ICD MSMN

ICD-MSMN

La implementación oficial de "Los sinónimos de código sí importan: red de coincidencia de múltiples sinónimos para la codificación automática de ICD" [ACL 2022]

Ambiente

Todos los códigos se prueban en Python 3.7, PyTorch 1.7.0. Necesita instalar opt_einsum para realizar cálculos de einsum. Se necesitan al menos 32 GB de GPU para entrenar la configuración completa de MIMIC-III.

Conjunto de datos

Solo ponemos varias muestras para cada conjunto de datos. Es necesario obtener licencias para descargar el conjunto de datos MIMIC-III. Una vez que obtenga el conjunto de datos MIMIC-III, siga caml-mimic para preprocesar el conjunto de datos. Debe obtener train_full.csv , test_full.csv , dev_full.csv , train_50.csv , test_50.csv , dev_50.csv después del preprocesamiento. Colóquelos en sample_data/mimic3 . Entonces deberías usar preprocess/generate_data_new.ipynb para generar un conjunto de datos en formato json.

Incrustación de palabras

Descargue word2vec_sg0_100.model de LAAT. Necesita cambiar la ruta de incrustación de palabras.

Usa nuestro código

MIMIC-III completo (1 GPU):

 CUDA_VISIBLE_DEVICES=0 python main.py --n_gpu 1 --version mimic3 --combiner lstm --rnn_dim 256 --num_layers 2 --decoder MultiLabelMultiHeadLAATV2 --attention_head 4 --attention_dim 512 --learning_rate 5e-4 --train_epoch 20 --batch_size 2 --gradient_accumulation_steps 8 --xavier --main_code_loss_weight 0.0 --rdrop_alpha 5.0 --est_cls 1  --term_count 4  --sort_method random --word_embedding_path word_embedding_path

MIMIC-III completo (8 GPU):

 NCCL_IB_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --nproc_per_node 8 --master_port=1212 --use_env  main.py --n_gpu 8 --version mimic3 --combiner lstm --rnn_dim 256 --num_layers 2 --decoder MultiLabelMultiHeadLAATV2 --attention_head 4 --attention_dim 512 --learning_rate 5e-4 --train_epoch 20 --batch_size 2 --gradient_accumulation_steps 1 --xavier --main_code_loss_weight 0.0 --rdrop_alpha 5.0 --est_cls 1  --term_count 4  --sort_method random --word_embedding_path word_embedding_path

MÍMIC-III 50:

 CUDA_VISIBLE_DEVICES=0 python main.py --version mimic3-50 --combiner lstm --rnn_dim 512 --num_layers 1 --decoder MultiLabelMultiHeadLAATV2 --attention_head 8 --attention_dim 512 --learning_rate 5e-4 --train_epoch 20 --batch_size 16 --gradient_accumulation_steps 1 --xavier --main_code_loss_weight 0.0 --rdrop_alpha 5.0 --est_cls 1 --term_count 8 --word_embedding_path word_embedding_path

Evaluar puntos de control

 python eval_model.py MODEL_CHECKPOINT

punto de control mimic3

punto de control imitar3-50

Citación

 @inproceedings{yuan-etal-2022-code,
    title = "Code Synonyms Do Matter: Multiple Synonyms Matching Network for Automatic {ICD} Coding",
    author = "Yuan, Zheng  and
      Tan, Chuanqi  and
      Huang, Songfang",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-short.91",
    pages = "808--814",
    abstract = "Automatic ICD coding is defined as assigning disease codes to electronic medical records (EMRs).Existing methods usually apply label attention with code representations to match related text snippets.Unlike these works that model the label with the code hierarchy or description, we argue that the code synonyms can provide more comprehensive knowledge based on the observation that the code expressions in EMRs vary from their descriptions in ICD. By aligning codes to concepts in UMLS, we collect synonyms of every code. Then, we propose a multiple synonyms matching network to leverage synonyms for better code representation learning, and finally help the code classification. Experiments on the MIMIC-III dataset show that our proposed method outperforms previous state-of-the-art methods.",
}