Many clinical studies measure inflammatory proteins such as interleukin-8 or TNF-α with ELISA (enzyme-linked immunosorbent assay, a plate-based antibody test that can detect proteins at picogram-per-millilitre levels). They then follow patients only until discharge, which may be a matter of days. That is enough to show that a cytokine predicts an in-hospital complication. It is not enough to say whether the same cytokine predicts death or heart failure one to five years later, and that longer-range question is the one that matters for follow-up care. Meanwhile, large retrospective electronic health record (EHR) cohorts do track patients for years, but they almost never contain research-grade cytokine measurements. The two data sources have complementary gaps, and no single cohort contains both the rich marker and the long outcome.
This note proposes to estimate the long-term prognostic value of an ELISA biomarker by fusing the two sources. The small cohort with ELISA values and short follow-up supplies the link from cytokine to the patient's short-term clinical state: the routine labs, vitals and events seen before discharge. The large cohort with routine EHR data and long follow-up supplies the link from that short-term state to long-term survival. Chaining the two links, while carrying through ELISA-specific noise such as detection limits and plate-to-plate drift, yields a long-horizon hazard estimate with honest uncertainty. The method rests on a testable assumption, which we call the bridging assumption: the cytokine's long-term effect passes through the short-term state that both cohorts observe. We make that assumption explicit and design the experiments to show when it breaks.
Core Claim
A long-horizon survival curve for an acutely measured ELISA marker can be recovered, without bias beyond a quantifiable amount, from two ingredients. The first is a short-follow-up ELISA cohort. The second is a separate long-follow-up routine-data cohort that shares a common short-term state $Z$. Write $X$ for the ELISA marker vector, $T$ for time to death, and $Z$ for the discharge-time state (routine inflammatory labs, organ-support indicators, in-hospital events). The bridged estimate is
$$ \hat S(t \mid x) = \int S_L(t \mid z)\, p_E(z \mid x)\, dz , $$
where $S_L(t \mid z)$ is the survival function learned in the long-follow-up cohort and $p_E(z \mid x)$ is the distribution of the short-term state given the marker, learned in the ELISA cohort. The estimate is exact when $T$ is independent of $X$ given $Z$. The bias grows with any direct pathway from marker to outcome that $Z$ does not capture.
What is new is not the chaining idea itself. It resembles surrogate-index reasoning from economics. What is new is applying it to immunoassay cohorts with limited follow-up, with three additions. First, $p_E$ is fitted under ELISA measurement structure, namely left-censoring at the lower limit of detection and plate random effects. Second, the uncertainty from both cohorts is propagated. Third, the method is validated against an oracle created by artificially truncating follow-up.
Why This Gap Is Real
The motivating evidence comes from ELISA cohort studies that stop at the hospital door. Yang et al.'s pancreatitis study measured IL-8, procalcitonin, CRP and TNF-α by ELISA in 180 patients and predicted in-hospital cardiovascular events with a logistic model (AUC 0.893 in training, 0.978 in a 36-patient validation set). It tracked nothing after discharge. Cao et al. found that plasma TNFR1/2 and urinary ICAM-1, but not CRP, tracked kidney injury in diabetic men. That study was purely cross-sectional, and the authors state that long-term renal outcomes could not be assessed. Retrospective cohorts with limited follow-up show the same limitation from the other side: Deng et al. studied 10,707 septic patients in MIMIC-IV with ICU-only mortality and noted survivor bias in length of stay.
Two methods papers constrain and supply the design. The multivariate Gaussian-process joint model links repeated albumin and calcium measurements to survival in 929 dialysis patients. It shows that carrying the last observation forward in a Cox model shrinks the hazard ratio for a 1 g/dL albumin drop from 5.1 to 3.4. The lesson we take is that ignoring measurement structure biases effects toward the null, so our $p_E$ must model assay noise explicitly. That model, however, needs the marker and long follow-up in the same patients, which is exactly what ELISA cohorts lack. Deep Survival Analysis shows that generative survival models can learn from very sparse EHR data (complete variables in only 1.4% of patient-months). It reaches a C-index of 73.11% versus 65.57% for the Framingham score, which makes it a reasonable candidate family for $S_L$. It does not handle immunoassays or cohort fusion.
Research Program
We will build the bridge in three parts.
The first part is the long-horizon model $S_L(t \mid z)$, fitted on a large retrospective EHR cohort. It uses a discharge-anchored time origin and the same $Z$ definition as the ELISA cohort, and it restricts $Z$ to variables that routine labs measure in both settings. We will fit both a flexible Weibull-mixture model and a deep survival model, and choose between them only by held-out calibration.
The second part is the marker-to-state model $p_E(z \mid x)$, fitted on the ELISA cohort. In this model, values below the detection limit enter as left-censored observations rather than being replaced with a fixed value such as the limit divided by √2. Plate identity enters as a random effect on the log-concentration scale. Replicate wells, when available, inform the assay's coefficient of variation. We fit it as a Bayesian model so that we can draw posterior samples of $z$.
The third part is uncertainty propagation. We draw samples jointly from both posteriors, or from bootstrap replicates of both cohorts, and integrate them to obtain interval estimates for long-horizon survival and hazard ratios.
A sensitivity parameter $\delta$ will add a direct log-hazard term $\delta \cdot x$ that bypasses $Z$. For each target quantity we will report how large $\delta$ would need to be to overturn the conclusion. This makes the bridging assumption auditable rather than hidden.
The main target application is the setting of Yang et al.: acute inflammatory illness (pancreatitis, sepsis), with ELISA taken at admission and follow-up that ends at discharge. The long-follow-up source will be MIMIC-IV, which is extractable with the MIMIC-IV processing pipeline. One caveat is essential. We will verify how far MIMIC-IV's recorded date of death extends past discharge before relying on it, and we will cap the horizon at what the data support.
Baselines
The naive reference point uses the ELISA cohort alone. We fit a Cox model on baseline ELISA values with follow-up ending at discharge, then read off whatever horizon the data allow. We also include its logistic-regression variant, which mirrors Yang et al., whose reported in-hospital AUC is 0.893 (training) and 0.978 (validation). For long-horizon survival there is no reported result on Lacuna for any bridged or extrapolated ELISA estimate. We count beating this baseline as producing long-horizon estimates at all with nominal interval coverage, where the naive model can only produce extrapolations.
The second baseline is parametric extrapolation: a Weibull model fitted to the short-follow-up ELISA cohort and extended to one year. Extrapolation is the standard way to get long-horizon numbers from truncated data, which makes it the most direct competitor. We fit it on identical data with the same covariates.
The third baseline uses routine markers only. A long-follow-up model on the EHR cohort uses CRP and white-cell count in place of cytokines. It answers the practical objection that ELISA adds nothing beyond routine labs. The bridge must show a gain in time-dependent C-index or net reclassification at one year over this model for the ELISA marker to earn its cost.
The strongest existing method is the multivariate Gaussian-process joint model. Because it cannot run without co-observed marker and outcome, we run it as an oracle on the semi-synthetic benchmark, where full follow-up is available for the marker subsample. Its reported advantage is a hazard ratio of 5.1 versus 3.4 for last-observation-carried-forward Cox. We count the bridge as successful if its hazard-ratio estimate falls within the oracle's 95% interval. It uses less information than the oracle, so matching the oracle, not beating it, is the goal.
Two ablations isolate the key ingredients. The first removes detection-limit and plate modelling from $p_E$ and substitutes limit/√2. The second shrinks $Z$ to admission-only variables, which tests whether the short-term trajectory carries the signal. All methods share covariates, the same data splits, and a fixed hyperparameter budget of 50 configurations chosen by the same validation criterion.
Why not just…?
Why not just extend follow-up by linking the ELISA cohort to death registries? When linkage is possible, it is the best design, and we do not argue otherwise. We use exactly such truncation-and-extension as ground truth in our validation. In many settings, though, linkage is legally or practically unavailable: single-centre Chinese hospital cohorts, older biobanks, or cross-border data. In those cases a bridge with a quantified bias bound is the alternative to saying nothing.
Why not just use CRP from the EHR and skip ELISA? This is a real threat. For some outcomes, CRP may carry all the long-term information. Cao et al. provide the counter-evidence: CRP was elevated but uncorrelated with kidney injury, while ELISA-measured TNFR1/2 tracked injury. Their study was cross-sectional, however. The routine-markers-only baseline settles the question directly. If the bridged cytokine model adds no one-year discrimination, we report that ELISA's added value is limited to the short term.
Why not just fit a joint longitudinal-survival model? It requires serial marker values and long follow-up in the same patients. Retrospective ELISA cohorts usually have one sample per patient and follow-up that ends at discharge. We include the joint model as an oracle where the data allow it, and we measure how far the bridge falls short of it.
Proposed Experiments
The decisive test is a semi-synthetic truncation benchmark on MIMIC-IV, built from the Sepsis-3 respiratory-failure extraction of Deng et al.. We choose a sparsely measured inflammatory lab as a stand-in for the ELISA marker. In a random subsample of about 200 patients, sized to match Yang's cohort, we keep that lab and add simulated ELISA artifacts: detection-limit censoring at the 20th percentile, plate shifts with a 15% coefficient of variation, and follow-up truncated at discharge. The remaining patients keep long follow-up but have the marker hidden. The oracle is a Cox model and a joint model fitted with both the marker and full follow-up on the full cohort.
Metrics are the bias of the one-year hazard ratio relative to the oracle, the coverage of 95% intervals, the time-dependent C-index and integrated Brier score at one year, and calibration slope.
We then plant a direct marker-to-outcome effect $\delta$ of increasing size to show where the bridge fails and whether the sensitivity analysis flags it. The bridging assumption would be refuted in practice if realistic values of $\delta$ produce hazard-ratio bias larger than the naive extrapolation does. In that case we would say so and recommend linkage.
Finally, we repeat the benchmark with albumin as the stand-in marker, following the joint-model setting, to check that results do not depend on one lab.
If partners can link a real ELISA pancreatitis or sepsis cohort to post-discharge outcomes, that cohort becomes the external test: bridge estimates are produced blind, before linkage, and compared after.
References
- Prognostic value of inflammatory biochemical markers (IL-8, PCT, CRP) for cardiovascular disease in patients with pancreatitis
- Inflammation and Kidney Injury in Diabetic African American Men
- Impact of long-acting glucocorticoids on ICU mortality in septic patients with acute respiratory failure: a MIMIC-IV based cohort study
- A Flexible Joint Longitudinal-Survival Modeling Framework for\n Incorporating Multiple Longitudinal Biomarkers
- Deep Survival Analysis
- An Extensive Data Processing Pipeline for MIMIC-IV