Reading Clearance Out of Clinical Notes to Pin Down Hidden Carriers in Hospital Infection Models

Models of how hospital infections such as MRSA (methicillin-resistant Staphylococcus aureus, a drug-resistant bacterium) spread are now built at fine spatial resolution. They split a state into hundreds of counties and individual facilities, and they move patients between these units using real transfer records. These models, however, learn only from billing claims, and claims record when someone is diagnosed but not when they stop carrying the bacterium. This proposal argues that the gap is more than an inconvenience. When a model sees only new diagnoses, it cannot tell a small pool of people who carry the bacterium for a long time from a large pool who carry it briefly, because both produce the same number of new cases. Those two worlds call for different interventions. Clinicians do write down the missing quantity: notes record repeat negative screens, completed decolonization, isolation precautions being lifted, and "history of MRSA" on readmission. We propose to extract each patient's carriage timeline from free-text notes, summarize it per facility as a distribution of carriage durations, and use that distribution to fix the parameter the claims-only models cannot identify.

The two-sentence version is this. Fine-grained infection models fit new-case counts well, but they cannot learn how long people stay infectious, so their estimates of which facility to target rest on an assumed number. We read that number, facility by facility, out of longitudinal clinical notes, and we test whether doing so changes which interventions the model recommends.

Core Claim

The model we build on is CALYPSO. It couples a neural calibration network with an SIRS metapopulation simulator. SIRS is a compartmental model in which people move from susceptible to infected to recovered and back to susceptible, and "metapopulation" means it runs separate copies of this model for 644 county and facility units and links them by travel. The authors note that claims lack recovery dates, so recovery is handled through literature-based assumptions. The deeper problem shows up in a simplified single-unit picture of an endemic infection, where new infections and clearances roughly balance:

$$ \text{incidence}_t \approx \gamma \, I_t , $$

Here $\gamma$ is the per-person clearance rate and $I_t$ is the number of people carrying the pathogen at time $t$. Claims observe the left side. Any pair ($\gamma$, $I$) with the same product fits the data equally well, and a calibration network that outputs both from case histories inherits this ambiguity. The size of the carrier pool $I$ is what drives the "silent bridge" that Miller et al. measured in claims: a family member's hospitalization raises household MRSA risk even without a diagnosed infection (IRR 1.44), and the risk grows with length of stay. It is also the quantity that decides how much an intervention at a given facility removes.

Our claim is that per-facility carriage-duration distributions extracted from longitudinal notes make $\gamma$ identifiable. With $\gamma$ fixed, the facility-targeting conclusions of a fine-grained metapopulation model should change measurably. Short-horizon forecast error may change much less, and we treat that difference as a prediction to test, not a weakness to hide.

Research Program

The work has three parts. Each draws on one of the input directions in a different role.

Patient-level carriage timelines from notes. This part reuses existing extraction methods but has to solve an open aggregation problem. HyDE classifies short masked contexts around disease mentions. It reports zero-shot AUPRC of 0.82–0.95 on unseen conditions and a 44-point F1 gain over rule-based labelers. It stops at single mentions, though, and leaves open how to combine them into a patient-level trajectory. We use a HyDE-style mention classifier to label each MRSA-related mention with a status: active infection, colonized, cleared or negative screen, historical, or ruled out. A long-context note reader then reconciles those mention labels into one interval-censored state sequence per patient. For the reader we follow ClinicalMamba, a state-space language model pretrained on multi-visit notes that handles up to 16k tokens. "Interval-censored" means clearance is known only to lie between the last positive evidence and the first negative evidence. Visits are irregular, so we borrow the onset-timing view of LATTE, which showed that billing codes lead or lag true onset, and we treat note dates as noisy, censored observations of the underlying state change. Dementia-R1 shows that a longitudinal note model trained first to extract verifiable intermediate quantities reasons better downstream. Our intermediate quantity is clearance time, and repeat cultures provide checkable anchors where they exist.

Facility-level duration distributions. Patient timelines are pooled within each spatial unit, either a care unit or a facility. We fit a censoring-aware survival model to get a carriage-duration distribution with uncertainty for each unit. Spatial granularity matters here. A long-term care facility and an acute ward plausibly differ in clearance, and a single statewide literature value removes exactly the facility differences that fine-grained models were built to capture.

Constrained calibration. In a CALYPSO-style pipeline, the calibration network keeps estimating transmission rates, while clearance becomes a prior centered on the extracted distribution for each unit, with width set by that distribution's uncertainty. We then rerun CALYPSO's facility-targeting analysis under this prior.

There is a data constraint we cannot avoid. The Virginia all-payer claims data cannot be linked to notes. We therefore develop the method on MIMIC-III, a de-identified single-hospital database with over 82,000 visits. There we treat care units as patches, use recorded inter-unit transfers as the travel matrix, and use structured microbiology results as partial ground truth. Full multi-facility validation needs a health-system partner whose notes and transfers cover several facilities. We present that stage as conditional on access.

Baselines

The naive reference is CALYPSO's own setup: an SIRS metapopulation model with clearance either set to literature values or learned freely by the calibration network from case histories. The authors report that this system cuts 4-week-ahead MSE by 18% and improves statewide $R^2$ by about 4.5% over an LSTM, and cuts MSE by over 95% relative to a mechanistic model calibrated by ensemble adjustment Kalman filter (EAKF). It is the strongest existing method on this question. We reimplement it on the same patches, transfer matrix and case series, and give it the same tuning budget as our variant. The only thing that changes between the two is the clearance prior. Beating it does not mean a lower MSE. It means recovering the true clearance rate and true intervention effects in simulation where it fails to, and, on real data, producing intervention rankings that agree better with held-out evidence on carriage, while keeping forecast error within its error bars.

The second baseline is clearance estimated from structured data only. It uses time from positive culture to first negative culture in structured microbiology results, under the same survival model. This ablation isolates what the notes add on top of structured data.

For the extraction step, the reference points are a rule-based keyword and negation labeler and HyDE's mention classifier without longitudinal reconciliation. We have no number on the file for MRSA carriage-status extraction ("no reported result found on Lacuna for MRSA clearance extraction from notes"). We therefore build a small expert-annotated set of patient timelines and evaluate on it, scoring both status accuracy and error in the estimated clearance date.

Why not just…?

Why not just let the calibration network learn the clearance rate, as CALYPSO already does? The identity above says incidence alone fixes only the product of clearance rate and carrier count. Added network capacity cannot recover what the data do not contain. We test this directly in simulation. We generate epidemics from known per-facility clearance rates on a CALYPSO-sized graph, fit the unconstrained model, and check whether its learned clearance values track the truth or stay at whatever value its initialization and training landed on.

Why not just use structured culture results instead of notes? This is the real threat. Where repeat cultures are routine, structured data may be enough. Our expectation is that most colonized patients never get a documented negative culture inside one system. For those patients, evidence of clearance or persistence appears only in text: outside screening results, decolonization, isolation being discontinued, or "history of MRSA" at readmission. The structured-only ablation settles the question. If its duration estimates and intervention rankings match the text-augmented ones, we report that notes are unnecessary and that the contribution reduces to using EHR clearance at all.

Why not just sweep clearance over its literature range and report sensitivity? If CALYPSO's facility rankings stay the same across the plausible range, the whole proposal is moot. We therefore run this sweep first, as a go/no-go check. If rankings do flip, a sweep tells a health department that the answer is uncertain but not which answer holds at its own facilities, and the extracted distributions supply that local answer.

Proposed Experiments

The first experiment is the sensitivity sweep. It asks whether the top-five facility targets in a CALYPSO-style model change as clearance varies across published ranges. The second is the simulation study. It measures recovery of clearance parameters and of true intervention effects under three conditions: free learning, structured-only priors, and text priors. In this study we deliberately corrupt the text-derived prior with censoring bias and extraction noise. CALYPSO reports that 3% noise in facility data significantly degrades predictions, so robustness to a wrong prior is a primary outcome. The third is extraction on MIMIC-III notes against expert-annotated timelines and against structured cultures where they exist, compared with rule-based and mention-only baselines. The fourth is the MIMIC-III unit-level model. We compare forecast error across the three clearance conditions and test held-out predictions of colonization status at readmission, which partly checks the inferred carrier pool. The decisive result combines several findings. The sweep shows that rankings are sensitive. Simulation shows that text priors recover truth where free learning cannot. Note-derived durations differ from structured-only durations at the facility level by more than their uncertainty, and the prior's benefit survives realistic extraction noise. If clearance-derived priors improve parameter recovery but notes add nothing over cultures, we will say so.