VISTA Architect: Graph-Based Health AI
- VISTA Architect is a graph-based health AI system that converts raw clinical documentation into a persistent, provenance-linked graph to support efficient longitudinal synthesis.
- The architecture employs a two-layer graph model—a source-faithful MEDS Graph and a clinically abstracted Timeline Object Architecture—ensuring deterministic, verifiable patient state reconstruction.
- By precomputing clinical narratives and enabling rapid, traceable queries, VISTA Architect outperforms traditional retrieval-augmented methods with high accuracy and auditability in thoracic oncology.
Searching arXiv for the specified paper and closely related context so the article can be grounded in current literature. VISTA Architect is a database-oriented health AI architecture for integrating LLMs with longitudinal electronic health records, designed to transform raw clinical documentation into a persistent, provenance-linked graph at ingestion rather than repeatedly reprocessing source text at query time. In the reported thoracic oncology deployment at Stanford Medicine, the system couples a source-faithful MEDS Graph with a clinically abstracted Timeline Object Architecture (TOA), enabling deterministic traversal over an organized patient state while preserving one-click return paths to exact source fragments for verification. The architecture was demonstrated on multidisciplinary tumor boards, where rapid and temporally accurate reconstruction of diagnoses, molecular findings, treatments, responses, adverse events, and imaging is operationally central (Kiiskinen et al., 21 Jun 2026).
1. Clinical rationale and architectural premise
Multidisciplinary thoracic oncology tumor boards require rapid, accurate reconstruction of longitudinal patient trajectories across fragmented EHR sources such as notes, labs, imaging, procedures, and medications. The motivating claim of VISTA Architect is that direct application of LLMs to whole EHRs is inefficient and error-prone in this setting, because long-context prompting degrades in temporal reasoning as input length grows, and standard retrieval-augmented generation retrieves semantically similar fragments without respecting longitudinal clinical structure. Deterministic relational queries preserve provenance, but they do not by themselves reconstruct a clinically meaningful patient-level narrative without bespoke synthesis logic (Kiiskinen et al., 21 Jun 2026).
The architectural response is a precomputation paradigm. Longitudinal EHRs are transformed at ingestion into a persistent graph with provenance, then abstracted once into clinically organized timelines and reusable current-state artifacts. Downstream interactions operate over this structured state and return to source text only for targeted verification. This design addresses cost and latency by amortizing the most expensive reasoning over the patient record rather than repeating raw-text processing for each question.
A common misunderstanding is to treat VISTA Architect as a variant of ordinary RAG. The system is explicitly framed instead as a graph database-oriented architecture in which temporal coherence, deduplication, and provenance are first-class features of the data model itself. This distinction is central to its reported superiority on time-varying tumor-board variables (Kiiskinen et al., 21 Jun 2026).
2. Two-layer graph model
The architecture consists of two graph layers with different epistemic roles. The first is the MEDS Graph, which preserves source structure and provenance. The second is the TOA, which represents clinically abstracted events and episodes aligned to occurrence dates rather than documentation timestamps.
| Layer | Purpose | Representative contents |
|---|---|---|
| MEDS Graph | Make the original EHR computationally addressable with full provenance | Person, Visit, Event, XMLFragment; VISIT_CONTAINS, IMAGE_REPORT_FOR, HAS_MEASUREMENT, PRECEDES |
| TOA | Represent the patient trajectory as deduplicated, temporally coherent events and episodes | TOA events, episodes, PRECEDES, CONTAINS, ANCHORED_BY, FOLLOWED_BY, SAME_DAY, SOURCED_FROM |
The MEDS Graph is source-faithful. In the thoracic oncology instantiation, its nodes include Person, Visit, Event, and XMLFragment. Event covers note, measurement/labs, imaging, procedures, and medications. XMLFragment nodes provide fine-grained source anchors, and provenance linkage preserves identifiers and references to the original document or measurement. Timestamps in this layer reflect documentation times such as note write time or lab post time. The reported median scale is 3,608 nodes per patient, with interquartile range 562–11,629 (Kiiskinen et al., 21 Jun 2026).
The TOA is clinically abstracted. Each event includes fields such as event_id, date, evidence_date, type/subtype, modality, site, laterality, compact description, priority, values, and source_event_refs. Episodes are organized as baseline, diagnosis, treatment_line, and post_oncological, with treatment lines additionally encoding line_number, regimen summary, intent, termination_reason, and clinical_context. This structure makes transitions such as progression or toxicity explicit at the episode level.
The distinction between date and evidence_date is fundamental. TOA dates correspond to clinical occurrence, whereas evidence_date records first documentation. This allows the architecture to align patient history to the chronology of care rather than to the order in which facts happened to be written into the record.
3. Graph-guided extraction, unification, and episode synthesis
TOA construction is performed by graph-guided LLM extraction over the MEDS Graph. Context is serialized in temporal order and partitioned into encounter-aligned chunks. For each chunk, the model returns JSON events with clinical occurrence dates and explicit source_event_refs, so every synthesized assertion remains traceable to MEDS nodes or XMLFragment anchors (Kiiskinen et al., 21 Jun 2026).
Unification is then performed over candidate events. A union-find merges mentions that share source_event_refs, and residual ambiguous merges are resolved by a sub-agent. Baseline_information facts are consolidated. After merging, fields are normalized by preferring clinical occurrence dates, consolidating descriptions, and preserving all provenance links. A radiology alignment safety net cross-checks TOA imaging against structured imaging dates and verifies missing anchors through a per-date sub-agent; PET-CT counts toward “Date of Last CT” by explicit rule.
The paper’s examples illustrate the clinical semantics of this synthesis. Multiple pathology reports can be merged into a single diagnostic event, such as adenocarcinoma with EGFR L858R, whose source_event_refs point simultaneously to a biopsy note, pathology PDF fragment, and molecular report, while the event date reflects biopsy occurrence rather than documentation order. Imaging events mentioned both in progress notes and radiology PDFs are deduplicated against structured imaging tables. These procedures are intended to minimize fragmentation and temporal inconsistency in longitudinal summaries (Kiiskinen et al., 21 Jun 2026).
Episode synthesis is performed by an LLM with domain rules. Baseline is anchored pre-first oncologic event; diagnosis includes the full workup; treatment_line includes in-line imaging, labs, and symptoms; termination events remain in the ending line; post_oncological captures hospice or comfort care. This segmentation imposes a clinically legible structure on the event stream and yields a state representation closer to tumor-board reasoning than raw document retrieval.
4. Ingestion pipeline, current-state artifacts, and query execution
The ingestion pipeline begins from OMOP-derived MEDS XML exports covering notes, labs and panels, imaging studies and reports, pathology and molecular testing, procedures, medications, and encounters. MEDS XML preserves raw structure; measurement panels are decomposed to individual lab values; radiology reports are linked to images; visits group events. Domain configuration defines event and episode vocabularies and variable dictionaries, which in the thoracic oncology demonstration include diagnosis, histology, metastasis/lymph nodes, molecular panel variables, ECOG, therapies, radiation, last CT date, allergies, and smoking (Kiiskinen et al., 21 Jun 2026).
From TOA events and episodes, the system emits current-state artifacts such as patient_info.json and summary.json. Each field is linked back to provenance through event-to-XMLFragment relations. This creates what the paper describes as an organized patient state comprising deduplicated TOA events, episodes, and display-ready fields. Query-time operation is then primarily deterministic traversal over this graph-resident state.
The reported retrieval characteristics are unusually asymmetric between retrieval and generation. Deterministic graph traversal using NetworkX typically takes 0.1–1.5 ms per operation, with provenance lookups under 10 ms end-to-end. Chat latency is reported as 1–3 s total, where most time is spent in planning and generation rather than in graph execution. This suggests that the core data-plane bottleneck has been shifted away from raw-record retrieval toward higher-level response construction (Kiiskinen et al., 21 Jun 2026).
The implementation uses Python, serializes graphs to GraphML, and in the research implementation keeps graphs in memory via NetworkX. The schema is described as compatible with Neo4j for population-scale deployment, including illustrative Cypher patterns for retrieving timeline events, episodes with anchors and contained events, and provenance for a specific event. This suggests a deployment path from research-scale in-memory analysis to graph-database-backed cohort operations, though the paper does not report population-scale Neo4j benchmarks.
5. Agentic interface and tumor-board workflow
VISTA Architect exposes an agentic interface whose tools are deterministic graph retrievers rather than unrestricted free-text search. Reported tools include imaging-date vectors, modality lists, ECOG and smoking status retrievers, driver mutation keyword search, treatment and drug exposure history, allergies and conditions, and code-status mentions, all tracked to MEDS provenance. A top-level agent plans graph traversals, invokes parallel sub-agents for chunk extraction, unification, episode synthesis, and display generation, and validates outputs against the graph before merging (Kiiskinen et al., 21 Jun 2026).
In the described tumor-board workflow, clinicians interact with TOA-derived views: an Overview, a Clinical Timeline organized into episodes with anchors and contributing events, an automated pre-tumor-board summary note, and a chat interface for provenance-backed questions. Typical interactions include inspecting the latest therapy and imaging, asking for evidence behind a metastasis claim, and generating a final board note.
Automation is presented as workflow compression rather than autonomous decision-making. In a held-out 30-patient cohort, the fully agentic build completed in about 2.2 minutes total, with mean 88 s per patient and maximum 131 s for ingestion, extraction, unification, episode synthesis, and display generation. Automated tasks included variable extraction across 15 tumor board–salient fields, timeline assembly, episode formation, pre-board summary note generation, and provenance linking (Kiiskinen et al., 21 Jun 2026).
A plausible implication is that VISTA Architect is less a conversational assistant than a structured clinical state compiler with a conversational surface. The architecture’s emphasis is on durable representation, deterministic traversal, and evidence-linked synthesis rather than on unconstrained natural-language reasoning over raw records.
6. Evaluation, provenance, and scope of generalization
The evaluation cohort comprised 1,180 thoracic oncology tumor-board patients from January 2020 to November 2025, de-identified, with 15 tumor board–salient variables per patient, totaling 17,700 evaluations. Reported accuracy was 96.4% (17,063/17,700), with mean quality 9.75/10 and 95% CI 96.1–96.7%. The most challenging field was Date of Last CT at 92.1%, attributed to internal versus external imaging ambiguity (Kiiskinen et al., 21 Jun 2026).
The paper states the accuracy formula as
and gives the binomial proportion confidence-interval form
while noting that Wilson intervals were reported in the paper.
On a matched 30-patient subset using the same LLMs and ground truth, a BM25 RAG baseline achieved 66.7–66.9% accuracy versus VISTA’s 96.9%, with VISTA outperforming RAG on every patient and showing the largest gaps on Previous Surgery, Metastasis, and Radiation Therapy. The paper further positions this performance against recent LLM extraction benchmarks, arguing that whole-record longitudinal synthesis benefits from encoding temporal structure and provenance at ingestion rather than at retrieval time (Kiiskinen et al., 21 Jun 2026).
Auditability is a defining feature. Every TOA event carries source_event_refs back to MEDS XMLFragment nodes, and SOURCED_FROM edges preserve audit trails across abstraction boundaries. The interface exposes “view source” traversal to raw EHR fragments, and chat answers cite linked evidence. Conflict resolution combines union-find deduplication, radiology alignment, and explicit episode rules such as keeping progression events within the terminating episode.
The principal limitations are explicit. Validation is currently confined to thoracic oncology at one institution with English-language records. Accuracy depends on source EHR completeness, and outside-institution records may be missing. Future work is identified as expanded validation, ablations of TOA components, explicit temporal normalization validation, denser retrieval baselines, multimodal integration including imaging and genomics, and prospective clinical workflow studies. The paper therefore supports modular adaptability across specialties through configurable event definitions, episode vocabularies, edge semantics, variable dictionaries, and agent tools, but it does not claim empirical generalization beyond the reported thoracic oncology setting (Kiiskinen et al., 21 Jun 2026).