-augmented generation (RAG) and a long-term memory mechanism (MemoRAG). The system is designed to support clinical decisions: it analyzes unstructured medical data—including discharge summaries, medical histories, laboratory parameters, and scanned PDF documents—and generates evidence-based recommendations consistent with current clinical guidelines.

The architecture includes three functional modules:
a) a data collection and preprocessing module that extracts structured clinical information using OCR and medical named entity recognition (Medical NER);
b) a diagnostic module based on a local LLM that evaluates the effectiveness of interventions and predicts individual relapse risk;
c) a recommendation generation module that uses reasoning LLM in a hybrid RAG/MemoRAG architecture to generate personalized and context-sensitive clinical advice.

The system operates entirely locally, without transferring data to cloud services, ensuring compliance with medical information protection laws and technological sovereignty. The development was carried out in partnership with the Federal State Budgetary Institution "St. Petersburg Research Institute of Ear, Throat, Nose, and Speech" (St. Petersburg Research Institute of ENT), which provided anonymized clinical data and conducted expert medical validation.

Medotor reduces the cognitive and operational burden on physicians, increases diagnostic accuracy and reproducibility, improves patient routing, and enhances the quality of specialized care. The solution's architectural principles ensure its scalability to other medical specialties and form the basis for subsequent formal registration as an intellectual property asset.

Repository: https://github.com/hse-scila/Medotor