2
$\begingroup$

I have been searching for 2 weeks and I got no where so far.

There is a list of diseases

Diabetes 
Heart Transplant
Fingertip amputation
Injury by sharp tools
.
.
.

and My dataset is a list of medical text reports.

the training dataset has diseases that can be generated from each record

example that I made

This patient has suffered a cut while using his Carving Chisel and led to losing the fingertip therefore we had to operate to sew the tip.....

from this report we get these diseases

1- Injury by sharp tools
2- Fingertip amputation
3- Sewing injury

another report results may have 3 or less or more diseases

I have searched a lot I found many examples about NLP classification

where a text will be classified into Sports, Politics, Culture, Science, etc.

I found NER where person names, locations, dates, etc can be extracted from a text.

But did not find anything for a single text could have multiple values (similar to my dataset)

I dont know where to start.

Could anyone please help me finding what is the name of this method of extracting list of issues from a text?

Edit

What else do I need to exclude the negations, if the report says

"This patient has stomach problem but not diabetes "

How can I make AI understand there is a negation (NOT) before diabetes so it should not be included?

So the result will be

stomach problem

as diabetes should be excluded because of the negation word

$\endgroup$
1
$\begingroup$

You might want to look at Multi-Label-Classification . If you have sufficient number of samples as your training data, you can build a model that can predict more than one label for a test sample. You can find more about the implementation at the sklearn page for the same here .

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks heaps, what else do I need to exclude the negations, if the report says "This patient has stomach problem but not diabetes " How can I make AI understand there is a negation (NOT) before diabetes so it should not be included? $\endgroup$ – asmgx Jul 24 '19 at 1:20
  • $\begingroup$ The idea is if you have enough samples like these, you can count on the models to learn this automatically. If you want to create features, you can create bigram features that will help the model distinguish diabetes vs not diabetes. $\endgroup$ – Gyan Ranjan Jul 25 '19 at 14:33
1
$\begingroup$

This task would be very close to topic modeling, which is usually addressed as a multi-label classification problem.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.