[Development of a multidisciplinary intelligent nexus for decision support in digestive malignancies and its...

We describe the technical architecture of the Huashan Multidisciplinary Intelligent Nexus for Decision Support in Digestive Malignancies (HS-MIND), developed using a reasoning-augmented large language model (ReAG-LLM) and a multi-agent system (MAS). We also summarize our approach to implementing the system at a single center and our experience using it in the perioperative management of older patients with digestive system malignancies. Built on a general-purpose medical AI platform, HS-MIND was customized for the management of digestive system malignancies and deployed on premises. Its core architecture comprises four layers: the data and knowledge layer, the reasoning augmentation and tool-use layer, the multi-agent collaboration layer, and the clinical application layer, with safety governance integrated throughout the workflow. By combining retrieval-augmented generation (RAG), controlled chain-of-thought (CoT) reasoning, tool-augmented reasoning, and multi-agent collaboration, the system grounds its outputs in evidence, integrates information across specialties, ranks alternative management pathways, and provides prompts for clinician review. In representative high-risk scenarios, HS-MIND can proactively identify perioperative risk factors, integrate information across specialties, and generate structured, evidence-grounded recommendations. HS-MIND has been used to synthesize clinical information and review risks before multidisciplinary discussions in 25 older patients with digestive system malignancies and multiple comorbidities. A module designed to detect conflicts involving cardiovascular risk or antithrombotic therapy was triggered in 5 cases. In 23 cases, the system's ranking of alternative treatment pathways was reviewed during multidisciplinary discussions and used to inform decision-making. Two cases were flagged for further review because of insufficient evidence or unresolved conflicts, with final adjudication by senior physicians. Using de-identified data from a previously treated 76-year-old patient with low rectal cancer and severe coronary artery disease, we illustrate the system's longitudinal reasoning while preserving the actual chronology of clinical events, thereby showing how it operates in a high-risk scenario involving competing clinical priorities. We also characterize the system's early clinical use across this single-center series of 25 cases, with the aim of informing clinical decision-making.