Graphical Recommender system

Hi! I’m making a prescription recommender system for doctors. Objectives: P(observation(symptoms) | disease(x), patient(p)) in real time, suggesting prescribing drugs on a speciality basis and much more as we keep on adding.
The implementations on the internet I’ve found are naïve and can’t really be useful to doctors. Therefore, building one from graphs and GNNs. The project is underway and looking for collaborators, hit me up if too find this interesting.


That is a massive project. You have to be able to read the entire medical codecs and link that with the correct Loinc codes (for observations) and ICD codes (for symptoms) and from there link to a prescription , treatment (surgery or others…)… I hope you have a big team to help you.

I find this interesting. Could you be sharing your progress as you go along? I would really want to see how this turn out

1 Like

I’ll share updates

not much. the data seems to be problem here and medical corpora is too large to do any mapping. I’m trying new ways, nothing with DL yet

1 Like

hi, @metamyth I am bit booked with project at the moment, however if you need help with graph databases, let me know I am working with neo4j for over 3 years now (and I highly recommend neo4j GDS for a check out) (I am not affiliated with the company)

Maybe I am understanding it incorrectly, I would imagine, this opens a huge can of worms legally and ethically. The way I am reading it, the goal is to go directly from symptoms to treatment without an actual diagnosis? if there is a diagnosis, you would start with what medicare and other insurances approve for each diagnosis.

Randomly picking a drug from PDR.
Acetazolamide , it has a lot of verbiage for drug interactions and adverse reactions.
Acetazolamide Tablets (acetazolamide) dose, indications, adverse effects, interactions... from

hi :wave:. I’ve been using neo4j at places in the model already. Not really much with GDS until now. I think something like embeddings could be useful. What do you think?

hey. I’m making this thing legit and making sure of legalities. Also, yes. The drug indications mean that diagnosis is being considered as well

1 Like

embeddings are a great start, also pagerank score could give you extra info about the underlying importance

1 Like

thx for the reply

Hi Amit,

I am interested in this work. Would you please share your progress so far? Did you achieve anything? How are you coordinating the team and managing the tasks? I would like to see the data if publicly available to quickly prototype something first as the starting point for discussions and deliberation that could be improved as we proceed with the work. Usually medications are given after diagnoses are finalised. So from Symptoms & Observations, diagnoses shall be predicted and likewise medication from the diagnoses. Let me know how to get involved.

Thanks a lot.

Hi @bilalUWE. I’m the only one who’s building it currently, We can have a short chat about your queries and starting out. You can schedule one with me here

1 Like

Hi @metamyth.

Thanks for the response. Cool.

I’ve created a discord server to keep track of the work on this idea. For all who wants to collaborate are free to join. The link to join the server is: Healthcare AI. We can continue on our discussions from there.

Pl join and let us have a chat.

1 Like