Model

This model uses K-BERT architecture and UER framework. In addition you can use it als multiclassification with multiple labels now. You can also use Pretrained models with transformer achitecture from huggingface just runing convert_bert_from_huggingface_to_uer.py

python3 run convert_bert_from_huggingface_to_uer.py \
    # path of model from huggingface
    --input_model_path {./path_of_model_from_huggingface } \
    # path of model you want to put, which later would be use in infCheXbert_model.ipynb
    # better to save in ./models folder
    --output_model_path ./models/models_name.bin \
    # you would better to check layers by code model.state_dict() from torch. 
    # the standard layers would be 12 
    --layer_num 12

and after training translate the model in huggingface model by running convert_bert_from_uer_to_huggingface.py with

python3 run convert_bert_from_huggingface_to_uer.py \
    # path of model you have just trained, normally in outputs folders
    --input_model_path {./path_of_model_trained} \
    # any place you want to save
    --output_model_path {any place you want put} \
    # should consist with layers_num in arguments of infCheXbert_model.ipynb. default 12
    --layer_num 12 

Brain(knowledge graphs):

the medical knowledge graphs used in this thesis consist of relation of labels () and anatomy medicine knowledge. they are formatted into a spo file which should be put into the folder ./brain/kgs. The file uses a medicine database created by Precision Medicine Knowledge Graph (PrimeKG). The relative article written by Payal Chandak*, Kexin Huang*, and Marinka Zitnik was public on Scientific Data 2023.

The dataset is hosted on Harvard Dataverse, you can download it with this link and then run kg_split.ipynb to manufacture the spo file. For more detail,see readme file in the folder ./brain. and the original project in github PrimeKG.

experiments and models

see experiments reports, wrong prediction reports, and training scripts in experiments folder. Normally, output checkpoints are stored under outputs