AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration
The source code for AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration
Welcome to test the prototype of our visualization tool. The clinical hidden status is built by our latest representation learning model ConCare https://github.com/Accountable-Machine-Intelligence/concare and AdaCare. The internationalised multi-language support will be available soon.
- Install python, pytorch. We use Python 3.7.3, Pytorch 1.1.
- If you plan to use GPU computation, install CUDA
We do not provide the MIMIC-III data itself. You must acquire the data yourself from https://mimic.physionet.org/. Specifically, download the CSVs. To run decompensation prediction task on MIMIC-III bechmark dataset, you should first build benchmark dataset according to https://github.com/YerevaNN/mimic3-benchmarks/.
After building the decompensation dataset, please save the files in decompensation
directory to data/
directory.
-
We provide the trained weights in
./saved_weights/AdaCare
. You can obtain the reported performance in our paper by simply load the weights to the model. -
You need to run
train.py
in test mode and input the data directory. For example,$ python train.py --test_mode=1 --data_path='./data/'
-
The minimum input you need to train AdaCare is the dataset directory and file name to save model. For example,
$ python train.py --data_path='./data/' --file_name='trained_model'
-
You can also specify batch size
--batch_size <integer>
, learning rate--lr <float>
and epochs--epochs <integer>
-
Additional hyper-parameters can be specified such as the dimension of RNN, dropout rate, etc. Detailed information can be accessed by
$ python train.py --help
-
When training is complete, it will output the performance of AdaCare on test dataset.