KAICD

A knowledge attention-based deep learning framework for automatic ICD coding.

Usage

1. How to preprocess the raw data

Firstly, you need to import the package.

from utils import *

Then you can instance the MIMIC3 object and do preprocessing.

mimic = MIMIC(ADMISSIONSPATH="ADMISSIONS.csv", NOTEEVENTSPATH="NOTEEVENTS.csv", 
              DIAGNOSESICDPATH="DIAGNOSES_ICD.csv", DICDDIAGNOSESPATH="D_ICD_DIAGNOSES.csv")
mimic.get_basic_data(outPath='mimic3/data.csv', noteOutPath='mimic3/note_corpus.txt', titleOutPath='mimic3/title_corpus.txt')

ADMISSIONSPATH is the path of "ADMISSIONS.csv" in MIMIC III.
NOTEEVENTSPATH is the path of "NOTEEVENTS.csv" in MIMIC III.
DIAGNOSESICDPATH is the path of "DIAGNOSES_ICD.csv" in MIMIC III.
DICDDIAGNOSESPATH is the path of "D_ICD_DIAGNOSES.csv" in MIMIC III.
outPath is the output path of preprocessed data.
noteOutPath is the output path of clinck notes corpus which will be used to train note embedding.
titleOutPath is the output path of ICD titles corpus which will be used to train title embedding.

After doing this, you can get 3 new files: data.csv, note_corpus.txt, title_corpus.txt.

2. How to train the model

Firstly, you need to import the package.

from utils import *
from model import *

Then you need to instance the data utils class and get the pretrained word embedding.

dataClass = DataClass(dataPath='mimic3/data.csv', dicddiagnosesPath='D_ICD_DIAGNOSES.csv', noteMaxLen=768)
dataClass.vectorize(noteFeaSize=320, titleFeaSize=192, iters=10, 
                    noteCorpusPath=‘mimic3/note_corpus.txt’, loadCache=True)

dataPath is the path of "data.csv".
dicddiagnosesPath is the path of "D_ICD_DIAGNOSES.csv".
noteMaxLen is the maximum number of keeped words in clinic notes. The words with low TF-IDF will be removed.
noteFeaSize is the embedding size of clinic notes.
titleFeaSize is the embedding size of ICD titles.
iters is the number of iterations in pretraining word embedding.

Finally, you can instance the model class and train the model.

model = KAICD(dataClass.classNum, dataClass.vector['noteEmbedding'], dataClass.vector['titleEmbedding'], dataClass.tokenizedTitle, dataClass.titleLen,
              noteFeaSize=320, titleFeaSize=192, hiddenSize=128, filterSize=448,
              embDropout=0.0, fcDropout=0.5)
model.train(dataClass, trainSize=128, batchSize=256, epoch=1000,
            threshold=0.2, earlyStop=30,
            savePath='model/KAICD', 
            metrics="MiF", report=["MiF", "MiAUC"])

noteFeaSize is the embedding size of clinic notes.
titleFeaSize is the embedding size of ICD titles.
hiddenSize is the hidden size of the BiRNN.
filterSize is the filter size of the CNN.
embDropout is the dropout rate after embedding layer.
fcDropout is the dropout rate before output layer.
threshold is the threshold in multi label prediction.
report is a list of reported indicators after each epoch.

3. How to do prediction

model.load(path="xxx.pkl", map_location="cpu", dataClass=dataClass)
model.to_eval_mode()
Y_prob_pre,Y = model.calculate_y_prob_by_iterator(dataClass.one_epoch_batch_data_stream(batchSize=64, type='valid', device=model.device))

path is the save path of your trained model.