Who Waits When a Patient Is Sped Up? Text-Labeled Capacity Spillovers in Hospital Causal Estimates

Hospitals increasingly estimate how many hours an operational change would save, such as faster consult responses or priority imaging, by having a large language model read each patient's notes and rebuild the chain of events that kept them in the hospital. These estimates quietly assume that speeding up one patient never slows down another. That assumption is called the stable unit treatment value assumption (SUTVA): each patient's outcome depends only on their own treatment. A hospital is a shared-capacity system. A CT scanner, a consult team, or a social worker that serves an expedited patient first serves someone else later. When that happens, the patient-level savings reported by current methods are partly time moved from one patient to another, and the hospital-wide benefit is smaller than the sum of individual savings.

This note proposes to measure that gap and correct for it. The key observation concerns where each kind of information lives. Structured timestamps show which patients were waiting for the same resource at the same time. Only the free-text notes say whether a given wait was actually capacity-bound ("awaiting MRI slot," "social work backlog") or clinically bound ("holding for repeat neuro exam," "family meeting Thursday"). Only capacity-bound waits can displace other patients. We use evidence-grounded text extraction to label the cause of each delay. We then combine those labels with a contention graph built from timestamps: a graph linking patients who were waiting for the same resource at overlapping times. From the two together we estimate a capacity-aware hospital-wide effect. The method is validated against simulated ground truth and against natural capacity shocks in real data.

Core Claim

The closest existing work, Expert-Guided g-computation with Large Language Models for Estimating Causal Effects on Timings, has a GPT-5 pipeline turn each patient's notes and orders into a patient-specific causal graph. It then simulates eleven quality-improvement interventions, reporting for example 18.9 hours saved per screened stay for consult escalation. Its authors state that the method relies on SUTVA. Our claim is that its overstatement is predictable and largely correctable. To a first approximation, the hospital-wide effect is the SUTVA estimate minus the displaced delay:

$$\text{GTE} \approx \sum_i \Delta_i \;-\; \sum_i \sum_{j \in C(i)} \pi_{ij}\, \delta_{ij}$$

The symbols are as follows. $\Delta_i$ is the patient-level time saved for patient $i$, as computed by the existing method. $C(i)$ is the set of patients who shared a queue with $i$ in the contention graph. $\pi_{ij}$ is the text-derived probability that the wait $i$ skips was capacity-bound. $\delta_{ij}$ is the extra wait patient $j$ incurs when $i$ is served earlier, estimated from historical queue data.

The new ingredient is $\pi_{ij}$. Without it, every shared queue looks binding and the correction is too large, which is the "saturated graph" failure. With $\pi_{ij} = 0$, we recover SUTVA. We predict that the corrected estimates for capacity-heavy interventions, such as imaging priority, will shrink substantially. Interventions acting mainly on clinically bound steps should barely move. If both kinds shrink alike, or neither shrinks, the text labels add nothing, and that result would be informative.

What Prior Work Establishes and Leaves Open

Three strands motivate the question. Interference, Bias, and Variance in Two-Sided Marketplace Experimentation shows that when treated units capture finite shared resources, controls are starved and the global effect is overestimated. Causal Inference under Interference and Model Uncertainty shows the same direction of error formally for linear models. Its "overestimation trap" says that ignoring same-sign spillovers inflates the estimate. Queue displacement has the opposite sign to the direct saving, so in hospitals the trap appears as inflated savings rather than inflated harm. Causal Inference Under Interference And Network Uncertainty shows that ignoring interference costs roughly an order of magnitude in bias (about 0.06 versus 0.006 in their simulations). Assuming everyone interferes instead inflates variance, so the interference structure has to be learned rather than assumed.

Two papers constrain the design. The estimator in Graph Machine Learning based Doubly Robust Estimator for Network Causal Effects (GDML) combines graph neural networks with double machine learning. It needs a set of evaluation patients whose neighborhoods do not overlap. Co-queue graphs in a busy ward are dense, so we expect heavy pruning of that set, and we will measure it rather than assume it away. Experimentation under Treatment Dependent Network Interference is relevant because reprioritizing a queue changes who is queued with whom; the treatment rewires the graph.

On the text side, we reuse methods rather than invent them. Self-Verification Improves Few-Shot Clinical Information Extraction forces a model to quote supporting text and prune unsupported items, raising GPT-4 F1 on MIMIC-IV ICD-10 coding from 0.487 to 0.533. Information Extraction from Clinical Notes: Are We Ready to Switch to Large Language Models? supplies a cost-matched comparison. Fine-tuned BERT comes within 1% of LLaMA-3-70B in-domain at roughly one twenty-eighth the latency, but trails it on unseen institutions (0.798 versus 0.872 F1 on i2b2). The motivation for reading notes at all, that delay reasons live in free text rather than structured fields, comes from The Age of Data-Driven Medicine. None of these extraction papers looks across patients, and none feeds extraction errors into a causal estimate.

Research Program

The program has three components.

Delay-cause labeling. For every waiting interval on a patient's critical path, such as the gap between an imaging order and its completion or between a consult request and the consultant's note, an extractor reads the surrounding notes. It labels the wait as capacity-bound, clinically bound, or indeterminate, and must quote verbatim evidence. Indeterminate cases get a calibrated probability rather than a forced label. We build the extractor on the self-verification recipe and train a BioClinicalBERT version through Kiwi for throughput. Clinicians annotate a stratified sample of several hundred delay episodes to measure labeling accuracy and calibration.

Contention graph and displacement rates. From order and completion timestamps we connect patients waiting concurrently for the same modality, consult service, or placement resource. We then estimate $\delta_{ij}$, how much a queue-mate's earlier service lengthens a patient's own wait. The estimate is restricted to intervals labeled capacity-bound and comes from historical variation in queue position, for example where order priority differs among concurrent requests.

Capacity-aware evaluation. For each of the eleven interventions in the egg-computation study, we rerun the patient-level simulation. Each simulated acceleration then propagates displacement to queue-mates through the equation above, and we update the contention graph when the intervention reorders the queue.

This is also an empirical validation exercise for estimators that have so far been tested only on synthetic graphs, citation networks, or village surveys. We will report exactly where they fail on real hospital topology.

Baselines

The do-nothing reference is the SUTVA estimate from egg-computation itself. On the ZSFG cohort it reports 18.9 hours per screened stay for consult escalation and 15.6 hours for disposition-critical imaging priority. Both are the numbers our correction targets. We rerun that pipeline unchanged on the same timelines, which keeps the comparison exact.

The saturated alternative treats everyone in the same unit and time window as interfering. It is implemented by setting all $\pi_{ij} = 1$.

The congestion-covariate baseline adds daily unit census to the downstream timing regressions.

The strongest existing network estimator is GDML, run on the structured contention graph with the same nuisance-model tuning budget we give our own outcome models. We will report its estimate, its confidence interval, and how many patients survive focal-set selection. The Greedy Debias algorithm from the uncertain-interaction-graph paper is a second network baseline. It uses edge probabilities from co-queue frequency, without text.

The key ablation is our full method with $\pi_{ij}$ replaced by a structured-only proxy, such as "wait exceeded the modality median." This isolates what reading the notes contributes. A second ablation swaps the self-verified extractor for one-pass extraction to measure how hallucinated labels propagate into effect estimates.

No reported result found on Lacuna gives hospital spillover magnitudes or a ground-truth global effect for any of these baselines. Beating them is therefore defined on our own validation sets. On the simulation, our method should reach lower absolute error on the hospital-wide effect than every baseline, with interval coverage near nominal. On real data, it should pass the capacity-shock check described below where the structured-only ablation does not.

Why not just…?

Why not just build the contention network from structured logs and skip the text? This is the most serious alternative. Structured logs do identify who shared a queue, and we use them for exactly that. What they cannot say is whether a long wait was binding capacity or a deliberate clinical pause. The egg-computation authors found that coarse, structure-level graphs misplaced effects by up to 14 hours relative to note-derived patient graphs. The structured-only ablation tests this directly. If it matches the full method on the simulation and on the capacity-shock check, we will report that text is unnecessary for this question.

Why not just run a discrete-event queueing simulation of the hospital? Such a simulation can represent displacement, but it needs every step classified as resource-consuming or not. That classification is what our extractor provides. We will run such a simulation fed with our labels as an additional comparator. If it matches our estimator, the contribution narrows to the labels and the validation protocol, which we would state plainly.

Why not just apply GDML or another network estimator off the shelf? These estimators assume a known, fixed graph and enough patients with non-overlapping neighborhoods. Co-queue graphs are dense and are rewired by the intervention itself. We expect focal-set pruning to discard most of the cohort and will report the surviving sample size as a finding.

Why not just use fine-tuned BERT for the labels? It may suffice. Hu et al. show near-parity in-domain at a fraction of the cost. We will compare it against the self-verified LLM on clinician-annotated episodes, using both labeling F1 and the downstream error in the hospital-wide effect. The BERT version is the deployment default unless cross-site transfer degrades it.

Decisive Experiments

Semi-synthetic test. We extend the 4,000-patient simulation from egg-computation with explicit finite servers for imaging and consults, so the true hospital-wide effect of each intervention is known. We vary the fraction of capacity-bound waits and inject label noise at rates matched to the measured extractor error. The test asks whether our corrected estimate tracks truth as SUTVA's bias grows, and how quickly accuracy degrades as label error rises.

Real-data capacity-shock falsification check. This runs on MIMIC-IV notes and timestamps, and on ZSFG timelines if the original authors can share them. Periods of reduced capacity, such as scanner downtime or holiday staffing, should lengthen waits labeled capacity-bound and leave waits labeled clinically bound essentially unchanged. If both respond equally, the labels carry no causal information and the program stops there.

Ranking comparison. We report how the ranking of the eleven interventions changes once displacement is subtracted. This connects the method back to the operational decision the original study was built to inform.