ACTRIMS 2026 | Agentic AI and the MedCP framework in MS research and beyond

Well, everyone of us is pretty well aware of the boom of AI technologies right now, from everyday use to research applications. So, agentic AI is a recent advancement in the AI field, and it’s different from previous attempts because it has autonomy. So agentic AI can plan a multi-step analysis executed by interacting with other sources like databases and tools, evaluate intermediate results and adapt a strategy along the way. So it’s like for us clinicians, it’s like the difference between asking a colleague a quick question or asking them to investigate something and to define a complete analysis and coming back with a more thorough result. So the data we presented at ACTRIMS are centered on MedCP, which is one declaration of agenetic AI. MedCP stands for Medical Modal Context Protocol, and it’s an infrastructure. It’s a framework that sits between the researcher and two clinical resources. So one is an electronic health record database. In our case, it is the University of California, San Francisco, EHR. And on the other side, there is SPOKE, which is a database of databases, meaning biological knowledge graph, which collects information on over 50 million curated concepts from 45 databases, meaning genes, diseases, proteins, pathways, drugs, and all the relationships between them. So the researcher can ask a question in natural language, in plain English, as we have now been used to do with common GPT-style AI tools. And the system allows to perform queries, to perform analysis, and to return complete results, including the processes and the steps that it has taken in order to get to the results. So it’s also transparent and reproducible. So we tested this system with 100 biomedical questions of different complexities, starting from simple lookups to actually emulation of clinical trials in real-world data. And we saw that advancing the knowledge that the LLM has through SPOKE, so through curated biomedical knowledge, allows us to inform more detailed results, diminish the hallucinations to improve the performance of this tool. And this is very crucial because as we know, AI can be prone to hallucination and reaching the parametric memory of the system with external sources of memory actually seems to be a useful way to address this issue. So, yeah, so these are the core advancements of our framework, which is, again, one of many frameworks that we will most likely meet in the next few months. So trying to advance and accelerate, democratize the access and analysis to big data for researchers through AI systems.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.