Reconstructing a patient's timeline from free-text notes means deciding, for many pairs of clinical events (a biopsy, a chemotherapy cycle, a relapse), whether one came before the other, overlapped it, or contained it. The number of candidate pairs grows with the square of the number of events. Current systems avoid that growth in two ways: they only compare events that sit close together in the text, or they rely on hand-drawn temporal "containers." Both choices throw away the long-range links that matter most for a longitudinal record. Large-scale 3D reconstruction in computer vision had the same problem with photo pairs. There, an effective fix was to treat pairwise checking as active link prediction on a sparse graph. The system sends an expensive verifier only the pairs whose answer it cannot already infer, and it uses the fact that matches chain together while non-matches do not.
We propose to carry that idea into clinical temporal extraction. Temporal relations have the same asymmetry as image matches. BEFORE and CONTAINS chain transitively: if A precedes B and B precedes C, then A precedes C. OVERLAP and "no relation" do not chain. We will build an event-pair selector that holds a partial timeline graph and computes what is already implied by chaining. It then spends a fixed budget of expensive verifier calls, made by a fine-tuned clinical encoder or an evidence-grounded language model, on the pairs most likely to add a relation that cannot be inferred. It gives priority to pairs that join parts of the timeline that are not yet connected. The claim is specific and falsifiable: at a fraction of the verifier calls, this selector recovers most of the relation F1 that exhaustive pairwise checking reaches, and it recovers far more cross-sentence and cross-note relations than proximity windows or containers at equal cost.
Core Claim
The claim has three parts. First, most gold temporal relations in a clinical document can be implied from a small, well-chosen set of verified relations. Second, a cheap scorer can learn to choose that set. Third, chaining over verified relations gives both extra recall and a built-in consistency check, because an inferred relation that contradicts a verified one marks a verifier error.
The closest prior work sits on each side of the gap and does not cover it. Match Graph Construction for Large Image Databases predicts which unverified image pairs to check. It does this with a graph Laplacian over verified matches, and it lowers the scores of candidates near failed verifications. Its verification hit ratio of 0.60 on 118,000 Flickr images beat 0.49 for the Building Rome heuristics. It was never applied to text, and it has no typed relation algebra; there is only "same scene" and "not same scene." Global-Aware Edge Prioritization for Pose Graph Initialization adds global message passing, spanning-tree selection, and a score boost for pairs that bridge distant subgraphs. It shows that greedy nearest-neighbor retrieval breaks into isolated islands: it reports over 75% correct reconstructions at $k=5$ on VisymScenes against below 60% for the retrieval baselines. We reuse its connectivity boost. Its setting has no directional, typed edges, and chaining over those edges is where our method gets most of its savings.
On the clinical side, Investigating the Challenges of Temporal Relation Extraction from Clinical Text motivates the proposal. It uses narrative containers to prune the quadratic space of pairs on THYME. It reaches precision 0.983 but recall of only 0.462 on CONTAINS, which is what one would expect if pruning cuts real links. Event Time Extraction and Propagation via Graph Attention Networks propagates fuzzy time intervals over a document event graph. It assumes that graph is given, and it reports about a 4% rate of ordering-constraint violations. We address the upstream question it leaves open, namely which edges to build. We also enforce ordering constraints exactly by chaining rather than softly.
Research Program
Each document or patient record starts with its gold or predicted event and time mentions. The system keeps a typed graph with three kinds of edges: verified relations, verified "no relation," and relations implied by chaining the transitive types. We compute the implied edges with standard interval-algebra closure restricted to the transitive relations.
A lightweight pair scorer rates each unverified pair $(i,j)$. It combines cheap signals: embeddings of the two event mentions, their text distance, the note boundary, section cues, and the current graph neighborhood. We turn its output into a selection score:
$$ s(i,j) = p_{\text{rel}}(i,j)\,\bigl(1 - p_{\text{implied}}(i,j)\bigr)\,\bigl(1 + \lambda\, d(i,j)\bigr). $$
Here $p_{\text{rel}}$ is the scorer's probability that the pair holds any relation. $p_{\text{implied}}$ is 1 if closure already fixes the pair, and otherwise the scorer's estimate that it is implied by nearly certain edges. $d(i,j)$ is the normalized graph distance between the two events' components in the current graph, set to its maximum when they are disconnected. $\lambda$ is the weight of the connectivity boost taken from the pose-graph work.
Pairs go to the verifier in batches, highest score first, until the budget $B$ runs out. After each batch the system recomputes closure. Pairs next to verified "no relation" edges have their scores lowered, following the match-graph paper's use of failed verifications. We train the scorer in a self-supervised way on the verifier's own outputs from training documents, which mirrors how the pose-graph work supervises on geometric verification. Gold labels are used only for evaluation.
When closure produces a contradiction, for example A BEFORE B together with B BEFORE A, the system does not guess. It sends the lowest-confidence verified edge in the cycle back for re-verification, spending part of the budget. This is how we will test whether exact constraint handling removes the violations that soft propagation leaves behind.
We use two verifiers so that the cost argument is tested in both a cheap setting and an expensive one. The first is a GatorTron-Base cross-encoder fine-tuned on pairs. The second is a prompted language model that must quote supporting text, following the evidence-grounding step of Self-Verification Improves Few-Shot Clinical Information Extraction. The quadratic cost only really hurts in the second setting and in multi-note records. We say this openly rather than claim savings where exhaustive checking is already cheap.
Baselines
The naive reference point is the proximity window used in most pairwise temporal systems. It verifies every pair within the same sentence or adjacent sentences, plus event–document-time pairs, with no chaining. It sets the recall floor at its own verifier cost. We will add closure to it as a second reference point.
The strongest existing method on the standard benchmark is the exhaustive verifier itself. Prompting Large Language Models for Clinical Temporal Relation Extraction reports F1 of 89.54% on 2012 i2b2 with hard prompting and unfrozen GatorTron-Base, above the previous 85.70%. We will reproduce that verifier and run it on all candidate pairs. The result is our upper reference and the target the budgeted method must approach. On THYME, the strongest reported number in our file is CONTAINS F1 0.629 (precision 0.983, recall 0.462) from narrative containers in Investigating the Challenges of Temporal Relation Extraction from Clinical Text. We will re-implement that container restriction around our own verifier so that the verifier is held fixed and only pair selection varies. We found no reported result on Lacuna for budgeted or cost-accounted temporal relation extraction on either corpus. The budget curves will therefore be new measurements, not comparisons against published numbers.
The retrieval baseline is greedy top-$k$ nearest neighbors over GatorTron event embeddings. It sends each event's $k$ most similar partners to the verifier, and we tune $k$ so its budget matches ours.
Two ablations isolate our ingredients. The first removes the $(1-p_{\text{implied}})$ term, so selection is by relation likelihood alone. The second removes the connectivity term ($\lambda=0$). We run each ablation with and without contradiction-triggered re-verification.
The rules for a fair comparison are the same for every method. All methods share the same event mentions, the same verifier weights, and the same scorer training data. Each method gets the same budget of verifier calls, reported both as a fraction of all candidate pairs and as verifier tokens. Tuning uses the same development split and the same number of search trials. Beating the baselines means two things. The first is statistically higher F1 than the window, container, and top-$k$ methods at every matched budget from 5% to 25%. The second is reaching within two F1 points of the exhaustive verifier at no more than a quarter of its calls on i2b2. On THYME, it means clearly raising CONTAINS recall above the container heuristic at equal cost.
Why not just…?
Why not just use proximity windows plus closure? This is the most serious threat. Many gold relations in discharge summaries are local, and closure over local edges already recovers some long-range ones. If windowed checking plus closure matches our method at equal budget, the active selector adds nothing, and we will say so. The container result, with recall of 0.462 at precision 0.983, suggests that local pruning misses many true links. What settles the question is the gap in cross-sentence and cross-note recall at matched budgets, which we report separately from overall F1.
Why not just retrieve top-$k$ neighbors by embedding similarity? Similarity does not track whether two events are temporally related. Two mentions of "chest pain," one of them negated, embed close together, while a biopsy and the surgery it led to may not. The pose-graph work found that greedy top-$k$ produces disconnected islands and falls below 60% on ambiguous scenes, where global selection stays above 75%. Our matched-budget comparison, together with a count of connected components in each method's final graph, tests whether the same failure appears in text.
Why not just run the verifier on every pair, since GatorTron-Base is small? For one discharge summary with a fine-tuned 345M encoder, exhaustive checking is affordable, and we do not claim otherwise. The argument rests on two regimes. One is multi-note patient records, where events number in the hundreds or more. The other is evidence-grounded language-model verifiers, whose per-pair cost the self-verification paper shows is substantial. If the savings appear only there, the contribution is narrower but still real, and the cost-versus-F1 curves for the language-model verifier will show whether it holds.
Why not just concatenate all the notes into a long-context model? Dementia-R1 reports that even 32B general models reason poorly over non-monotonic multi-visit trajectories without structured intermediate targets. The i2b2 comparison also shows GatorTron-Base outperforming LLaMA3-8B and MeLLaMA-13B on this task. We still include a single-pass long-context extraction run as a secondary reference on the multi-note setting, so this is measured rather than assumed.
Proposed Experiments
The first experiment runs on 2012 i2b2 (310 discharge summaries, over 60,000 relations). We plot relation F1 against verifier budget for every method and both verifiers, and report the budget each method needs to reach 95% of exhaustive F1. We also report the hit ratio, meaning the fraction of verifier calls that return a relation not already implied, in the sense used by the match-graph paper.
The second experiment runs on THYME. We report CONTAINS precision, recall, and F1, broken down into nominal and verbal events, because the THYME study identifies nominal events as the main source of lost recall. Where records can be grouped by patient, we build multi-note graphs and report cross-note relation recall separately.
The third experiment measures consistency. We compute the rate of ordering-constraint violations in the final timelines with and without contradiction-triggered re-verification, and compare it with the roughly 4% reported for soft propagation in Event Time Extraction and Propagation via Graph Attention Networks.
The decisive check is simple. If the active selector does not beat windows plus closure on cross-sentence and cross-note recall at matched budget, the transfer from image matching has failed. The idea would then reduce to temporal closure, which is already known.
References
- Match Graph Construction for Large Image Databases
- Global-Aware Edge Prioritization for Pose Graph Initialization
- Investigating the Challenges of Temporal Relation Extraction from Clinical Text
- Event Time Extraction and Propagation via Graph Attention Networks
- Self-verification improves few-shot clinical information extraction
- Prompting Large Language Models for Clinical Temporal Relation Extraction
- Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis