ICD MSMN

ICD-MSMN

Implementasi resmi "Sinonim Kode Penting: Jaringan Pencocokan Sinonim Ganda untuk Pengodean ICD Otomatis" [ACL 2022]

Lingkungan

Semua kode diuji dengan Python 3.7, PyTorch 1.7.0. Perlu menginstal opt_einsum untuk perhitungan einsum. Setidaknya diperlukan GPU 32GB untuk melatih pengaturan penuh MIMIC-III.

Kumpulan data

Kami hanya menempatkan beberapa sampel untuk setiap dataset. Seseorang perlu mendapatkan lisensi untuk mengunduh kumpulan data MIMIC-III. Setelah Anda mendapatkan kumpulan data MIMIC-III, ikuti caml-mimic untuk melakukan praproses kumpulan data tersebut. Anda harus mendapatkan train_full.csv , test_full.csv , dev_full.csv , train_50.csv , test_50.csv , dev_50.csv setelah prapemrosesan. Silakan letakkan di bawah sample_data/mimic3 . Maka Anda harus menggunakan preprocess/generate_data_new.ipynb untuk menghasilkan kumpulan data format json.

Penyematan kata

Silakan unduh word2vec_sg0_100.model dari LAAT. Anda perlu mengubah jalur penyematan kata.

Gunakan kode kami

MIMIC-III Penuh (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 Penuh (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

MIMIK-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

Evaluasi pos pemeriksaan

 python eval_model.py MODEL_CHECKPOINT

pos pemeriksaan mimik3

pos pemeriksaan meniru3-50

Kutipan

 @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.",
}