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

Technical Summary: VISTA Architect

Problem Statement
Clinicians, particularly in multidisciplinary tumor boards, face significant challenges reconstructing coherent patient narratives from fragmented electronic health records (EHRs). EHR data is organized for documentation rather than clinical reasoning, scattering patient trajectories across isolated notes, labs, and imaging reports. Manually integrating these data points is time-consuming (up to 4 hours daily per clinician) and prone to cognitive errors.

Existing AI approaches struggle with this scale. While Large Language Models (LLMs) excel at summarizing single documents, they degrade in performance when applied to longitudinal records spanning years and thousands of events. Direct long-context prompting often fails to capture temporal relationships and incurs high latency and cost. Retrieval-Augmented Generation (RAG) systems retrieve text fragments based on semantic similarity rather than clinical or temporal structure, often missing critical longitudinal connections. Furthermore, dense retrieval models frequently underperform simple lexical baselines in specialized domains, and even optimized RAG pipelines require substantial computation per query while achieving suboptimal accuracy compared to structured approaches.

Methodology: VISTA Architect
VISTA Architect is a database-oriented AI system designed to integrate LLMs with longitudinal EHRs by transforming raw documentation into a persistent, provenance-linked knowledge graph. The architecture avoids repeated raw-text processing at query time by precomputing a structured representation of the patient state. It consists of three sequential layers:

1. Source-Faithful MEDS Graph: The system first ingests EHR data (represented here as MEDS XML) and converts it into a hierarchical graph. This layer preserves the granular structure of the source data (patients, encounters, notes, measurements, procedures, etc.) as nodes and edges, maintaining full provenance links back to the original documents. This step is deterministic and does not require an LLM.
2. Timeline Object Architecture (TOA) Graph: Built atop the MEDS Graph, this layer uses graph-guided LLM extraction to synthesize a clinically abstracted timeline. Instead of prompting an LLM with the entire raw record, the system traverses the MEDS Graph to present compact, graph-derived segments. The LLM extracts standardized, structured clinical event objects (e.g., diagnosis, treatment start, adverse events). Crucially, the TOA Graph infers the actual time of clinical occurrence, distinguishing it from the documentation timestamp, thereby creating a coherent, deduplicated timeline of the patient's biomedical course.
3. Episode and Current-State Artifacts: The TOA event stream is further compressed into higher-order clinical structures called "episodes" (e.g., baseline history, treatment lines, post-oncological phases). This layer also generates current-state artifacts, such as structured variable snapshots and summaries. These are precomputed reasoning outputs stored directly in the graph.

Agentic Interface and Retrieval
The system exposes these layers through an agentic AI bridge. When a query is made, the agent operates over the precomputed TOA layer and current-state artifacts. If additional detail is required, the agent traverses clinically meaningful edges in the graph to retrieve targeted source evidence from the MEDS Graph or original documents. This shifts the retrieval paradigm from searching raw text to navigating a structured clinical map. The system supports both on-demand cohort preparation and conversational clinical reasoning via a shared interface.

Key Results
The system was evaluated in the context of thoracic oncology tumor boards at Stanford Medicine across a cohort of 1,180 patients.

• Accuracy: VISTA Architect achieved 96.4% accuracy (mean score 9.75/10) on 15 tumor board-salient variables (17,700 total evaluations). This included high performance on demographics (100%), tumor characteristics (e.g., Histology 97.8%), and treatment history.
• Comparison with RAG: In a matched comparison on a 30-patient subset, VISTA Architect (96.9% accuracy) significantly outperformed a standard BM25 RAG baseline (66.9% accuracy). The gap was most pronounced on variables requiring temporal resolution of multiple mentions (e.g., Previous Surgery, Metastasis, Radiation Therapy), where RAG accuracy dropped to 33–53%, while VISTA maintained >96%.
• Efficiency: The fully agentic build pipeline reduced preparation time for a 30-patient cohort to approximately 2.2 minutes (mean 88 seconds per patient), a 34-fold speedup over sequential processing. Graph traversal for queries occurred in under 100ms.
• Cross-Patient Retrieval: The system demonstrated the ability to perform cross-patient similarity searches using precomputed graph artifacts without reprocessing raw EHRs, retrieving clinically relevant matches based on biology, treatment trajectory, and narrative summaries.

Significance and Claims
The paper claims that VISTA Architect addresses the fundamental tension between the scale of longitudinal patient records and the capacity of current AI methods. Its primary contribution is not the replacement of SQL or relational databases, but the encoding of a "clinically meaningful, provenance-linked hierarchical tree-like compression" of the longitudinal record.

By moving the expensive task of longitudinal synthesis to the ingestion phase, the system allows downstream applications to operate over an organized patient state, traversing to source evidence only when necessary. This approach mitigates the temporal reasoning degradation and hallucination risks associated with direct long-context prompting and the disconnected context of standard RAG. The authors position VISTA Architect as a model for connecting large document repositories to LLM-powered applications, offering a flexible architecture adaptable to other specialties through customizable event definitions and episode structures, provided that rigorous validation is conducted in those specific domains.