A Russian Medical language understanding Benchmark is the set of NLP tasks on medical textual data for the Russian language.

This repository contains code and data to reproduce the results of the paper RuMedBench: A Russian Medical Language Understanding Benchmark.

Video from the AIME 2022 conference

• RuMedTop3* is the task for diagnosis prediction from a raw medical text, including patient symptoms and complaints.

• RuMedSymptomRec* Given an incomplete medical text, the task is to recommend the best symptom to check or verify.

• RuMedDaNet is the yes/no question answering task in the range of medical-related domains (pharmacology, anatomy, therapeutic medicine, etc).

• RuMedNLI is the natural language inference task in the clinical domain. The data is the full translated counterpart of MedNLI data.

• RuMedNER is the task of named entity recognition in drug-related user reviews. The data is from the RuDReC repo.

• ECG2Pathology is the task of assessment the quality of multilabel classification on ECG signals from the PTB-XL dataset.

*Both tasks are based on the RuMedPrime dataset.

We have implemented several baseline models; please see details in the paper.

Accuracy is the base metric for all tasks evaluation. For some tasks, additional metrics are used:

• RuMedTop3 and RuMedSymptomRec - Hit@3
• RuMedNER - F1-score

Test results:

We define the overall model score as mean over all metric values (with prior averaging in the case of two metrics).

* this is implementation from the paper RuBioRoBERTa: a pre-trained biomedical language model for Russian language biomedical text mining (repository).

You can find the extension of this benchmark (with closed test sets) on the MedBench platform.

Please refer to the code/ directory.

If you have any questions, please post a Github issue or email the authors.

@misc{blinov2022rumedbench,
    title={RuMedBench: A Russian Medical Language Understanding Benchmark},
    author={Pavel Blinov and Arina Reshetnikova and Aleksandr Nesterov and Galina Zubkova and Vladimir Kokh},
    year={2022},
    eprint={2201.06499},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}