By Julian Strietzel, Nils Rehtanz, Tim Baum, Marc Sinner
The rasa chatbot of Patient Report Insights Summary Module (PRISM) is a conversational AI chatbot that provides insights into patient health data. The chatbot is designed to help healthcare professionals quickly and easily access important information about their patients' health. PRISM uses natural language processing (NLP) to understand user queries and provide relevant information in a conversational format.
To run the chatbot have a look at the Technical guide to run the project.
Interested about the project? Have a look at the final presentation of the project.
Have a look at the in depth features of this chatbot in
the Documentation
about expected intents and answers. We layed a special focus on relevant health data insights supported by graphics
within the blood pressure domain, including trend changes, daily patterns and outliers.
Additionally, it provides several fallback handling mechanisms to ensure a smooth user experience using different LLM
techniques. Those can be triggered by out-of-scope intents or by pre-pending the user input with ask gpt or ask defog.
The PRISM chatbot is built using the Rasa framework, which provides tools for building conversational AI applications. The chatbot consists of two main components: the Rasa Core, for NLU, conversation handling, and dialogue management, and the Rasa Action Server, for custom actions and API integrations. It is used to interact with the user and provide in depth analysis of the patient data.
This project was developed as part of the Natural Language Processing (NLP) SEBA pracitcal course at the Technical University of Munich. Within this project we were challenged to develop efficient methods to provide medical information to medical professionals and evaluate the potential of a conversational AI chatbot in the healthcare domain. Our second complementary approach was to use tailored LLM or rule-based approaches to provide highly focused summaries of the patient data.
Special thanks goes to our supervisors Phillip Schneider and Juraj Vladika for their guidance and support throughout the project.
Additionally we would like to thank the team of Alma Phil especially the amazing Lutz Frick for constantly challenging us and providing the data, aswell as the opportunity to work on this project.