OHDSI Home | Forums | Wiki | Github

Drug to disease - Cross domain mapping

Implementers
, , , ,
#1

Hello Folks,

I would like to know whether it is possible to know the disease from drug name?

For ex: if a person has Metformin, we can say that it is prescribed for diabetes.

I understand that there might be cases where drugs might be prescribed in combination for other diseases as well. But am looking for a simple most frequently prescribed drugs for a disease.

When we hear the name Metformin, we know that it’s for Diabetes.

Either by using OHDSI concept tables/tools or outside do we have any relationship tree like thingy which will help us know what drugs are prescribed to what diseases?

For ex: In our ATHENA, we see “CONCEPT_HIERARCHY” table which shows the relationship in the same domain. I am looking for something similar across domains.

Is it possible? if not ready-made option/utility, you can guide me to place where I can find something related to this is also helpful. I can see whether I can write any code to get the way I want

Anyone has any experience doing this?

1 Like
#2

@Akshay:

Do you have a use case for that?

This subject has been discussed many - times - before. Bottom line: It doesn’t work.

#3

Here is a use case for Procedure to Condition domain mappings or indication coding. This is not for drugs but the same concept. Also not OMOP but could be.

“Identifying Cancer-Directed Surgeries in Medicare Claims: A Validation Study Using SEER-Medicare Data”

https://ascopubs.org/doi/pdfdirect/10.1200/CCI.18.00093

Why does this not “work”? We just had the use case of characterizing ‘active’ cancer treatment for COVID-19 patients and had to cobble something like this together. Would be nice for these kinds of tools to be in the OHDSI vocabulary toolbox.

Sometimes you don’t only want the data to talk to you but you want to talk to the data.

#4

The study measured sensitivity and specificity (without calling it that way) of procedures that are used for cancer treatment (i.e. cancer treatment procedures). It found out, from the data, that some procedures have a higher sensitivity and specificity than others. Mastectomy for example is pretty much only used for breast cancer. There is no other condition that requires the breasts to be removed. For hysterectomy it is different. Cancer is one indication to remove the uterus, endometriosis, fibroids and prolapse are others. So, this use case you are quoting is totally legit: For each treatment tell me the range of indications, and the strength of association.

What folks want is for the vocabulary to tell them the indication from a priori knowledge perspective in absolute terms. But that can’t exist as a binary solution: It depends where you set the cutoff for specificity and sensitivity.

The other problem is the granularity of the condition. Where do you set the line? Any tumor, including benign? Cancer? Cancer by organ (schema as NAACCR calls it)? By histology? By stage? Prostate cancers of different stages have very different treatments, and chasing that would be an arduous task.

Bottom line: If you want indications you need an analytic like in that paper. I don’t think there is an a priori solution, no matter how much empathy I have for this desire.

t