VISTA Architect: A graph database-oriented health AI system demonstrated in multidisciplinary tumor boards

Clinical decision-making in multidisciplinary tumor boards requires the rapid synthesis of longitudinal patient records that are often fragmented across disparate notes, lab results, and imaging reports. Traditional AI approaches, such as direct long-context prompting or standard retrieval-augmented generation (RAG), often struggle with the temporal complexity and sheer volume of these records, leading to high latency, increased costs, and potential inaccuracies. The authors sought to develop a more efficient, structured architecture to integrate large language models (LLMs) with longitudinal electronic health records (EHRs) to improve the speed and accuracy of clinical case preparation.

VISTA Architect introduces a two-tiered graph database architecture that moves away from repeated raw-text processing. First, it ingests EHR data into a source-faithful MEDS Graph, which preserves the granular structure and provenance of the original records. Second, it builds a Timeline Object Architecture (TOA) on top of this, using graph-guided LLM extraction to synthesize deduplicated, temporally coherent clinical events and episodes. By pre-computing this clinical state, the system allows downstream agentic interfaces to perform complex queries and retrieve targeted source evidence without needing to re-process the entire patient history for every request.

The researchers evaluated VISTA Architect using a cohort of 1,180 patients in thoracic oncology tumor boards at Stanford Medicine. The system achieved 96.4% accuracy across 15 tumor board-salient clinical variables, outperforming both a matched BM25 RAG baseline and existing benchmarks for LLM-based clinical extraction. Furthermore, an agentic interface enabled the preparation of patient cases in approximately 2.2 minutes, significantly reducing the manual effort required by clinicians while maintaining high precision.