mld3 / fiddle-experiments Goto Github PK

Hi Shengpu,

I have summarized some bugs in the mimic3_experiments directory. You may check them while available.

1_data_extraction

extract_data.py

Exceptions:

1. Line 251: pandas._libs.tslibs.np_datetime.OutOfBoundsDatetime: Result is too large for pandas.Timedelta. Convert inputs to datetime.datetime with 'Timestamp.to_pydatetime()' before subtracting.

Suggestions:

1. Replace x.INTIME with x.INTIME.to_pydatetime().

LabelDistributions.ipynb

Exceptions:

1. Line 44: FileNotFoundError
2. Line 54: pandas._libs.tslibs.np_datetime.OutOfBoundsDatetime: Result is too large for pandas.Timedelta. Convert inputs to datetime.datetime with 'Timestamp.to_pydatetime()' before subtracting.

Suggestions:

1. Replace open('config.yaml') with open('../config.yaml')
2. Replace x.INTIME with x.INTIME.to_pydatetime()

InclusionExclusion.ipynb

Exceptions:

1. Line 29: pandas._libs.tslibs.np_datetime.OutOfBoundsDatetime: Result is too large for pandas.Timedelta. Convert inputs to datetime.datetime with 'Timestamp.to_pydatetime()' before subtracting.

Suggestions:

1. Replace x.INTIME with x.INTIME.to_pydatetime()

PopulationSummary.ipynb

Exceptions:

1. Line 24: KeyError
2. Line 26: FileNotFoundError
3. Line 68: pandas._libs.tslibs.np_datetime.OutOfBoundsDatetime: Result is too large for pandas.Timedelta. Convert inputs to datetime.datetime with 'Timestamp.to_pydatetime()' before subtracting.

Suggestions:

1. Replace set_index('ICUSTAY_ID') with set_index('D')
2. The file pop.mortality_benchmark.csv is not exist
3. Replace x.INTIME with x.INTIME.to_pydatetime()

2_apply_FIDDLE

Suggestion: I think it's better to include FIDDLE module in this directory. After that, there are some other bugs.

README.md

Exceptions:

1. Line 41: FileNotFoundError

Suggestion:

1. There is no file named make_features.py

run_make_all.sh

exceptions:

1. output_dir is required
2. FileNotFoundError

Suggestion:

1. You should set the output_dir for each run, since it's required in run.py

2. Since the dir features/outcome=mortality,T=48.0,dt=1.0 is replaced by features/benckmark,outcome=mortality,T=48.0,dt=1.0 in 1_data_extraction/run_prepare_all.sh, this script is not able to run:

   OUTCOME=mortality
   T=48.0
   dt=1.0
   python run.py \
       --data_fname="$DATAPATH/features/outcome=$OUTCOME,T=$T,dt=$dt/input_data.p" \

   Since the file pop.mortality_benchmark.csv is not exist, this script is not able to run:

   python run.py \
       --data_fname="$DATAPATH/features/benchmark,outcome=mortality,T=48.0,dt=1.0/input_data.p" \
       --population="$DATAPATH/population/pop.mortality_benchmark.csv" \

3_ML_models

lib/data.py

Exceptions:

1. Line 75, 121: FileNotFoundError
2. Line 123, 124: Directory not exist

Suggestion:

1. The file pop.mortality_benchmark.csv is not exist
2. The directory features/outcome=mortality,T=48.0,dt=1.0 is not exist and replaced by features/benckmark,outcome=mortality,T=48.0,dt=1.0

config.yaml

Exceptions:

1. Line 21: The feature_dimension of ARF 4.0 is not 4143

Suggestion:

1. Set to 4381

run_deep_eval.py

Exceptions:

1. Line 57: import error

Suggestion:

1. Replace from sklearn.externals.joblib import Parallel, delayed with from joblib import Parallel, delayed

I want to run an experiment to assess whether a model trained on MIMIC is able to generalize on eICU. Are the FIDDLE features comparable as it is? If not, is it possible to carve out a subset that is comparable across the datasets?

Hi,
Thanks for open-sourcing the code. However, it seems that the code snippets hosted in https://github.com/MLD3/FIDDLE-experiments/tree/jamia-replication/eicu_experiments/1_data_extraction only preprocess a subset of tables mentioned in the paper, namely "medication", "nurseCharting", 'patient', 'lab', 'respiratoryCare', 'intakeOutput' tables.
Am I missing something?

Thanks,
YD

I tried to reproduce the FIDDLE experiments, however, the output X.npz is not a sparse matrix (and thus won't load using spicy.sparse.load_npz(), so I used lumpy.load()). X.npz contains:

X['data']: a long vector of only 1's
X['shape']: a vector describing the correct dimensions of the expected output tensor
X['fill_value']: a vector with just a single zero in it
X['coords']: a vector with 3 rows and the same number of columns as the length of X['data']

Is this an error or do I need to process this output first in order to get the sparse N x L x D tensor? I did not see anything in the documentation or paper regarding this. Cheers.

I am trying to reproduce your results using these instructions: https://github.com/MLD3/FIDDLE-experiments/tree/master/mimic3_experiments.

However, when running the IHM_Benchmark notebook, I get an error [Errno 2] No such file or directory 'train_listfile.csv'

I am not sure at which step the csv was supposed to be created since the other steps do not appear to generate it anywhere. Any help would be greatly appreciated.

Hi!

I tried to replicate the eICU experiments with the descretize option turned off, but got an error in the FIDDLE code saying "TypeError: bad operand type for unary ~: 'float' ". I adjusted the FIDDLE code and eventually it worked, but then I got a dimension error in the process of training the CNN, where it said that matrix 1 could not multiply with matrix 2 because they were not the right shape.

Do you have any idea in what direction I could go to fix this problem? Thank you so much!