Bridging Clinical Guidelines with Real-time Patient Data: An AI Knowledge Graph Framework for Anesthesia and Intensive Care

In intensive care units and perioperative settings, anesthesiologists must make fast, high-stakes decisions based on multiple sources of patient data. AI-based support systems have begun to emerge that recommend treatment actions based on patient vitals, clinical notes, laboratory results, and/or demographic information. Yet despite being able to predict patient outcomes with reasonable accuracy, they do not account for the fact that physicians rely on established medical guidelines and their own experience observing similar patient responses to make decisions. Acting largely as black boxes, traditional AI models do not convey the reasoning behind their predictions and, subsequently, fail to capture the trust of healthcare practitioners. We take a different approach by developing a knowledge graph-based system that explicitly links clinical guidelines with multi-modal patient information, presenting attending physicians with relevant guidelines matched to a patient’s condition in real-time. To achieve this, we first construct a knowledge graph from textual guidelines that connects entities in the form of patient characteristics and treatment options; and then train specialized encoders that map patient data to that same entity space. We further target another source of information that clinicians rely on, namely their own experience attending similar cases. We distill this complex data by learning a second representational space that allows the retrieval of similar cases via contrastive learning, defining our positive pairs (that are brought close together in the embedding space) based on the guidelines that the knowledge graph deems applicable. Our system is designed to learn in a largely unsupervised manner, minimizing the need for time-intensive manual annotations, and can be improved over time with user feedback via reinforcement learning. To ensure real-world applicability, we will collaborate closely with domain experts in designing the system’s interface and interoperability requirements. Prospective testing at the University Hospital Vienna will allow us to evaluate the impact of our proposed approach on clinical efficiency and adherence to best practices. This is the first system to integrate structured guideline recommendations with multi-modal patient data, combining knowledge graph reasoning with recent advances in self-supervised learning and integrating adaptive learning principles to promote sustainability and ease-of-use.