NLP_admissionNotes

The objective of this project is to develop an NLP model capable of extracting diagnoses and diseases from admission notes. Admission notes are created by doctors when a patient is newly admitted to a hospital and typically contain personal information such as age, gender, and various medical conditions. However, the process of entering this information into the hospital's medical system, such as the Hospital Information System (HIS), can be time-consuming and challenging. Therefore, the aim is to train an advanced NLP model that can efficiently and accurately extract relevant diagnoses and diseases from these admission notes. The dataset containing pseudonymized admission notes is publicly available here: http://trec-cds.org/topics2021.xml

Explanation

./hyperparameter_tuning_results.txt contains results from hyperparameter exploration.

./datasets contains all the data I used for testing and training.

./datasets/origin contains all raw data. It is publicly available in http://trec-cds.org/topics2021.xml

./datasets/testing contains all preprocessed data I used for testing the models.

./datasets/training contains all preprocessed data I used for training the models.

./scripts contains all scripts (in Python) which I use for data processing, analysis, training and using the pretrained models.

./scripts/text_extractor.py is used for converting the original .xml file into a csv file with columns: text_id, text

./scripts/entity_extractor.py is used for knitting together two .csv files into one, which can then be used for .conll file preparation.

./scripts/conll_prep.py is used for knitting together two .csv files, one with annotated entities and one generated from text_extractor.py

./scripts/nlp_apply.py takes a pretrained model and applies it to the input text. Represents the result in readable manner.

./scripts/training/nlp_train.py generates an nlp model and automatically tests it. Saves the model to specified folder. Test and training data must be in .conll format.

./scripts/training/nlp_train_stemming.py is the same as ./scripts/training/nlp_train.py, but includes stemming in the preprocessing steps of the pipeline.

./scripts/training/hyperparameter_tuning.py is used for hyperparameter optimization. Implemented Grid Search through a pre-defined search space. Cross-Validation is used for the generated models - evaluated by f1-score. Report is automatically generated.

./scripts/training/hyperparameter_tuning_stemming.py is used for hyperparameter optimization, includes stemming.

./scripts/training/*.conll are used for training and testing respectively. These files have been generated from conll_prep.py

./scripts/training/ner_graphs is used for storing the graphs generated during training the model.

./scripts/training/ner_logs is used for storing the logs generated during training the model.

Here are some great tutorials I used:

https://www.johnsnowlabs.com/the-ultimate-guide-to-building-your-own-ner-model-with-python/

https://github.com/JohnSnowLabs/spark-nlp-workshop/blob/master/tutorials/Certification_Trainings/Healthcare/1.3.prepare_CoNLL_from_annotations_for_NER.ipynb

Dataset Explanation

The dataset contains 75 pseudonymized admission notes. Admission notes always have a similar structure, containing personal data and diagnoses of a single patient. Our goal is to extract these diagnoses and diseases with an NLP model.

Example for an admission note:

"Patient is a 45-year-old man with a history of anaplastic astrocytoma of the spine complicated by severe lower extremity weakness and urinary retention s/p Foley catheter, high-dose steroids, hypertension, and chronic pain. The tumor is located in the T-L spine, unresectable anaplastic astrocytoma s/p radiation. Complicated by progressive lower extremity weakness and urinary retention. Patient initially presented with RLE weakness where his right knee gave out with difficulty walking and right anterior thigh numbness. MRI showed a spinal cord conus mass which was biopsied and found to be anaplastic astrocytoma. Therapy included field radiation t10-l1 followed by 11 cycles of temozolomide 7 days on and 7 days off. This was followed by CPT-11 Weekly x4 with Avastin Q2 weeks/ 2 weeks rest and repeat cycle."

Approach

Goals in data preparation

For training the nlp model, we require two .csv files: one containing the text, and one containing the entities, along with all kinds of metadata.

The text file has following columns: text_id, text

The entities file has following columns: text_id, entity, begin, end, chunk

As a part of data-preprocessing, all text is converted to lower case.

First step: Converting the .xml file

For that, I used the script ./scripts/text_extractor.py - nothing interesting going on here.

Second step: Extracting the entities

I tried doing that manually, but that was way too much work for my liking - especially extracting begin and end of each word. I gave up on that after the first text!

So I tried a different approach. I just wrote each entity I found into a custom .csv file and connected them to their text_id. This means that I got a handcrafted .csv file with the columns: text_id, entity

Then we use the ./scripts/entity_extractor.py script for filling out the rest!

How do we deal with overlaps, though? For instance, we might have "diabetes mellitus" and "diabetes" in the same word.

Interestingly, the NLP model training can handle that just fine - it even yields greater results!

Third step: Data Splitting

Since we only have a small dataset (75 entries), I used the first 70 entries for training, and the last 5 entries for testing.

Fourth step: .conll file preparation

For that, I used the ./scripts/conll_prep.py script, inspired from https://github.com/JohnSnowLabs/spark-nlp-workshop/blob/master/tutorials/Certification_Trainings/Healthcare/1.3.prepare_CoNLL_from_annotations_for_NER.ipynb

First Training Session

The first training session took place with 7 epochs, a learning rate of 0.004, a batch size of 32, and a validation split of 20%. It already produced great results, despite no hyperparameter optimization. It yields an f1-score of about 60%. Results may look like:

entity	token
O	Patient
O	is
O	80yo
O	,
O	has
O	been
O	recently
O	diagnosed
O	with
B-DIAGNOSIS	type
I-DIAGNOSIS	2
I-DIAGNOSIS	diabetes
I-DIAGNOSIS	mellitus
O	.
O	Suffers
O	from
B-DIAGNOSIS	proteinuria
O	and
B-DIAGNOSIS	liver
I-DIAGNOSIS	cirrhosis

B-DIAGNOSIS -> beginning of disease entity

I-DIAGNOSIS -> inside of disease entity

O -> no entity

Hyperparameter Tuning and Optimization

Turns out, this is a little more tricky with spark-nlp than previously thought. Although the NerDLApproach class offerst automatic percentile folds for cross validation (with f1-score, recall etc being logged in every epoch!!), there is no way to automatically perform hyperparameter tuning with this method. We cannot even extract the f1-score. That's kinda disappointing :(

I thought that the CrossValidator-class would be of help for that. However, it doesn't work with our predefined pipeline! It seems to not be built for NER-models after all... Which is unfortunate, and I think that the spark-nlp developer team should take this into consideration for future development. We need both embeddings and graph_builder before our nerTagger. However, as soon as we use the CrossValidator, the vocabulary is not yet ready for validation.

We are in need for a different approach!!

When testing with out test-dataset, we have both ground_truth and prediction columns. We can use the "positive-positive"'s, the "positive-negative"'s, and so on, for manually calculating the f1-score: https://towardsdatascience.com/the-f1-score-bec2bbc38aa6

But we got a function for that! Saves us at least some work.

Normally, we would split the original dataset into three subsets: training, validation (for hyperparameter tuning) and testing (for evaluation of the final model). Our dataset is really small, however. We need the data for training, since diagnoses are of rather subjective nature. We don't even have a subset for validation, so we are really prone to overfitting our model. Rather than going with the holdout approach, cross-validation would be much more suitable, altough more complicated to implement from scratch.

But that's exactly what we're doing, using Pandas and KFold from the sk-learn package: https://towardsdatascience.com/effortless-hyperparameters-tuning-with-apache-spark-20ff93019ef2

Now we need to define a search space for our hyperparameters. From experience, I found the following grid to make sense:

learning rate: [0.001, 0.003, 0.005, 0.01, 0.1], hidden layers: [10, 20], batch size: [8, 16, 32], maximum epochs: [7, 10]

Finally, we iterate through the search space via the grid search approach - "fuck around and find out"

Run the script ./scripts/training/hyperparameter_tuning.py, grab a coffee, take a walk, maybe even start a family... it's gonna take forever to finish.

Analysis took about 3h, and once aborted due to a network error. The results can be seen in "./hyperparameter_tuning_results.txt". I was able to make a few observations during training, which has allowed me to skip a few extra iterations.

A bigger batch size meant less performance. A batch size of 32 meant terrible performance, which can be probably traced back to the fact that the dataset is way too small for such a batch size. A smaller batch size often means independence from the other data in the dataset, which helps us to avoid overfitting. A batch size too small can have the adverse effect of underfitting, though.

The f1-score also got better the more epochs we had. The number of hidden layers did not necessarily have any significant effect on the performance.

Another interesting observation was that with a higher learning rate, the f1-score dropped significantly. It performed much better with lower learning rates.

The final hyperparameters is:

f1-score: 0.9506947550389957 | Learning Rate: 0.003 | Hidden Layers: 10 | Batch Size: 8 | Maximum Epochs: 10

A model trained with these parameters yields surprisingly good results with the test data - it even recognizes diagnoses which I forgot to include in data preprocessing...

token	ground_truth	prediction
The	O	O
patient	O	O
is	O	O
a	O	O
34-year-old	O	O
obese	B-DIAGNOSIS	B-DIAGNOSIS
woman	O	O
who	O	O
comes	O	O
to	O	O
the	O	O
clinic	O	O
with	O	O
weight	O	O
concerns	O	O
.	O	O
She	O	O
is	O	O
165	O	O
cm	O	O
tall	O	O
,	O	O
and	O	O
her	O	O
weight	O	O
is	O	O
113	O	O
kg	O	O
(	O	O
BMI	O	O
:	O	O
41.5	O	O
).	O	O
In	O	O
the	O	O
past	O	O
,	O	O
she	O	O
unsuccessfully	O	O
used	O	O
anti	O	O
obesity	B-DIAGNOSIS	O
agents	O	O
and	O	O
appetite	O	O
suppressants	O	O
.	O	O
She	O	O
is	O	O
complaining	O	O
of	O	O
sleep	B-DIAGNOSIS	B-DIAGNOSIS
apnea	I-DIAGNOSIS	I-DIAGNOSIS
,	O	O
PCO	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
dissatisfaction	O	O
with	O	O
her	O	O
body	O	O
shape	O	O
.	O	O
She	O	O
is	O	O
a	O	O
high-school	O	O
teacher	O	O
.	O	O
She	O	O
is	O	O
married	O	O
for	O	O
5	O	O
years	O	O
.	O	O
She	O	O
doesn't	O	O
use	O	O
any	O	O
contraceptive	O	O
methods	O	O
for	O	O
the	O	O
past	O	O
4	O	O
months	O	O
and	O	O
she	O	O
had	O	O
no	O	O
prior	O	O
pregnancies	O	O
.	O	O
She	O	O
rarely	O	O
exercises	O	O
and	O	O
movement	O	O
seems	O	O
to	O	O
be	O	O
hard	O	O
for	O	O
her	O	O
.	O	O
She	O	O
is	O	O
not	O	O
able	O	O
to	O	O
complete	O	O
the	O	O
four-square	O	O
step	O	O
test	O	O
in	O	O
less	O	O
than	O	O
15	O	O
seconds	O	O
.	O	O
She	O	O
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
any	O	O
drugs	O	O
.	O	O
Her	O	O
BP	O	O
:	O	O
130/80	O	O
,	O	O
HR	O	O
:	O	O
195/min	O	O
and	O	O
her	O	O
BMI	O	O
is	O	O
:	O	O
41.54	O	O
.	O	O
Her	O	O
lab	O	O
results:FBS	O	O
:	O	O
98	O	O
mg/dlTG	O	O
:	O	O
150	O	O
mg/dlCholesterol	O	O
:	O	O
180	O	O
mg/dlLDL	O	O
:	O	O
90	O	O
mg/dlHDL	O	O
:	O	O
35	O	O
mg/dlShe	O	O
is	O	O
considering	O	O
a	O	O
laparoscopic	O	O
gastric	O	B-DIAGNOSIS
bypass	O	I-DIAGNOSIS
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
16-year-old	O	O
girl	O	O
recently	O	O
diagnosed	O	O
with	O	O
myasthenia	B-DIAGNOSIS	B-DIAGNOSIS
gravis	I-DIAGNOSIS	I-DIAGNOSIS
class	O	I-DIAGNOSIS
IIa	O	I-DIAGNOSIS
.	O	O
She	O	O
complains	O	O
of	O	O
diplopia	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
weakness	B-DIAGNOSIS	B-DIAGNOSIS
affecting	O	O
in	O	O
her	O	O
upper	O	O
extremities	O	O
.	O	O
She	O	O
had	O	O
a	O	O
positive	O	O
anti-AChR	O	O
antibody	O	O
test	O	O
,	O	O
and	O	O
her	O	O
single	O	O
fiber	O	O
electromyography	O	O
(	O	O
SFEMG	O	O
)	O	O
was	O	O
positive	O	O
.	O	O
She	O	O
is	O	O
on	O	O
acetylcholinesterase	O	O
inhibitor	O	O
treatment	O	O
combined	O	O
with	O	O
immunosuppressants	O	O
.	O	O
But	O	O
she	O	O
still	O	O
has	O	O
some	O	O
symptoms	O	O
.	O	O
She	O	O
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
illicit	O	O
drugs	O	O
.	O	O
She	O	O
is	O	O
not	O	O
sexually	O	O
active	O	O
,	O	O
and	O	O
her	O	O
menses	O	O
are	O	O
regular	O	O
.	O	O
Her	O	O
physical	O	O
exam	O	O
and	O	O
lab	O	O
studies	O	O
are	O	O
not	O	O
remarkable	O	O
for	O	O
any	O	O
other	O	O
abnormalities.BP	O	O
:	O	O
110/75Hgb	O	O
:	O	O
11	O	O
g/dlWBC	O	O
:	O	O
8000	O	O
/mm3Plt	O	O
:	O	O
300000	O	O
/mlCreatinine	O	O
:	O	O
0.5	O	O
mg/dlBUN	O	O
:	O	O
10	O	O
mg/dlBeta	O	O
hcg	O	O
:	O	O
negative	O	O
for	O	O
pregnancy	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
3-day-old	O	O
female	O	O
infant	O	O
with	O	O
jaundice	B-DIAGNOSIS	B-DIAGNOSIS
that	O	O
started	O	O
one	O	O
day	O	O
ago	O	O
.	O	O
She	O	O
was	O	O
born	O	O
at	O	O
34w	O	O
of	O	O
gestation	O	O
and	O	O
kept	O	O
in	O	O
an	O	O
incubator	O	O
due	O	O
to	O	O
her	O	O
gestational	O	B-DIAGNOSIS
age	O	I-DIAGNOSIS
.	O	O
Vital	O	O
signs	O	O
were	O	O
reported	O	O
as	O	O
:	O	O
axillary	O	O
temperature	O	O
:	O	O
36.3°C	O	O
,	O	O
heart	O	O
rate	O	O
:	O	O
154	O	O
beats/min	O	O
,	O	O
respiratory	O	O
rate	O	O
:	O	O
37	O	O
breaths/min	O	O
,	O	O
and	O	O
blood	O	O
pressure	O	O
:	O	O
65/33	O	O
mm	O	O
Hg	O	O
.	O	O
Her	O	O
weight	O	O
is	O	O
2.1	O	O
kg	O	O
,	O	O
length	O	O
is	O	O
45	O	O
cm	O	O
,	O	O
and	O	O
head	O	O
circumference	O	O
32	O	O
cm	O	O
.	O	O
She	O	O
presents	O	O
with	O	O
yellow	B-DIAGNOSIS	O
sclera	I-DIAGNOSIS	O
and	O	O
icteric	B-DIAGNOSIS	B-DIAGNOSIS
body	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
Her	O	O
liver	O	O
and	O	O
spleen	O	O
are	O	O
normal	O	O
to	O	O
palpation.Laboratory	O	O
results	O	O
are	O	O
as	O	O
follows	O	O
:	O	O
Serum	O	O
total	O	O
bilirubin	O	O
:	O	O
21.02	O	O
mg/dLDirect	O	O
bilirubin	O	O
of	O	O
2.04	O	O
mg/dLAST	O	O
:	O	O
37	O	O
U/LALT	O	O
:	O	O
20	O	O
U/LGGT	O	O
:	O	O
745	O	O
U/LAlkaline	O	O
phosphatase	O	O
:	O	O
531	O	O
U/LCreatinine	O	O
:	O	O
0.3	O	O
mg/dLUrea	O	O
:	O	O
29	O	O
mg/dLNa	O	O
:	O	O
147	O	O
mEq/LK	O	O
:	O	O
4.5	O	O
mEq/LCRP	O	O
:	O	O
3	O	O
mg/LComplete	O	O
blood	O	O
cell	O	O
count	O	O
within	O	O
the	O	O
normal	O	O
range.She	O	O
is	O	O
diagnosed	O	O
with	O	O
neonatal	B-DIAGNOSIS	B-DIAGNOSIS
jaundice	I-DIAGNOSIS	I-DIAGNOSIS
that	O	O
may	O	O
require	O	O
phototherapy	O	O
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
53-year-old	O	O
man	O	O
complaining	O	O
of	O	O
frequent	O	O
headaches	B-DIAGNOSIS	B-DIAGNOSIS
,	O	O
generalized	O	O
bone	B-DIAGNOSIS	B-DIAGNOSIS
pain	I-DIAGNOSIS	I-DIAGNOSIS
and	O	O
difficulty	O	O
chewing	O	O
that	O	O
started	O	O
6	O	O
years	O	O
ago	O	O
and	O	O
is	O	O
getting	O	O
worse	O	O
.	O	O
Examination	O	O
shows	O	O
bilateral	O	B-DIAGNOSIS
swellings	B-DIAGNOSIS	I-DIAGNOSIS
around	O	O
the	O	O
molars	O	O
.	O	O
The	O	O
swellings	B-DIAGNOSIS	B-DIAGNOSIS
have	O	O
increased	O	O
since	O	O
his	O	O
last	O	O
examination	O	O
.	O	O
Several	O	O
extraoral	O	O
lesions	B-DIAGNOSIS	O
of	I-DIAGNOSIS	O
the	I-DIAGNOSIS	O
head	I-DIAGNOSIS	O
and	I-DIAGNOSIS	O
face	I-DIAGNOSIS	O
are	O	O
detected	O	O
.	O	O
The	O	O
swellings	B-DIAGNOSIS	O
are	O	O
non-tender	O	O
and	O	O
attached	O	O
to	O	O
the	O	O
underlying	O	O
bone	O	O
.	O	O
Further	O	O
evaluation	O	O
shows	O	O
increased	O	O
uptake	O	O
of	O	O
radioactive	O	O
substance	O	O
as	O	O
well	O	O
as	O	O
an	O	O
increase	O	O
in	O	O
urinary	O	O
pyridinoline	O	O
.	O	O
The	O	O
serum	O	O
alkaline	O	O
phosphatase	O	O
is	O	O
300	O	O
IU/L	O	O
(	O	O
the	O	O
normal	O	O
range	O	O
is	O	O
44	O	O
-	O	O
147	O	O
IU/L	O	O
).	O	O
The	O	O
patient's	O	O
sister	O	O
had	O	O
the	O	O
same	O	O
problems	O	O
.	O	O
She	O	O
was	O	O
diagnosed	O	O
with	O	O
Paget's	B-DIAGNOSIS	B-DIAGNOSIS
disease	I-DIAGNOSIS	I-DIAGNOSIS
of	I-DIAGNOSIS	I-DIAGNOSIS
bone	I-DIAGNOSIS	I-DIAGNOSIS
when	O	O
she	O	O
was	O	O
52	O	O
years	O	O
old	O	O
.	O	O
The	O	O
diagnosis	O	O
of	O	O
Paget's	O	B-DIAGNOSIS
Disease	O	I-DIAGNOSIS
of	O	I-DIAGNOSIS
Bone	O	I-DIAGNOSIS
is	O	O
confirmed	O	O
and	O	O
Bisphosphonate	O	O
will	O	O
be	O	O
started	O	O
as	O	O
first-line	O	O
therapy	O	O
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
55-year-old	O	O
man	O	O
who	O	O
was	O	O
recently	O	O
diagnosed	O	O
with	O	O
Parkinson's	B-DIAGNOSIS	B-DIAGNOSIS
disease	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
He	O	O
is	O	O
complaining	O	O
of	O	O
slowness	B-DIAGNOSIS	B-DIAGNOSIS
of	I-DIAGNOSIS	I-DIAGNOSIS
movement	I-DIAGNOSIS	I-DIAGNOSIS
and	O	O
tremors	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
His	O	O
disease	O	O
is	O	O
ranked	O	O
as	O	O
mild	O	O
,	O	O
Hoehn-Yahr	B-DIAGNOSIS	B-DIAGNOSIS
Stage	I-DIAGNOSIS	I-DIAGNOSIS
I	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
His	O	O
past	O	O
medical	O	O
history	O	O
is	O	O
significant	O	O
for	O	O
hypertension	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
hypercholesterolemia	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
He	O	O
lives	O	O
with	O	O
his	O	O
wife	O	O
.	O	O
They	O	O
have	O	O
three	O	O
children	O	O
.	O	O
He	O	O
used	O	O
to	O	O
be	O	O
active	O	O
with	O	O
gardening	O	O
before	O	O
his	O	O
diagnosis	O	O
.	O	O
He	O	O
complains	O	O
of	O	O
shaking	O	B-DIAGNOSIS
and	O	I-DIAGNOSIS
slow	O	I-DIAGNOSIS
movement	O	I-DIAGNOSIS
.	O	O
He	O	O
had	O	O
difficulty	O	O
entering	O	O
through	O	O
a	O	O
door	O	O
,	O	O
as	O	O
he	O	O
was	O	O
frozen	O	O
and	O	O
needed	O	O
guidance	O	O
to	O	O
step	O	O
in	O	O
.	O	O
His	O	O
handwriting	O	O
is	O	O
getting	O	O
smaller	O	O
.	O	O
He	O	O
is	O	O
offered	O	O
Levodopa	O	O
and	O	O
Trihexyphenidyl	O	O
.	O	O
He	O	O
is	O	O
an	O	O
alert	O	O
and	O	O
cooperative	O	O
man	O	O
who	O	O
does	O	O
not	O	O
have	O	O
any	O	O
signs	O	O
of	O	O
dementia	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
He	O	O
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
any	O	O
illicit	O	O
drugs	O	O
.	O	O

As we can see, hyperparameter tuning is extremely important - we achieved an f1-score of about 95%, while in our first model we had an f1-score of about 60%.

Stemming

Stemming might prove useful for us since it can dramatically enhance our model's vocabulary, despite us having quite little training data.

Here's a good use-case. We got the diagnosis "seasonal allergy" and "seasonal allergies". Via stemming, we reduce these words to their stem. Our model would only have to memorize "allergi".

The implementation is relatively straightforward: https://sparknlp.org/api/python/reference/autosummary/sparknlp/annotator/stemmer/

In testing the new model's performance, I realized that stemming might not even be the best choice - the f1-score slightly decreased. Hyperparameter Tuning might have to be repeated.

The training algorithm where stemming is added to the pipeline can be found here: ./scripts/training/nlp_train_stemming.py

Hyperparameter Optimization cont'd

I added a new script which should perform the same Grid Search etc. for our stemming algorithm. Unfortunately, it got the same optimal hyperparameters, but with worse f1-score:

f1-score: 0.939280697131209 | Learning Rate: 0.003 | Hidden Layers: 10 | Batch Size: 8 | Maximum Epochs: 10

Adding More Entities

This time, I manually added 3 more entities to our training data. These include...

AGE: This should extract the age of our patient.

GENDER: ...gender of our patient.

NEGATIVE: Sometimes, diseases are entered, but negated. We do not want "allergy" highlighted as a diagnosis if it is negated, e.g. as "Patient is 80yo and has no allergies.".

First results already look somewhat promising:

entity	token
O	this
B-AGE	70
I-AGE	years
I-AGE	old
B-GENDER	gentleman
O	is
O	positive
O	for
B-DIAGNOSIS	seasonal
I-DIAGNOSIS	allergies
O	.
O	has
B-NEGATIVE	no
I-NEGATIVE	history
I-NEGATIVE	of
I-NEGATIVE	dm

Other input is just terrible:

entity	token
O	patient
O	is
O	80-year-old
O	woman
O	and
O	has
O	tested
O	negative
O	to
O	dm
O	type
O	1

Next steps to polish our model:

Cleaning up our training data. We need to make sure that it's safe and ready for training. Remember the "shit in, shit out" principle...

And some more Hyperparameter Tuning!

Maybe Stemming will make a good comeback?

Hyperparameter Tuning cont'd cont'd

That wasn't really spectacular. Whether with or without stemming, it produced the same results as before. However, I realized that the more epochs we have, the better our model would perform. But how much of that is overfitting? We would find out if we had more testing data... Anyways, I let all parameters run through with 5 epochs. Then, I picked the best performing params, and let them have 30 epochs. I found an increase of performance in a range of 5% to 7% - quite remarkable, actually. Stemming would still slack behind, though. These are the results without stemming:

f1-score: 0.920768469152548 | Learning Rate: 0.003 |Hidden Layers: 10 |Batch Size: 8 | Maximum Epochs: 30

And with stemming:

f1-score: 0.9022492993322529 | Learning Rate: 0.003 | Hidden Layers: 10 | Batch Size: 16 | Maximum Epochs: 30

Eventually, the final model would perform quite well on test data:

token	ground_truth	prediction
The	O	O
patient	O	O
is	O	O
a	O	O
34-year-old	B-AGE	B-AGE
obese	B-DIAGNOSIS	B-DIAGNOSIS
woman	B-GENDER	B-GENDER
who	O	O
comes	O	O
to	O	O
the	O	O
clinic	O	O
with	O	O
weight	O	B-DIAGNOSIS
concerns	O	I-DIAGNOSIS
.	O	O
She	B-GENDER	B-GENDER
is	O	O
165	O	O
cm	O	O
tall	O	O
,	O	O
and	O	O
her	O	O
weight	O	O
is	O	O
113	O	O
kg	O	O
(	O	O
BMI	O	O
:	O	O
41.5	O	O
).	O	O
In	O	O
the	O	O
past	O	O
,	O	O
she	B-GENDER	B-GENDER
unsuccessfully	O	O
used	O	O
anti	O	O
obesity	O	O
agents	O	O
and	O	O
appetite	O	O
suppressants	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
complaining	O	O
of	O	O
sleep	B-DIAGNOSIS	B-DIAGNOSIS
apnea	I-DIAGNOSIS	I-DIAGNOSIS
,	O	O
PCO	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
dissatisfaction	B-DIAGNOSIS	B-DIAGNOSIS
with	O	O
her	O	O
body	O	O
shape	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
a	O	O
high-school	O	O
teacher	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
married	O	O
for	O	O
5	O	O
years	O	O
.	O	O
She	B-GENDER	B-GENDER
doesn't	O	O
use	O	O
any	O	O
contraceptive	O	O
methods	O	O
for	O	O
the	O	O
past	O	O
4	O	O
months	O	O
and	O	O
she	B-GENDER	B-GENDER
had	O	O
no	B-NEGATIVE	B-NEGATIVE
prior	I-NEGATIVE	I-NEGATIVE
pregnancies	I-NEGATIVE	I-NEGATIVE
.	O	O
She	B-GENDER	B-GENDER
rarely	O	O
exercises	O	O
and	O	O
movement	O	O
seems	O	O
to	O	O
be	O	O
hard	O	O
for	O	O
her	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
not	O	O
able	O	O
to	O	O
complete	O	O
the	O	O
four-square	O	O
step	O	O
test	O	O
in	O	O
less	O	O
than	O	O
15	O	O
seconds	O	O
.	O	O
She	B-GENDER	B-GENDER
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
any	O	O
drugs	O	O
.	O	O
Her	O	O
BP	O	O
:	O	O
130/80	O	O
,	O	O
HR	O	O
:	O	O
195/min	O	O
and	O	O
her	O	O
BMI	O	O
is	O	O
:	O	O
41.54	O	O
.	O	O
Her	O	O
lab	O	O
results:FBS	O	O
:	O	O
98	O	O
mg/dlTG	O	O
:	O	O
150	O	O
mg/dlCholesterol	O	O
:	O	O
180	O	O
mg/dlLDL	O	O
:	O	O
90	O	O
mg/dlHDL	O	O
:	O	O
35	O	O
mg/dlShe	O	O
is	O	O
considering	O	O
a	O	O
laparoscopic	O	O
gastric	O	O
bypass	O	O
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
16-year-old	B-AGE	B-AGE
girl	B-GENDER	B-GENDER
recently	O	O
diagnosed	O	O
with	O	O
myasthenia	B-DIAGNOSIS	B-DIAGNOSIS
gravis	I-DIAGNOSIS	I-DIAGNOSIS
class	I-DIAGNOSIS	I-DIAGNOSIS
IIa	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
She	B-GENDER	B-GENDER
complains	O	O
of	O	O
diplopia	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
weakness	B-DIAGNOSIS	B-DIAGNOSIS
affecting	O	I-DIAGNOSIS
in	O	O
her	O	O
upper	O	O
extremities	O	O
.	O	O
She	B-GENDER	B-GENDER
had	O	O
a	O	O
positive	O	O
anti-AChR	O	O
antibody	O	O
test	O	O
,	O	O
and	O	O
her	O	O
single	O	O
fiber	O	O
electromyography	O	O
(	O	O
SFEMG	O	O
)	O	O
was	O	O
positive	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
on	O	O
acetylcholinesterase	O	O
inhibitor	O	O
treatment	O	O
combined	O	O
with	O	O
immunosuppressants	O	O
.	O	O
But	O	O
she	B-GENDER	B-GENDER
still	O	O
has	O	O
some	O	O
symptoms	O	O
.	O	O
She	B-GENDER	B-GENDER
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
illicit	O	O
drugs	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
not	O	O
sexually	O	O
active	O	O
,	O	O
and	O	O
her	O	O
menses	O	O
are	O	O
regular	O	O
.	O	O
Her	O	O
physical	O	O
exam	O	O
and	O	O
lab	O	O
studies	O	O
are	O	O
not	O	O
remarkable	O	O
for	O	O
any	O	O
other	O	O
abnormalities.BP	O	O
:	O	O
110/75Hgb	O	O
:	O	O
11	O	O
g/dlWBC	O	O
:	O	O
8000	O	O
/mm3Plt	O	O
:	O	O
300000	O	O
/mlCreatinine	O	O
:	O	O
0.5	O	O
mg/dlBUN	O	O
:	O	O
10	O	O
mg/dlBeta	O	O
hcg	O	O
:	O	O
negative	O	O
for	O	O
pregnancy	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
3-day-old	B-AGE	B-AGE
female	B-GENDER	B-GENDER
infant	O	O
with	O	O
jaundice	B-DIAGNOSIS	B-DIAGNOSIS
that	O	O
started	O	O
one	O	O
day	O	O
ago	O	O
.	O	O
She	B-GENDER	B-GENDER
was	O	O
born	O	O
at	O	O
34w	O	O
of	O	O
gestation	O	O
and	O	O
kept	O	O
in	O	O
an	O	O
incubator	O	O
due	O	O
to	O	O
her	O	O
gestational	O	O
age	O	O
.	O	O
Vital	O	O
signs	O	O
were	O	O
reported	O	O
as	O	O
:	O	O
axillary	O	O
temperature	O	O
:	O	O
36.3°C	O	O
,	O	O
heart	O	O
rate	O	O
:	O	O
154	O	O
beats/min	O	O
,	O	O
respiratory	O	O
rate	O	O
:	O	O
37	O	O
breaths/min	O	O
,	O	O
and	O	O
blood	O	O
pressure	O	O
:	O	O
65/33	O	O
mm	O	O
Hg	O	O
.	O	O
Her	O	O
weight	O	O
is	O	O
2.1	O	O
kg	O	O
,	O	O
length	O	O
is	O	O
45	O	O
cm	O	O
,	O	O
and	O	O
head	O	O
circumference	O	O
32	O	O
cm	O	O
.	O	O
She	B-GENDER	B-GENDER
presents	O	O
with	O	O
yellow	B-DIAGNOSIS	O
sclera	I-DIAGNOSIS	B-DIAGNOSIS
and	O	O
icteric	B-DIAGNOSIS	B-DIAGNOSIS
body	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
Her	O	O
liver	O	O
and	O	O
spleen	O	O
are	O	O
normal	O	O
to	O	O
palpation	O	O
.	O	O
Laboratory	O	O
results	O	O
are	O	O
as	O	O
follows	O	O
:	O	O
Serum	O	O
total	O	O
bilirubin	O	O
:	O	O
21.02	O	O
mg/dLDirect	O	O
bilirubin	O	O
of	O	O
2.04	O	O
mg/dLAST	O	O
:	O	O
37	O	O
U/LALT	O	O
:	O	O
20	O	O
U/LGGT	O	O
:	O	O
745	O	O
U/LAlkaline	O	O
phosphatase	O	O
:	O	O
531	O	O
U/LCreatinine	O	O
:	O	O
0.3	O	O
mg/dLUrea	O	O
:	O	O
29	O	O
mg/dLNa	O	O
:	O	O
147	O	O
mEq/LK	O	O
:	O	O
4.5	O	O
mEq/LCRP	O	O
:	O	O
3	O	O
mg/LComplete	O	O
blood	O	O
cell	O	O
count	O	O
within	O	O
the	O	O
normal	O	O
range	O	O
.	O	O
She	B-GENDER	B-GENDER
is	O	O
diagnosed	O	O
with	O	O
neonatal	B-DIAGNOSIS	O
jaundice	I-DIAGNOSIS	B-DIAGNOSIS
that	O	O
may	O	O
require	O	O
phototherapy	O	O
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
53-year-old	B-AGE	B-AGE
man	B-GENDER	B-GENDER
complaining	O	O
of	O	O
frequent	O	O
headaches	B-DIAGNOSIS	B-DIAGNOSIS
,	O	O
generalized	O	O
bone	B-DIAGNOSIS	B-DIAGNOSIS
pain	I-DIAGNOSIS	I-DIAGNOSIS
and	O	O
difficulty	B-DIAGNOSIS	B-DIAGNOSIS
chewing	I-DIAGNOSIS	I-DIAGNOSIS
that	O	O
started	O	O
6	O	O
years	O	O
ago	O	O
and	O	O
is	O	O
getting	O	O
worse	O	O
.	O	O
Examination	O	O
shows	O	O
bilateral	O	O
swellings	B-DIAGNOSIS	B-DIAGNOSIS
around	O	O
the	O	O
molars	O	O
.	O	O
The	O	O
swellings	B-DIAGNOSIS	B-DIAGNOSIS
have	O	O
increased	O	O
since	O	O
his	O	O
last	O	O
examination	O	O
.	O	O
Several	O	O
extraoral	O	O
lesions	B-DIAGNOSIS	O
of	I-DIAGNOSIS	O
the	I-DIAGNOSIS	O
head	I-DIAGNOSIS	O
and	I-DIAGNOSIS	O
face	I-DIAGNOSIS	O
are	O	O
detected	O	O
.	O	O
The	O	O
swellings	B-DIAGNOSIS	B-DIAGNOSIS
are	O	O
non-tender	O	O
and	O	O
attached	O	O
to	O	O
the	O	O
underlying	O	O
bone	O	O
.	O	O
Further	O	O
evaluation	O	O
shows	O	O
increased	O	O
uptake	O	B-DIAGNOSIS
of	O	O
radioactive	O	O
substance	O	O
as	O	O
well	O	O
as	O	O
an	O	O
increase	B-DIAGNOSIS	B-DIAGNOSIS
in	I-DIAGNOSIS	I-DIAGNOSIS
urinary	I-DIAGNOSIS	I-DIAGNOSIS
pyridinoline	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
The	O	O
serum	O	O
alkaline	O	O
phosphatase	O	O
is	O	O
300	O	O
IU/L	O	O
(	O	O
the	O	O
normal	O	O
range	O	O
is	O	O
44	O	O
-	O	O
147	O	O
IU/L	O	O
).	O	O
The	O	O
patient's	O	O
sister	O	O
had	O	O
the	O	O
same	O	O
problems	O	O
.	O	O
She	B-GENDER	B-GENDER
was	O	O
diagnosed	O	O
with	O	O
Paget's	B-DIAGNOSIS	B-DIAGNOSIS
disease	I-DIAGNOSIS	I-DIAGNOSIS
of	I-DIAGNOSIS	I-DIAGNOSIS
bone	I-DIAGNOSIS	I-DIAGNOSIS
when	O	O
she	B-GENDER	B-GENDER
was	O	O
52	B-AGE	O
years	I-AGE	O
old	I-AGE	O
.	O	O
The	O	O
diagnosis	O	O
of	O	O
Paget's	B-DIAGNOSIS	B-DIAGNOSIS
Disease	I-DIAGNOSIS	I-DIAGNOSIS
of	I-DIAGNOSIS	O
Bone	I-DIAGNOSIS	O
is	O	O
confirmed	O	O
and	O	O
Bisphosphonate	O	O
will	O	O
be	O	O
started	O	O
as	O	O
first-line	O	O
therapy	O	O
.	O	O
The	O	O
patient	O	O
is	O	O
a	O	O
55-year-old	B-AGE	B-AGE
man	B-GENDER	B-GENDER
who	O	O
was	O	O
recently	O	O
diagnosed	O	O
with	O	O
Parkinson's	B-DIAGNOSIS	B-DIAGNOSIS
disease	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
He	B-GENDER	B-GENDER
is	O	O
complaining	O	O
of	O	O
slowness	B-DIAGNOSIS	B-DIAGNOSIS
of	I-DIAGNOSIS	I-DIAGNOSIS
movement	I-DIAGNOSIS	I-DIAGNOSIS
and	O	O
tremors	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
His	O	O
disease	O	O
is	O	O
ranked	O	O
as	O	O
mild	O	O
,	O	O
Hoehn-Yahr	B-DIAGNOSIS	B-DIAGNOSIS
Stage	I-DIAGNOSIS	I-DIAGNOSIS
I	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
His	O	O
past	O	O
medical	O	O
history	O	O
is	O	O
significant	O	O
for	O	O
hypertension	B-DIAGNOSIS	B-DIAGNOSIS
and	O	O
hypercholesterolemia	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
He	B-GENDER	B-GENDER
lives	O	O
with	O	O
his	O	O
wife	O	O
.	O	O
They	O	O
have	O	O
three	O	O
children	O	O
.	O	O
He	B-GENDER	B-GENDER
used	O	O
to	O	O
be	O	O
active	O	O
with	O	O
gardening	O	O
before	O	O
his	O	O
diagnosis	O	O
.	O	O
He	B-GENDER	B-GENDER
complains	O	O
of	O	O
shaking	B-DIAGNOSIS	B-DIAGNOSIS
and	O	I-DIAGNOSIS
slow	B-DIAGNOSIS	I-DIAGNOSIS
movement	I-DIAGNOSIS	I-DIAGNOSIS
.	O	O
He	B-GENDER	B-GENDER
had	O	O
difficulty	O	O
entering	O	O
through	O	O
a	O	O
door	O	O
,	O	O
as	O	O
he	B-GENDER	B-GENDER
was	O	O
frozen	O	O
and	O	O
needed	O	O
guidance	O	O
to	O	O
step	O	O
in	O	O
.	O	O
His	O	O
handwriting	O	O
is	O	O
getting	O	O
smaller	O	O
.	O	O
He	B-GENDER	B-GENDER
is	O	O
offered	O	O
Levodopa	O	O
and	O	O
Trihexyphenidyl	O	O
.	O	O
He	B-GENDER	B-GENDER
is	O	O
an	O	O
alert	O	O
and	O	O
cooperative	O	O
man	B-GENDER	B-GENDER
who	O	O
does	O	O
not	O	O
have	O	O
any	O	O
signs	O	O
of	O	O
dementia	B-DIAGNOSIS	B-DIAGNOSIS
.	O	O
He	B-GENDER	B-GENDER
does	O	O
not	O	O
smoke	O	O
or	O	O
use	O	O
any	O	O
illicit	O	O
drugs	O	O
.	O	O

Future

-> ???