This code is for medical home page ICD-10 main diagnose code classification base on electronic medical record.
10 classes accuracy 81 classes accuracy
LSTM 0.8326 0.5355
LSTM + feat_emb 0.8515 0.5671
LSTM + feat_emb + keyword_atten 0.8728 0.5964
LSTM + feat_emb + keyword_atten + second_atten 0.8950
baseline(HA-GRU) 0.9151
Because the whole data is private, it is not available for public. We offer the sample data, which is seen in sample_data.csv The data includes 7 feature text data (chief complaint, present illness history, examination report, first progress note, operation record, ward-round records, discharge records) and the label(ICD-10 diagnose code). All the chinese text data have been tokenized by jieba tokenizer.
Code for dividing the whole data into three dataset: train data, validation data, test data.
Code for data process, including clean text, build vocabulary, transform sentences into word index representation and extract the topic word(key words for each classes).
Code for the model. The model structure is shown in the below picture.
Code for training and testing model. just use the command:
python run_model.py
filter_num = 2000 # 数据过滤参数，滤掉样本数小于filter_num的类
valid_rate = 0.1 # 验证集比例
test_rate = 0.1 # 测试集比例
seq_length = 1500 # 序列长度
keyword_num = 10 # 关键词（特征词）个数
test_rate = 0.1 # 测试集比例
hidden_dim = 100 # LSTM的隐层神经元个数（输出维度）
word_emb_dim = 50 # 词向量维度
feature_emb_dim = 10 # 特征名称embedding维度
keep_prob = 0.8 # dropout保留比例
num_layers = 1 # LSTM层数
batch_size = 20 # 每个batch的大小
learning_rate = 0.001 # 学习率
num_epochs = 2 # 训练数据迭代次数
add_feature = True # 是否加特征名embedding