Ehr transformer ICU example #245

ellabarkan · 2023-01-08T11:30:08Z

.

…ic values)

This reverts commit b46b8ad.

This reverts commit b35af44.

mosheraboh

Thanks Ella.
I've reviewed the files you mentioned (didn't get into all the details though).
Looks good, please see few comments inline

mosheraboh · 2023-01-08T16:01:49Z

examples/fuse_examples/multimodality/ehr_transformer/ops_read_cinc.py

+from fuse.utils.ndict import NDict
+
+
+class OpReadDataframeCinC(OpBase):


Can we use OpReadDataFrame from fuse as is? What do you think?
If we should process the dataframe, we can do it before we pass is to OpReadDataframe

Following your suggestion, I moved the creation of the patients dataframe to the "loading stage" and using OpReadDataFrame now.

mosheraboh · 2023-01-08T19:49:13Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+from ops_read_cinc import OpReadDataframeCinC
+from fuse.data.utils.export import ExportDataset
+
+SOURCE = r"C:/D_Drive/Projects/EHR_Transformer/PhysioNet/predicting-mortality-of-icu-patients-the-physionetcomputing-in-cardiology-challenge-2012-1.0.0/predicting-mortality-of-icu-patients-the-physionet-computing-in-cardiology-challenge-2012-1.0.0"


We need to read the path from env variable and provide instructions on how to download the data

Updated configuration file to read from the environment variable and the main in dataset.py

mosheraboh · 2023-01-08T19:52:23Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+
+
+class OpAddBMI(OpBase):
+    def __call__(self, sample_dict) -> Any:


Relevant to all ops: if you think that the op might be useful for other datasets as well (and also for readability).
It's better to also get the input keys and output keys. in this case key_in_height, key_in_weight, key_out_bmi.

Added parameters of input and output keys

mosheraboh · 2023-01-08T19:53:06Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+        d_static = sample_dict["StaticDetails"]
+
+        if ("Height" in d_static.keys()) & ("Weight" in d_static.keys()):
+            height = d_static["Height"]


Do we expect specific units, maybe we should specify it in call comments.

Added information in the comment , please check if it in the right place

mosheraboh · 2023-01-08T19:54:47Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+
+        d_visits_sentences = sample_dict["VisitSentences"]
+        n_visits = len(d_visits_sentences)
+        # TODO add filtering patients with small number of visits - add to configuration


If you put this op in static_pipeline, you can return None in order to filter.

Moved the filtering patients on the number of visits to the raw data loading method together with filtering based on time (less than X hours in the hospital)

mosheraboh · 2023-01-08T21:53:46Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+
+        d_static = sample_dict["StaticDetails"]
+
+        if ("Height" in d_static.keys()) & ("Weight" in d_static.keys()):


if not? is it a sample you want to keep or not?
If yes, maybe you should add -1 as a dummy value?

I prefer not to add BMI with -1 to the dict of static values in case it can't be calculated from missing Height or Weight.
It will be further converted to nan during percentile calculations and will be skipped in the next steps of trajectories build.

mosheraboh · 2023-01-08T21:55:00Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+        d_static = sample_dict["StaticDetails"]
+
+        if ("Height" in d_static.keys()) & ("Weight" in d_static.keys()):
+            height = d_static["Height"]


Since it's ndict, you can also access it as sample_dict["StaticDetails.Height"]

mosheraboh · 2023-01-08T21:55:35Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+
+        d_static = sample_dict["StaticDetails"]
+
+        if ("Height" in d_static.keys()) & ("Weight" in d_static.keys()):


"Height" in d_static
will also work

mosheraboh · 2023-01-08T21:59:16Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+        # convert continuous measurements to categorical ones based on defined percentiles
+
+        # mapping static clinical characteristics (Age, Gender, ICU type, Height, etc)
+        for k in sample_dict["StaticDetails"].keys():


for k in sample_dict["StaticDetails"]:
Will also work

mosheraboh · 2023-01-08T22:01:57Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+
+
+class OpMapToCategorical(OpBase):
+    def __call__(self, sample_dict, percentiles: dict) -> Any:


maybe rename percentiles to bins here?

Do you mean in all places in the code?

…mer_example

…iAI/fuse-med-ml into ehr_transformer_example

…; other misc changes

…into ehr_transformer_example

… comments to files

mosheraboh · 2023-01-13T06:46:04Z

This example is based on Vadim (@floccinauc) research.
@ellabarkan implemented the dataset and tokenizer, and I implemented the model and training script.
@michalozeryflato @SagiPolaczek , I'm merging now cause we need it for Sunday.
But please review and add comments/questions to anyone of us.

mosheraboh · 2023-01-13T06:51:10Z

@ellabarkan , we still need to create a unittest. Can you please work on it in a new PR next week?

SagiPolaczek

With some delay I did a quick review and added few small comments inline 😄

The code looks nice and clean! I think a small unittest can be useful here to make sure the example won't break in the future.

SagiPolaczek · 2023-01-16T17:14:30Z

examples/fuse_examples/multimodality/ehr_transformer/config.yaml

+  aux_next_vis_classification: ${aux_next_vis_classification}
+  next_vis_loss_weight: 0.1
+
+  # uncomment to track in clearml


I think it's a better practice to use flags. for example:

config:

clearml: track: 1 # 0 or 1 project_name: "ehr_transformer" task_name: ${name} tags: "fuse_example" reuse_last_task_id: True continue_last_task: False

and in the code:

if cfg.clearml.track: # blablabla

SagiPolaczek · 2023-01-16T17:14:49Z

examples/fuse_examples/multimodality/ehr_transformer/config.yaml

+  #   continue_last_task: False
+
+
+  # uncomment for SGD


like the above :)

SagiPolaczek · 2023-01-16T17:17:42Z

examples/fuse_examples/multimodality/ehr_transformer/dataset.py

+STATIC_FIELDS = ["Age", "Gender", "Height", "ICUType", "Weight"]
+
+
+class OpAddBMI(OpBase):


Cool :)

Suggestion:
if you think its useful - we can start an ops_clincal.py file and store it over there for future reuse (might apply also for some UKBB ops)

SagiPolaczek · 2023-01-16T17:24:17Z

examples/fuse_examples/multimodality/ehr_transformer/main_train.py

+    :param track_clearml: optional - to track with clearml provide arguments to start_clearml_logger()
+    """
+
+    if track_clearml is not None:


this is where we can fit the flag (as a continuation for the previous comment)

SagiPolaczek · 2023-01-16T18:27:51Z

fuse/data/ops/ops_common.py

@@ -521,3 +521,16 @@ def __call__(self, sample_dict: NDict, key: str, value: Any) -> Union[None, dict
        """
        sample_dict[key] = value
        return sample_dict
+
+
+class OpSetIfNotExist(OpBase):


Maybe add it to the ops list in the data/REAME in a future PR

simona-rc and others added 18 commits October 2, 2022 08:42

Initial example of EHR Transformer based on BERT from Hugging Face

b37dd0b

Fix lightening trainer parameters

d1149c6

Add preamble to python files

7cd0410

Preparation for moving OrigBertFuse to fuse core

1441b73

changes

e298cad

changes

b143fe6

changes

1ee6960

changes

f82f344

changes

be3f0aa

changes

a4c07c0

changes

376f70e

changes (fixed percentile generation to include both dynamic and stat…

6b6d6b7

…ic values)

added digitization Op

2ba0162

fixed bugs (running large dataset) and added dropping of short patients

64671e1

added trajectories generation

1e92d8c

added trajectories generation + comments

7fe75b2

added corpus

0a2d128

added corpus

0951d0b

ellabarkan requested a review from mosheraboh January 8, 2023 11:30

ellabarkan added 6 commits January 8, 2023 17:09

black formatter applyied

b35af44

black formatter applied

b46b8ad

Revert "black formatter applied"

cbec12a

This reverts commit b46b8ad.

Revert "black formatter applyied"

29f6e37

This reverts commit b35af44.

black formatter applied

f44df46

new black version formatter applied

56f048c

mosheraboh reviewed Jan 8, 2023

View reviewed changes

ellabarkan added 4 commits January 9, 2023 16:45

fixes following pull request

1710c6a

updated Op of generation trajectories, added postions, and indexes

087acdf

updated Op of generation trajectories, added postions, and indexes

e7051e8

bug fixes in generating trajectory of visits Op

19cb648

Moshe Raboh Moshiko.Raboh@ibm.com and others added 24 commits January 10, 2023 06:49

Merge branch 'master' of github.com:IBM/fuse-med-ml into ehr_transfor…

7fb6c60

…mer_example

less

77736c7

fixes

8b5477d

Merge branch 'ehr_transformer_example' of https://github.com/BiomedSc…

e4219f0

…iAI/fuse-med-ml into ehr_transformer_example

moved WordVocab to local "utils.py"; added path to data pkl to config…

24e61c5

…; other misc changes

fixed bug in passing percentile arg

6d60137

misc fixes

60747b7

Merge branch 'ehr_transformer_example' of github.com:IBM/fuse-med-ml …

d35d525

…into ehr_transformer_example

single head script with vanila transformer

bfd4d31

added option of adding static details as a first special visit

b038bd8

added Readme

8f8efcf

deleted code of the old version of transforment implementation, added…

fcd324a

… comments to files

more comments and flake8 fixes added

ea04710

add auxilary heads

b094f3e

add bert support

eecc052

fixes of Readme files and better configuration of pickle file

c2b17ff

black reformatting

1e48f70

updating figures

2fe806c

document

332dae8

flake8 fix

808d6f4

add transformers to dependnecy lst

c6e569e

Merge branch 'master' into ehr_transformer_example

12bece7

cleanup

440d7d0

change default to bert

b8e2689

mosheraboh requested review from SagiPolaczek and michalozeryflato January 13, 2023 06:46

mosheraboh merged commit 3390fda into master Jan 13, 2023

SagiPolaczek reviewed Jan 16, 2023

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Ehr transformer ICU example #245

Ehr transformer ICU example #245

ellabarkan commented Jan 8, 2023 •

edited by mosheraboh

mosheraboh left a comment

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023 •

edited

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023 •

edited

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh Jan 8, 2023

ellabarkan Jan 9, 2023

mosheraboh commented Jan 13, 2023

mosheraboh commented Jan 13, 2023

SagiPolaczek left a comment

SagiPolaczek Jan 16, 2023

SagiPolaczek Jan 16, 2023

SagiPolaczek Jan 16, 2023

SagiPolaczek Jan 16, 2023

SagiPolaczek Jan 16, 2023

		from fuse.utils.ndict import NDict


		class OpReadDataframeCinC(OpBase):



		class OpAddBMI(OpBase):
		def __call__(self, sample_dict) -> Any:


		d_static = sample_dict["StaticDetails"]

		if ("Height" in d_static.keys()) & ("Weight" in d_static.keys()):



		class OpMapToCategorical(OpBase):
		def __call__(self, sample_dict, percentiles: dict) -> Any:

		STATIC_FIELDS = ["Age", "Gender", "Height", "ICUType", "Weight"]


		class OpAddBMI(OpBase):

Ehr transformer ICU example #245

Ehr transformer ICU example #245

Conversation

ellabarkan commented Jan 8, 2023 • edited by mosheraboh

mosheraboh left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ellabarkan Jan 9, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ellabarkan Jan 9, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mosheraboh commented Jan 13, 2023

mosheraboh commented Jan 13, 2023

SagiPolaczek left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ellabarkan commented Jan 8, 2023 •

edited by mosheraboh

ellabarkan Jan 9, 2023 •

edited

ellabarkan Jan 9, 2023 •

edited