Spoken Dialogue System for Rx Medications (sds4rx)

This application uses the Mindmeld framework to build a conversational agent that inquires and collects information about a patient's medication record. The use case for this application is medication reconciliation, which is the process of creating an accurate record of a patient's current medication regimen.

This agent aims to accomplish the following tasks:

• Collect medication brand, primary ingredients, dosage, and frequency from natural language
• Inquire user for missing information about their medications
• Cross check information with RxNorm, a database of standardized perscription drug information maintained by the US Library of Medicine

This application is currently under development.

NLP Model Hiearchy

The medication domain is to support the following intents:

• add_medication — Patient reports a series of new medications that are to be added to their record.
• update_record — Agent interviews patient about medications on their existing record and updates it as needed.
• start_over — Clears information from the current conversation and goes back to the start message.

The general domain is to support the following intents:

• greet — Begins the interaction. Greet the patient and inform them about the agent's functionality.
• exit — Ends the current interaction and say goodbye to the patient.
• help — State the agent's functionality and recomend several actions they can take in case the patient gets stuck.
• unsupported — Patient has inquired for information outside the scope of the agent's knowledge.

File structure

• medication_history_examples — Examples of medication reconciliation dialogue from Vanderbilt University Medical Center's training videos.
• sds4rx — Mindmeld application
  • data — JSON knowledge base of all known medications
  • domains — Contains annotated example queries for all intents, sorted by domain.
  • entities — Contains training data for entity resolution.
  • scripts — custom Python scripts used to generate data in the three folders described previously.
    • rxnorm.py — Collects data from RxNorm API
    • kb_generatory.py — Generates a JSON knowledge base given a list of ingredients using information from RxNorm.
    • entity_generator.py — Generates entity mapping files given a JSON knowledge base.
    • frequencies_gen.py — Generates natural language phrases for describing frequencies.

Current Status & Limitations

• This application's knowledge base is limited to the 50 most perscribed medications in the US (Fuentes, 2018).
• Representative training data is currently being generated to train the NLP classifiers. (related Mindmeld documentation)

Future Work

• Include a knowledge base component that directly interfaces with the RxNorm API instead of a locally-stored database.