Inquiry: possibility of sharing used patient splits

[manuelburger]

Dear Chang Lu and colleagues,

I am a master's student in data science and working on graph neural networks in the health care domain. I would love to reference and compare to your work. To get the best comparison estimate possible, it would be essential to know what patient split you have used for the published results here: https://www.ijcai.org/proceedings/2021/0486.pdf

Your code here:

CGL/run_preprocess.py

Line 45 in b8b8833

train_pids, valid_pids, test_pids = split_patients(

and here:

CGL/preprocess/build_dataset.py

Line 6 in b8b8833

def split_patients(patient_admission: dict, admission_codes: dict, code_map: dict, seed=6669) -> (np.ndarray, np.ndarray, np.ndarray):

indicates that a random split has been used.

Would it be feasible for you to share the MIMIC-III patient ids (i.e. the SUBJECT_ID) for the three splits train/validation/test used in the published results?

Thank you very much for sharing and making your work openly available.

Best regards,
Manuel
from ETH Zürich

[LuChang-CS]

Hi Manuel, thank you for your interests in our project. As you mentioned, in the function of split_patients, there is a default argument seed=6669 and

np.random.seed(seed)

In run_preprocess.py, I do not specify the seed so it will use the default seed. Normally, we can get the same split of patient ids. It is safe to use your own split.

Additionally, it is possible that the final prediction results are not the same under different software (e.g., CUDA version) or hardware (e.g., GPU version) settings.