Timeline Object Architecture: Multidomain Insights

• Timeline Object Architecture (TOA) is an approach that converts heterogeneous temporal data into standardized, auditable timeline objects across domains.
• In clinical applications, TOA replaces documentation nodes with event-centric abstractions, improving temporal fidelity and deduplication in patient records.
• In vision-language systems, TOA defines explicit temporal mappings over embedding spaces, enabling interpretable ordering and effective time analysis.

Searching arXiv for the cited papers to ground the article in the current literature. arxiv_search(query="(Kiiskinen et al., 21 Jun 2026)", max_results=5) Timeline Object Architecture (TOA) is an architectural pattern for making temporal structure explicit by converting heterogeneous source material into standardized timeline objects that can be ordered, aggregated, queried, and, in some instantiations, traced back to source evidence. The available arXiv literature uses the term in at least two technical senses. In health AI, TOA is the clinical abstraction layer in VISTA Architect, where it converts longitudinal electronic health record documentation into a deduplicated, temporally coherent, provenance-linked timeline of patient care. In vision-language modeling, a TOA design grounded in TIME10k defines an explicit timeline representation over embedding spaces, turning latent temporal information into a scalar chronological mapping and associated ordering utilities (Kiiskinen et al., 21 Jun 2026, Tekaya et al., 22 Oct 2025).

1. Clinical TOA as an event-centric abstraction layer

In VISTA Architect, TOA exists to address a fundamental mismatch between how EHRs are written and how clinicians reason. EHRs are documentation-centric, redundant, and timestamped by when something was recorded; clinical reasoning is event-centric, phase-structured, and organized around when something occurred. TOA therefore replaces documentation-order nodes with clinical-occurrence events and episode structures, yielding an organized patient state that downstream agents can traverse deterministically. Every TOA event remains linked by provenance edges to the source-faithful MEDS Graph, so synthesized facts remain auditable rather than detached summaries (Kiiskinen et al., 21 Jun 2026).

The stated goals of TOA are fourfold: reduce longitudinal redundancy by merging repeated mentions across notes and reports; restore temporal fidelity by inferring clinical occurrence dates; provide clinically meaningful aggregation via episodes such as baseline, diagnosis, treatment lines, and post-oncological phases; and anchor every assertion to source evidence so that verification remains possible. In this formulation, TOA is not merely a summarization layer. It is a graph of standardized timeline objects representing clinical occurrences rather than documentation mentions, with explicit temporal normalization through the distinction between occurrence date and evidence date (Kiiskinen et al., 21 Jun 2026).

A common misconception is to treat TOA as interchangeable with retrieval-augmented generation or long-context prompting. The VISTA formulation explicitly positions TOA against both. Direct long-context prompting and classical RAG repeatedly process raw records at query time, often miss temporal relationships, and have difficulty maintaining consistent event ordering or deduplication. TOA instead precomputes clinical synthesis once at ingestion and exposes a persistent temporal state for later traversal (Kiiskinen et al., 21 Jun 2026).

2. Formal object model, episodes, and graph semantics

The clinical TOA in VISTA Architect sits above a source-faithful MEDS Graph. The MEDS Graph preserves granular EHR structure with provenance, including patients, encounters, notes, measurements, images, and procedures; in the thoracic cohort, its per-patient median node count was 3,608 with interquartile range 562–11,629. TOA is the abstraction layer above that source plane, storing standardized event nodes and episodic nodes with edges encoding temporal order, containment, anchoring, and provenance (Kiiskinen et al., 21 Jun 2026).

Formally, an event is modeled as

$$e = (\mathrm{type}, t_{\mathrm{start}}, t_{\mathrm{end}}, a, P, p),$$

where $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$, $t_{\mathrm{start}}$ and $t_{\mathrm{end}}$ are occurrence-time bounds, $a$ is a structured attribute map, $P$ is a set of provenance pointers into the MEDS Graph, and $p$ is an optional confidence or quality score used during synthesis. TOA stores both date and evidence_date per event to preserve temporal normalization (Kiiskinen et al., 21 Jun 2026).

Element	Key fields or edges	Function
Event node	event_id, date, evidence_date, source_event_refs, type, subtype, modality, site, laterality, description, priority, values	Standardized clinical occurrence
Episode node	episode_id, kind, start_date, end_date, anchor_event_id, event_ids, and for treatment lines line_number, treatment, intent, clinical_context, termination_reason	Phase-structured temporal aggregation
Core edges	PRECEDES, CONTAINS, ANCHORED_BY, FOLLOWED_BY, SAME_DAY, SOURCED_FROM	Ordering, membership, anchoring, co-occurrence, provenance

The episode model is central to TOA’s phase structure. Episode kinds are baseline, diagnosis, treatment_line, and post_oncological. Treatment-line episodes additionally encode line number, regimen string, intent, clinical context, and termination reason. Edge constraints are equally explicit: episodes are time-monotonic; an event that terminates a treatment line belongs to the terminating episode; and FOLLOWED_BY yields an acyclic sequence. This graph semantics turns a timeline into a navigable object space rather than a flat list of dated facts (Kiiskinen et al., 21 Jun 2026).

3. Extraction, deduplication, and temporal normalization

At ingestion, VISTA Architect traverses the MEDS Graph, serializes longitudinal content in temporal order, partitions it into encounter-aligned chunks of approximately 120k characters by default, and uses LLMs to extract structured events. The prompts require the model to distinguish occurrence date from documentation date, attach provenance through source_event_refs, separate baseline_information, critical_information, and adverse_effects, extract imaging separately per modality, and include laboratory events only when abnormal or decision-informing. Production used GPT-4.1 for chunk extraction, while the agentic variant used Gemini 3.5 Flash; episode synthesis later segments unified events into baseline, diagnosis, numbered treatment lines, and post-oncological phases (Kiiskinen et al., 21 Jun 2026).

The ingestion pipeline is explicitly staged. Source serialization and chunk planning are followed by per-chunk event extraction to JSON, a unifier for deduplication, a radiology alignment safety net, episode synthesis, and generation of patient_info.json, episodes.json, and summary.json. These artifacts are stored rather than recomputed per query. Traversal begins from TOA when answering questions and descends to MEDS only when provenance-led verification is required (Kiiskinen et al., 21 Jun 2026).

Deduplication is formalized through a similarity function that combines type equality, attribute overlap, date proximity, and provenance intersection:

$$\mathrm{sim}(e_i,e_j)=w_p \cdot \mathbf{1}[|P_i \cap P_j|>0] + w_t \cdot \mathbf{1}[\mathrm{type}_i=\mathrm{type}_j] + w_a \cdot J(a_i,a_j) + w_d \cdot \kappa(|t_i-t_j|).$$

The deduplication decision is

$$D(e_i,e_j)=\mathbf{1}[\mathrm{sim}(e_i,e_j)\ge \tau].$$

In practice, the primary merge is a union-find keyed by provenance, with disjoint sets initialized over events and any pair with shared provenance unioned. Canonical events are then created by attribute-wise reduction. A secondary merge handles events without shared provenance using type, attributes, and date proximity; residual ambiguous cases are resolved by a sub-agent. Temporal normalization favors explicit time-of-occurrence stated in imaging or report headers and signed pathology report dates, falling back to nearest documentation if ambiguous. Date ranges are treated as intervals, and month-only dates are normalized to the first day of the month with a flag. Temporal ordering further defines a partial order over events, with same-date ties broken by clinical causality rules or represented through SAME_DAY edges. In the thoracic profile, broader episode grouping used $\pm 90$ days around imaging anchors and co-occurrence merging used $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$0 days for imaging-related events (Kiiskinen et al., 21 Jun 2026).

4. Query plane, amortized computation, and thoracic-oncology evaluation

A defining property of the clinical TOA is that it is precomputed once at ingestion. Queries then operate over the TOA data plane and descend into MEDS only for targeted verification. The paper expresses the cost contrast as

$$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$1

and

$$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$2

with a break-even point $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$3 satisfying

$$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$4

Measured values in the thoracic cohort were a mean ingestion time of 88 seconds per patient for the fully agentic build, deterministic TOA traversal typically under 10 ms for graph execution, and end-to-end chat latency of 1–3 seconds dominated by generation rather than retrieval. The BM25 RAG baseline required fresh retrieval over raw EHR XML and generation for each variable, with approximately 67% accuracy and longer per-query latencies (Kiiskinen et al., 21 Jun 2026).

The query model is deterministic and graph-oriented. Downstream agents and dashboards read patient_info.json, episodes.json, and TOA events directly. Typical workflows include retrieving the latest tumor stage from patient_info, traversing episodes to the latest diagnosis episode and following PRECEDES edges to the most recent staging assessment event, listing treatment-line episodes by line number with treatment intent and termination reason, or scanning imaging events with modality in {CT, PET-CT} and returning the maximum occurrence date. Source verification is accomplished by following SOURCED_FROM back to the exact radiology, pathology, or XML fragment in MEDS (Kiiskinen et al., 21 Jun 2026).

The demonstration setting was multidisciplinary thoracic oncology tumor boards at Stanford Medicine, where precise reconstruction of longitudinal lung cancer histories is critical. Across 1,180 patients, 15 tumor-board–salient variables, and 17,700 total evaluations, VISTA Architect achieved 96.4% accuracy, corresponding to 17,063 correct answers and a mean score of 9.75/10, with 95% confidence interval 96.1–96.7%. Challenging variable performance included 92.1% for Date of Last CT, 99.1% for radiation therapy, and 97.8% for histology. On a 30-patient held-out cohort, the agentic build matched accuracy while reducing preparation to approximately 2.2 minutes total through full intra- and inter-patient parallelization. In a matched 30-patient comparison to BM25 RAG, VISTA achieved 96.9% correct versus 66.7–66.9% for RAG, with per-patient superiority significant by exact sign test $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$5 and Wilcoxon signed-rank one-sided $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$6–$$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$7. The paper reports that RAG dropped particularly on time-varying variables requiring temporal resolution, including Previous Surgery, Metastasis, and Radiation Therapy (Kiiskinen et al., 21 Jun 2026).

5. TOA in vision-LLMs

A distinct use of the term appears in work on temporal awareness in vision-LLMs. “A Matter of Time: Revealing the Structure of Time in Vision-LLMs” introduces TIME10k, a benchmark of 10,091 images curated from Wikipedia and Wikimedia Commons with year-level temporal ground truth across six categories: Cars, Mobile Phones, Ships, Instruments, Aircraft, and Weapons and Ammunition. Evaluating 37 VLMs, the study reports that temporal information is structured along a low-dimensional, non-linear manifold in the embedding space. Kernel PCA projections of year-specific time embeddings form a smooth chronologically ordered curve, and image embeddings projected into the same subspace co-locate with that curve. Dimensionality analysis indicates that around 13 dimensions suffice to capture most temporal signal for both CLIP and EVA-CLIP (Tekaya et al., 22 Oct 2025).

Grounded in those methods and findings, the technical design of a TOA for VLMs defines a Timeline Object $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$8 with an embedding function $$\mathrm{type} \in \{\mathrm{imaging}, \mathrm{diagnostic}, \mathrm{treatment}, \mathrm{surgery}, \mathrm{lab}, \mathrm{adverse\_effect}, \mathrm{baseline\_information}, \mathrm{critical\_information}, \ldots\}$$9, a timeline mapping $t_{\mathrm{start}}$0 from embedding space to scalar time, calibration anchors $t_{\mathrm{start}}$1, an uncertainty estimator $t_{\mathrm{start}}$2, and optional manifold assets $t_{\mathrm{start}}$3 such as a projection map, k-nearest-neighbor graph, principal curve, or Bézier curve. The paper’s temporal probing baseline predicts

$t_{\mathrm{start}}$4

while two explicit timeline extraction methods are emphasized: a one-dimensional UMAP timeline optimized to maximize Spearman’s $t_{\mathrm{start}}$5 between projected order and chronological order, and a Bézier-curve timeline in either the original embedding space or a projected subspace. In the Bézier formulation, time embeddings are projected onto a fitted curve by

$t_{\mathrm{start}}$6

and interpolative year assignment is

$t_{\mathrm{start}}$7

Empirically, the paper reports near-perfect ranking alignment for some timeline constructions, with $t_{\mathrm{start}}$8, $t_{\mathrm{start}}$9, and $t_{\mathrm{end}}$0, while Bézier timelines achieve near-probing inference speed and competitive or superior Time-Adaptive Accuracy relative to prompt-based baselines. This suggests that, outside health AI, TOA can also denote an explicit temporal representation layer built over learned multimodal manifolds rather than over provenance-linked clinical graphs (Tekaya et al., 22 Oct 2025).

6. Generalization, related lineages, and unresolved questions

The clinical TOA is modular rather than intrinsically thoracic-oncology-specific. The architecture is described as adaptable to other specialties through customizable event vocabularies, episode structures, ontologies, and tool surfaces. Examples given include echocardiography and catheterization for cardiology, seizure events and brain MRI for neurology, acute admission phases, rehabilitation episodes, chronic management cycles, and specialty-specific deterministic retrievers. The stated requirements are a source representation convertible to nodes, edges, and timestamps; prompt profiles for event extraction and episode synthesis; domain-specific grouping windows and merging heuristics; and an agentic interface backed by a PHI-compliant LLM serving stack. At the same time, validation beyond thoracic oncology remains future work, and the architecture’s specialty-agnostic design is presented as requiring empirical confirmation per domain (Kiiskinen et al., 21 Jun 2026).

Several limitations are explicitly identified. Edge cases include conflicting events such as multiple external reports with inconsistent dates, missing timestamps with month-only documentation, and sparse provenance for remote care. External notes and outside studies may be missing in OMOP exports, affecting variables such as Date of Last CT and some molecular results. Deferred validation includes ablations isolating TOA layers, explicit temporal-normalization evaluation against chart-reviewed event dates, expanded clinician adjudication beyond the 30-patient subset, and stronger RAG baselines including dense retrieval and long-context models. Prospective multi-site studies, multimodal integration with DICOM, WSI, and genomics, and cohesive cross-patient agent surfaces are planned to assess workflow impact, user satisfaction, and decision quality at scale (Kiiskinen et al., 21 Jun 2026).

A second misconception is that TOA necessarily sacrifices traceability for compactness. The VISTA formulation argues the opposite: synthesized events remain auditable because every event is linked through SOURCED_FROM to MEDS XML fragments, and verification is designed as downward traversal from abstraction to source. A third misconception is that there is a single canonical TOA schema. The current literature instead shows domain-specific instantiations: a provenance-linked graph of events and episodes in longitudinal clinical AI, and a manifold-aware temporal object over embedding spaces in vision-language modeling. A plausible implication is that TOA is best understood as a family of architectures that externalize time into reusable objects, with the exact object semantics determined by the source domain and intended inference tasks.

A related but distinct lineage appears in distributed systems research on secure timelines. “Secure History Preservation Through Timeline Entanglement” defines a secure timeline as a tamper-evident historic record of the states through which a system goes throughout its operational history, and extends secure timelines across multiple, mutually distrustful services through timeline entanglement, which links undeniably the past of one timeline to the future of another. The paper presents Timeweave as a service development framework for timeline entanglement and reports efficient deployment in a loosely-coupled distributed system of a few hundred services with overhead of roughly 2–8% of the processing resources of a PC-grade system. Although this work does not present a TOA in the clinical or VLM sense, it foregrounds the same broader concerns with explicit temporal structure, ordering guarantees, and persistent temporal proofs [0202005].