深度学习>> Explainable-Automated-Medical-Coding>> 返回
项目作者： acadTags

项目描述 ：
Implementation and demo of explainable coding of clinical notes with Hierarchical Label-wise Attention Networks (HLAN)
高级语言： Python
项目地址: git://github.com/acadTags/Explainable-Automated-Medical-Coding.git


Explainable Automated Medical Coding

This project proposes an explainable automated medical coding approach based on Hierarchical Label-wise Attention Network and label embedding initialisation. The approach can be applied to multi-label text classification in any domains.

Detailed explanation of the approach is in

  • Explainable Automated Coding of Clinical Notes using Hierarchical Label-Wise Attention Networks and Label Embedding Initialisation, accepted for Journal of Biomedical Informatics, Oct 2020. Preprint is available on ArXiv.

A part of the results (especially regarding label embedding initialisation) was orally and virtually presented in HealTAC 2020 with slides available.

Update:

  • Hierarchical evaluation CoPHE by @modr00cka: added --do_hierarchical_evaluation flag for hierarchical evaluation. (6 Sep 2021)
  • Code refactoring CS598DL4H by @dmcguire81 and @chrispebble: a cleaned implementation of the repository on HLAN/HA-GRU/HAN with Google Colab setup. (12 May 2022)

Hierarchical Label-wise Attention Network



The key computation graph is implemented in def inference_per_label(self) in ./HLAN/HAN_model.py.

Requirements

  • Python 3 (tested on versions 3.6. and 3.7.)
  • Tensorflow 1.* (tested on versions 1.4.1, 1.8.0, 1.13.1, 1.14.0, and 1.15.5)
  • Numpy
  • scikit-learn for implementing evaluation metrics
  • Gensim 3.* (tested on 3.8.3) for pre-training word and label embeddings with the word2vec algorithm
  • NLTK for tokenisation
  • Spacy 2.3.2 (before 3.x) for customised rule-based sentence parsing
  • TFLearn for sequence padding

How to Train on New Data

  • First, format your data where each line has the format doc__label__labelA labelB labelC, for details see datasets. The data can be either split to train[-validation]-test (each split as a single file) or without split (only one data file).
  • Second, prepare word embeddings and the optional label embeddings using Gensim package, using existing embeddings or those trained from your texts (e.g. using script in embeddings, also see the notebook from caml-mimic for embedding from MIMIC-III. The trained embeddings from MIMIC-III can be downloaded from onedrive (3.5G with other files)).
  • Third, add a new data block (if FLAGS.dataset == "YOUR_DATASET_NAME":) with variables specified in HAN_train.py. Please read closely the example code block and comments provided. For MIMIC-III dataset settings, use the existing data blocks in the code.
  • Finally, run commands (e.g. python HAN_train.py --dataset YOUR_DATASET_NAME) with arguments, see details in Training the models.

Jupyter Notebook Demo with MIMIC-III ICD Coding

  • First, ensure that you have requested the MIMIC-III dataset. Place the files D_ICD_DIAGNOSES.csv and D_ICD_PROCEDURES.csv under the knowledge_bases folder.

  • Second, download the files in checkpoints, cache_vocabulary_label_pik, and embeddings folders from Onedrive (link (3.5G with other files)).

  • Third, run the Jupyter Notebook demo demo_HLAN_viz.ipynb and try with your own discharge summaries or those in the MIMIC-III dataset. By setting the to_input in the notebook as True, the notebook will ask you to input or paste a discharge summary; otherwise, you can save your discharge summaries, each in a line, under the ..\dataset\ folder and replace the filename_to_predict to the filename (see Section 2.part2 in the notebook). After running, the predictions are displayed with label-wise attention visualisations. The attention visualisations are further stored as .xlsx files in the ..\explanation\ folder.

Content

  • ./HLAN/HAN_train.py contains code for configuration and training
  • ./HLAN/HAN_model.py contains the computational graph, loss function and optimisation
  • ./HLAN/data_util_gensim.py contains code for input and target generation
  • ./HLAN/demo_HLAN_viz.ipynb and ./HLAN/model_predict_util.py contains code for the demo based on Jupyter Notebook and the helper functions
  • ./HLAN/evaluation_setup.py and ./HLAN/multi_level_eval.py contains code from CoPHE for the hierarchical evaluation of multi-label classification
  • ./embeddings contains self-trained word2vec embeddings: word embeddings and label embeddings
  • ./datasets contains the datasets used
  • ./checkpoints contains the checkpoints of HLAN, HA-GRU, and HAN models trained from the author on the MIMIC-III datasets
  • ./explanations contains the Excel sheets displaying the attention visualisation, generated after running the demo in ./HLAN/demo_HLAN_viz.ipynb
  • ./knowledge_bases contains knowledge sources used for label subsumption relations and the ICD code description files
  • ./cache_vocabulary_label_pik stores the cached .pik files about vocabularies and labels
  • ./results-HEALTAC 2020 contains the CNN, CNN+att, Bi-GRU, BERT results with label embedding initilisation

Key Configurations and Further Details

Reproducing results from the paper

After getting access to MIMIC-III, obtain the data split from CAML using their preprocessing script. Then follow the steps in “How to Train on New Data”.

Training the models

--dataset is set to mimic3-ds-50 for MIMIC-III-50, mimic-ds for MIMIC-III, and mimic-ds-shielding-th50 for MIMIC-III-shielding.

To use label embedding initialisation (+LE), set --use_label_embedding to True; otherwise, set it to False.

All the --marking_ids below are simply for better marking of the command, which will appear in the name of the output files, can be changed to other values and do not affect the running.

Run HLAN

To train with the MIMIC-III-50 dataset

  1. python HAN_train.py --dataset mimic3-ds-50 --batch_size 32 --per_label_attention=True --per_label_sent_only=False --num_epochs 100 --report_rand_pred=False --running_times 1 --early_stop_lr 0.00002 --remove_ckpts_before_train=False --use_label_embedding=True --ckpt_dir checkpoint_HLAN+LE_50/ --use_sent_split_padded_version=False --marking_id 50-hlan --gpu=True

Run HA-GRU

This is by changing --per_label_sent_only to True while keeping --per_label_attention as True.

To train with the MIMIC-III-50 dataset

  1. python HAN_train.py --dataset mimic3-ds-50 --batch_size 32 --per_label_attention=True --per_label_sent_only=True --num_epochs 100 --report_rand_pred=False --running_times 1 --early_stop_lr 0.00002 --remove_ckpts_before_train=False --use_label_embedding=True --ckpt_dir checkpoint_HAGRU+LE_50/ --use_sent_split_padded_version=False --marking_id 50-hagru --gpu=True

Run HAN

This is by changing --per_label_attention to False. The --batch_size is changed to 128 for this model in the experiment.

To train with the MIMIC-III-50 dataset

  1. python HAN_train.py --dataset mimic3-ds-50 --batch_size 128 --per_label_attention=False --per_label_sent_only=False --num_epochs 100 --report_rand_pred=False --running_times 1 --early_stop_lr 0.00002 --remove_ckpts_before_train=False --use_label_embedding=True --ckpt_dir checkpoint_HAN+LE_50/ --use_sent_split_padded_version=False --marking_id 50-han --gpu=True

Other Configurations

For all the models above, you can set the learning rate (--learning_rate), number of epochs (--num_epochs), early stop learning rate (--early_stop_lr), and other configurations when you run the command, or set those in the *_train.py files.

By setting running_times as k, it will report averaged results and standard deviations with k runs. For example, --running_times 10.

For hierarchical evaluation results using CoPHE, add the flag --do_hierarchical_evaluation=True.

Check the full list of configurations in HAN_train.py.

To view the changing of training loss and validation loss, replacing $PATH-logs$ to a real path.

  1. tensorboard --logdir $PATH-logs$

Label embedding initialisation



Key part of the implementation of label embedding initiailisation is in the two functions def assign_pretrained_label_embedding_per_label (for HLAN and HA-GRU) and def assign_pretrained_label_embedding (for HAN) in ./HLAN/HAN_train.py.

Besides, below is the implementation of label embedding initiailisation on top of the model.py from the caml-mimic GitHub project.

  1. # based on https://github.com/jamesmullenbach/caml-mimic/blob/master/learn/models.py
  2. def _code_emb_init(self, code_emb, code_list):
  3.         # code_emb is a Gensim Word2Vec model loaded from pre-trained label embeddings
  4.         # code_list is a list of code having the same order as in multi-hot representation (sorted by frequency from high to low)
  5.         code_embs = Word2Vec.load(code_emb)
  6.         # bound for random variables for Xavier initialisation.
  7.         bound = np.sqrt(6.0) / np.sqrt(self.num_labels + code_embs.vector_size)  
  8.         weights = np.zeros(self.classifier.weight.size())
  9.         n_exist, n_inexist = 0, 0
  10.         for i in range(self.num_labels):
  11.             code = code_list[i]
  12.             if code in code_embs.wv.vocab:
  13.                 n_exist = n_exist + 1
  14.                 vec = code_embs.wv[code]
  15.                 #normalise to unit length
  16.                 weights[i] = vec / float(np.linalg.norm(vec) + 1e-6) 
  17.                 #additional standardisation for BERT models:
  18.                 #standardise to the same as the originial initilisation in def _init_weights(self, module) in https://huggingface.co/transformers/_modules/transformers/modeling_bert.html
  19.                 #weights[i] = stats.zscore(weights[i])*self.initializer_range # self.initializer_range = 0.02
  21.             else:
  22.                 n_inexist = n_inexist + 1
  23.                 #using the original xavier uniform initialisation for CNN, CNN+att, and BiGRU
  24.                 weights[i] = np.random.uniform(-bound, bound, code_embs.vector_size);
  25.                 #or using the original normal distribution initialisation for BERT
  26.                 #weights[i] = np.random.normal(0, std, code_embs.vector_size);
  27.         print("code exists embedding:", n_exist, " ;code not exist embedding:", n_inexist)
  29.         # initialise label embedding for the weights in the final linear layer
  30.         self.classifier.weight.data = torch.Tensor(weights).clone()
  31.         print("final layer: code embedding initialised")

Dataset and preprocessing

We used the MIMIC-III dataset with the preprocessing steps from caml-mimic to generate the two dataset settings MIMIC-III and MIMIC-III-50. We also created a MIMIC-III-shielding dataset based on the NHS shielding ICD-10 codes. See details in the datasets page.

Pre-training of label embeddings

We used the Continous Bag-of-Words algorithm (CBoW) in Gensim word2vec (see gensim.models.word2vec.Word2Vec, on all label sets in the training data. Codes for training word and label embeddings are available in train_word_embedding.py and train_code_embedding.py.

Additional libraries for other models

For CNN, CNN+att, Bi-GRU models:

For BERT models:

Other details

  • Using pre-trained BioBERT models: See answer from https://github.com/huggingface/transformers/issues/457#issuecomment-518403170.
  • Fine-tuning BERT with long documents: We adapted the sliding window approach from SimpleTransformers for multi-label classification. The idea is to treat a long document (discharge summaries in this project) as separate documents within the token length limit (sharing the same set of labels) for training. During the testing stage, the BERT model outputs the averaged prediction score of the separated documents. The results of MIMIC-III-50 were based on this adaptation. The results of MIMIC-III were based on first 512 tokens only due to a memory usage above the 60G limit.

Acknowledgement