OntoSPM-LLM

Surgical training and clinical practice traditionally rely on experience, textbooks, and case studies. Current AI-based systems still exhibit significant limitations for medical applications: insufficient reliability and explainability can lead to inaccurate or misleading conclusions. Moreover, most AI models do not leverage medical ontologies, resulting in limited semantic grounding. Consequently, there is a need for an AI system capable of generating medically validated and interpretable responses.

The project OntoSPM-LLM aims to bridge this gap by combining Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) and external knowledge sources to establish a semantically grounded knowledge base for surgery. 

In the long term, such knowledge could also serve as a foundation for robotic assistance in the operating room. Leveraging ontologies ensures that system responses remain consistent, explainable, and evidence-based. As a first step, the project develops a natural language dialogue system for medical domains, enabling surgeons to discuss clinical questions with a competent AI system, deepen domain knowledge, and, ideally, support more informed decision-making.

The project addresses the following research questions:

To what extent can medical ontologies be integrated into LLMs through RAG such that structured ontological information becomes available for precise query answering? How can symbolic and data-driven methods be combined to form a semantically robust knowledge representation that correctly interprets surgical terminology, procedures, and workflows?

Building on this, how can a knowledge-based dialogue interface be designed that supports both general surgical knowledge dissemination and case-specific queries involving annotated clinical data?

Finally, how can such a system be rigorously evaluated in collaboration with surgical experts, and to what degree does an ontology-assisted LLM outperform existing AI-based solutions in terms of error rates, answer quality, and user satisfaction in educational as well as clinical scenarios?

The scientific approach is based on integrating medical ontologies with LLMs using Retrieval-Augmented Generation. Ontological models are extended and formalized for machine access, allowing structured domain knowledge to be embedded in the RAG pipeline. The LLM component provides dialogic interaction and semantic processing of user queries. The methodology requires continuous interdisciplinary research between computer scientists and surgical experts for the curation of terminologies, workflows, and case studies, as well as for validating the resulting knowledge representations.

In this project, we will develeop new methods fort he integration of ontologies and LLMs to gain novel insights towards the combination of symbolic and data-driven AI. These will result in the development of a functional prototype of a knowledge-based dialogue interface as a demonstrator. This prototype will showcase the feasibility of RAG-based, ontology-assisted conversational AI for surgical knowledge dissemination and case-based interaction. We will also asses possibilities to use this technology in other domains, e.g. manufacturing, in future projects.