An Open MIMIC-IV Benchmark for Latent Transmission and Far-UVC Dose Evaluation

Hospitals cannot see their main transmission engine. City-scale simulation of community-associated MRSA found that asymptomatic, colonized individuals drive over 95% of new colonization events — roughly 2.1 million colonizations versus only 116,000 clinical infections over ten years in Chicago — while hospitals contributed just 7.8% of spread (Modeling the transmission of community-associated methicillin-resistant Staphylococcus aureus: a dynamic agent-based simulation). Clinical predictive models built on electronic health records inherit this blindness because their outcome labels come from diagnostic tests that only symptomatic or suspected patients receive. Meanwhile, 222-nm far-UVC lamps can continuously disinfect occupied rooms — proteins in the dead outer skin layer and tear film absorb the photons before they reach living cell nuclei — achieving a Killing Log above 3 on S. aureus, E. coli, and P. aeruginosa in 15 seconds where a conventional 254-nm mercury lamp needs more than 10 minutes (Study on disinfection effect of a 222-nm UVC excimer lamp on object surface). But no study connects the physics of that lamp to real patient flows, unobserved carriers, or the confounded way we would measure its effect from records.

This proposal builds the missing bridge as a single artifact: a standardized, open benchmark on MIMIC-IV that (a) defines a reproducible latent-transmission task with fixed cohort extraction and fixed splits, (b) couples the PALS latent-spreader inference model to a dose-dependent far-UVC exposure field so that the physics of disinfection enters the transmission model, and (c) ships a causal audit that tests whether any observational evaluation of the intervention would survive the processes-of-care confounding that famously fabricated a near-50% mortality reduction from a random coin flip in the TREWS audit (Randomized Clinical Trials or Convenient Controls: TREWS or FALSE?).

Core Claim

Latent transmission models, physics-based disinfection dosing, and retrospective causal evaluation have never been composed, and each has a specific hole the others fill. PALS (Learning the Probability of Activation in the Presence of Latent Spreaders) infers which unobserved contacts are contagious — it achieved an AUROC of 0.705 on a nurse-sharing network across 20,147 admissions — but it ran on a custom, unreleased single-center cohort, treats the network as static, and contains no environmental intervention. The far-UVC carrier study (Study on disinfection effect of a 222-nm UVC excimer lamp on object surface) quantifies the contrast-toxicity tradeoff precisely — human safety and germicidal speed both come from protein absorption — but only on steel carriers in a chamber. And the standardized-pipeline literature, from NerfBaselines' exposure of ranking distortions from undocumented downscaling (NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods) to the open MIMIC-IV processing pipeline (An Extensive Data Processing Pipeline for MIMIC-IV), has never defined a transmission task at all. Our claim is that once these three are composed on public data, a concrete question becomes answerable that none can address alone: how much far-UVC irradiance, degraded by the organic-matrix shielding penalty that Ning et al. measured (15 s in buffer versus 60 s in protein-rich broth, a 4× dose penalty), is needed to meaningfully suppress transmission driven by carriers nobody tests, and would a retrospective before/after comparison of such a deployment even measure the truth?

The physics here is the safety mechanism, not a footnote. The same protein absorption that makes 222 nm safe for occupied rooms means that pathogens shed in fecal, mucous, or biofilm matrices are shielded. The proposal's central quantitative deliverable is a dose-to-impact curve: simulated transmission events averted per unit of installed irradiance, with the organic-load penalty explicitly varied, so that the safe-operation window (efficacy achieved without exceeding skin and corneal exposure limits) is computed rather than assumed.

Research Program

The first component is the benchmark itself, which we call the MIMIC-IV Transmission Task. It reuses the ingestion machinery of the existing pipeline (An Extensive Data Processing Pipeline for MIMIC-IV) — ICD-9/ICD-10 mapping, NDC medication normalization, percentile outlier trimming, leakage-free forward-fill imputation, and 2-hour time-window discretization — because we should not reinvent standardized preprocessing, and the pipeline's own limitation (strict tailoring to MIMIC-IV, no transmission tasks) is exactly the gap we fill. From the hospital module's ward-transfer records we construct a co-location contact network over time: two patients are connected during overlapping stays in the same ward within aligned time windows. We acknowledge this is coarser than PALS's nurse-sharing graph; the benchmark defines the network from what MIMIC-IV actually records, and network definition becomes an explicit, documented benchmark variable rather than a hidden engineering choice — the precise failure mode that NerfBaselines and BARS (BARS: Towards Open Benchmarking for Recommender Systems) showed inflates rankings. Outcomes are hospital-acquired C. difficile and MRSA, defined by positive laboratory tests more than 48 hours after admission, with fixed temporal train/test splits following the discipline of Wiens et al. (Learning Data-Driven Patient Risk Stratification Models for Clostridium difficile). Evaluation reports AUROC, PR-AUC, and calibration together, because the COP-IV experience — 86.86% AUROC collapsing to 17.24% PR-AUC under class imbalance (Revisiting Clinical Outcome Prediction for MIMIC-IV) — shows a single metric misleads.

The second component couples the benchmark to physics. We extend PALS's generative model — each contact $j$ carries a latent contagiousness variable $z_j \sim \mathrm{Bernoulli}(\sigma(u^\top x_j))$ inferred by mean-field variational inference — with a time-varying environmental survival term. In the baseline model, a contact's shedding is modulated only by clinical features (broad-spectrum antibiotics like ciprofloxacin get large positive learned weights in PALS). In ours, the effective infectious contact rate acquires a factor $\exp(-\kappa \, D_{ij})$, where $D_{ij}$ is the far-UVC dose accumulated on the shared surface or air volume between patients $i$ and $j$, and $\kappa$ is a pathogen inactivation coefficient. $D_{ij}$ is computed from a room-geometry irradiance model parameterized directly by Ning et al.'s measurements — 1796 µW/cm² at 32 cm from the excimer lamp, 5 to 10 times lower required dose than a 254-nm lamp — with a line-of-sight shadowing map and an organic-burden multiplier of 4 (the PBS-to-TSB penalty) applied to surfaces where the model infers heavy bioburden. The inference problem then jointly estimates latent spreader states, susceptibility parameters, and, in a deployment scenario, the effective inactivation rate, so posterior uncertainty on $\kappa$ reflects how much of transmission is environmentally mediated versus contact-mediated.

The third component is the causal audit, and it is where the retrospective-cohort direction does real work rather than decoration. No hospital in MIMIC-IV had far-UVC installed, so we do two things honestly. First, we apply the TREWS audit machinery — counterfactual "coin-flip" arms and backdoor adjustment for test-ordering behavior — to an observable intervention-like process in the records, contact precautions, to calibrate how badly naive before/after comparisons mislead in this exact data. Second, we run simulated deployments on the real retrospective cohort (the far-UVC dose field applied to real patient trajectories) and evaluate them twice: once with naive outcome comparison, once with propensity-score adjustment of the kind validated by Tewari et al. (Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach), where unadjusted comparisons were heavily biased by age, comorbidity, and race (20.0% versus 31.0% surgery rates for Black versus white patients) until matching recovered a valid 68% mortality reduction for surgery. The deliverable is a protocol showing exactly which measured confounders an eventual real deployment study must record, and quantifying the residual bias attributable to unmeasured ones.

Baselines

The naive reference point is the ward-level colonization-pressure proxy: the daily fraction of patients in a unit with confirmed infection, used as the exposure variable. Makar et al. explicitly benchmarked this against PALS and found it fails because it is blind to untested carriers; no numerical AUROC for the pressure proxy is reported on Lacuna, so the benchmark's first job is to produce one under fixed splits. The strongest existing method is PALS itself, which reported an AUROC of 0.705 on its nurse-sharing network over 20,147 admissions in a private cohort (Learning the Probability of Activation in the Presence of Latent Spreaders); we re-implement it, verify we reproduce its qualitative attribution behavior (antibiotics positive, contact precautions negative), and run it on the public MIMIC-IV networks with the same feature budget, tuning budget, and variational-inference settings that we give every competing method, so any gap from 0.705 reflects cohort and network differences made explicit rather than engineering choices. The second strong competitor is the data-driven susceptibility-only model of Wiens et al., which reported AUROC 0.81 against 0.71 for a 14-variable expert score across 69,568 admissions, with 38% higher sensitivity and 50% higher PPV at the 95th-percentile threshold (Learning Data-Driven Patient Risk Stratification Models for Clostridium difficile); we run its feature-extraction approach on identical MIMIC-IV windows. The key ablation is our own model with the dose term disabled ($D_{ij} \equiv 0$): it isolates whether adding environmental physics improves inference of latent carriers even when no lamp is installed — which it should, since shadowed and disinfected rooms should attenuate the inferred contact-spreading signal, sharpening attribution of transmission to truly contagious contacts rather than ambient contamination. For the intervention analyses, the naive unadjusted before/after comparison is itself a baseline, and the decisive result is that our audit reproduces the TREWS-style inflation on it before collapsing under adjustment.

Why not just…?

Why not just run a big agent-based simulation of far-UVC in a hospital? A simulation in the style of the Chicago MRSA model — 3 million agents, 4.125 × 10¹² contact events (Modeling the transmission of community-associated methicillin-resistant staphylococcus aureus: a dynamic agent-based simulation) — could sweep irradiance levels without touching a single EHR. The problem is anchoring: ABM parameters (shedding rates, contact intensities, bioburden distributions) would be guessed, and the model would never confront the messiness that real records introduce — informative testing, MNAR labs, coarse ward-level contacts. Our design uses the retrospective cohort for exactly what it is uniquely good for: calibrating susceptibility weights and validating that latent inference beats observed-case proxies on outcomes that actually happened. We will, however, use a small ABM layer for the community spillover question (whether in-hospital disinfection protects discharged households, motivated by Miller et al.'s finding that a hospitalized family member raises household MRSA risk, Hospitalizations among family members increase the risk of MRSA infection in a household), so the alternatives are complementary here, not adversarial.

Why not just use the existing Wiens-style daily risk predictor and skip latent inference? This is the most serious challenge. The daily-risk model achieved AUROC 0.82 at Michigan and 0.75 at MGH with a Brier score of 0.01 and a 5-day median lead time (A Generalizable, Data-Driven Approach to Predict Daily Risk of Clostridium difficile Infection at Two Large Academic Health Centers) — better discrimination than PALS's 0.705. But it answers a different question: it predicts who will test positive, which is a mix of true susceptibility and clinician testing behavior, and it represents exposure as a static unit average. It cannot attribute transmission to specific contacts, so it cannot tell you where a lamp matters or whether an intervention broke a chain. Our benchmark will report both model families on the same task, and the honest prediction is that the risk predictor wins on AUROC for test-positive prediction while PALS-style models win on contact-level attribution — if that prediction fails, the latent component's value claim fails with it, and we will say so.

Why not just compare hospital units with and without UV disinfection retrospectively? Because the TREWS audit already showed what happens: patients whose clinicians respond to suspicion get more tests and more care, and a purely random alert appeared to cut mortality from 6.4% to 3.3% ($p < 0.001$), an artifact that vanished once lactate-ordering behavior was adjusted for (odds ratios 2.21 and 1.6 in the two cohorts) (Randomized Clinical Trials or Convenient Controls: TREWS or FALSE?). A units-with-lamps versus units-without comparison would confound lamp installation with everything else that differs between units. The benchmark's audit component exists precisely to test, on synthetic deployments over real records, how large this bias would be and which measured covariates neutralize it — turning "just compare units" from a mistake into a protocol with a known error bar.

Why not just stick with 254-nm mercury lamps in unoccupied rooms? The file's own numbers close this door: the 254-nm lamp cannot run in occupied rooms at all, because unlike 222 nm it penetrates living epidermis and cornea (Mitigating Phototoxicity in High-Intensity Optical Applications), and in occupied rooms is precisely where asymptomatic carriers are continuously shedding. On speed it loses by roughly 40-fold (more than 10 minutes versus 15 seconds to a Killing Log above 3 in buffer conditions) (Study on disinfection effect of a 222-nm UVC excimer lamp on object surface). The one place the simple alternative bites is organic load, where 222 nm slows to 60 seconds — which is exactly why organic burden is an explicit variable in our dose-to-impact analysis rather than an assumption of clean surfaces.

Proposed Experiments and Decisive Checks

Experiment one establishes the benchmark: on the MIMIC-IV transmission task with fixed temporal splits, we run the colonization-pressure proxy, the Wiens-style susceptibility model, PALS, and our physics-coupled PALS, all under identical feature and tuning budgets, reporting AUROC, PR-AUC, and Brier calibration per method. Decisive outcome: the pressure proxy's AUROC (no reported result found on Lacuna for it; the benchmark produces the first number), PALS-family models beating it, and the $D_{ij} \equiv 0$ ablation quantifying the physics term's contribution. Experiment two produces the dose-to-impact curve: simulated far-UVC deployments across irradiance levels and organic-burden multipliers (1× to 4× per Ning et al.'s PBS-to-TSB penalty), reporting transmission events averted against cumulative human exposure dose, identifying whether the safe-operation window survives heavy bioburden. Decisive outcome: either a viable window exists at Ning-measured irradiances, or the 4× penalty closes it — both are publishable, and the second would be an important negative result for occupied-room far-UVC claims. Experiment three runs the causal audit: naive before/after comparison of simulated deployments versus propensity-adjusted estimates, plus the coin-flip-style arm on contact precautions, quantifying observational bias directly on MIMIC-IV. Decisive outcome: the audit reproduces the TREWS-style inflation pattern; if it does not, that itself calibrates how much MIMIC-IV's testing behavior differs from the TREWS settings and bounds the transferability of audit conclusions. All code, cohort definitions, splits, and irradiance parameters ship with the benchmark, in the spirit of OGB and BARS, so that the next transmission model is compared rather than asserted.