VISTA Architect preps oncology tumor boards in 2.2 minutes

A new paper on arXiv introduces VISTA Architect, a system that prepares thoracic oncology tumor board summaries from longitudinal electronic health records in about 2.2 minutes for a 30-patient held-out cohort. Across 1,180 patients, the authors report 96.4% accuracy, a mean of 9.75 out of 10 (95% CI 96.1-96.7%) on 15 tumor board-salient variables over 17,700 evaluations, surpassing what they call a matched BM25 RAG baseline.

The architectural bet is what makes this worth reading rather than just another LLM-on-EHR pitch. Instead of running retrieval-augmented generation at query time, VISTA does the synthesis once at ingestion and stores it in what the authors describe as a persistent, provenance-linked knowledge graph. There are two layers. A source-faithful MEDS Graph preserves the granular EHR structure with full provenance. A Timeline Object Architecture sits above it and uses graph-guided LLM extraction to produce a concise timeline of deduplicated, temporally coherent clinical events. Downstream queries read from an organized patient state and traverse to source documentation only when detailed verification is needed.

For clinicians, the implication is the prep time. A thoracic oncology multidisciplinary board has a familiar bottleneck: someone spends hours per case stitching together pathology, imaging, lines of therapy, and recent encounters. Compressing that to minutes is a real productivity claim, and the provenance link is what lets a clinician trace any synthesized fact back to its source note rather than trusting an opaque summary.

The honest caveat is one the paper itself flags. The system was demonstrated at a single institution, Stanford Medicine, configured for thoracic oncology, and the authors write plainly that validation beyond thoracic oncology remains future work. The 15 variables were chosen by the team, and what the abstract does not give you is the variable list itself, the ingestion cost per patient, or how the precomputed graph stays current as new notes arrive.

If the approach holds up beyond this demo, the people who benefit first are academic cancer centers drowning in chart review and trial-matching teams that want to surface eligible patients across longitudinal records at speed. The authors say the modular design adapts to other specialties through customizable event definitions, episode structures, and agentic tools, so the direction worth watching is whether the precompute-into-a-graph pattern gets picked up beyond oncology.