Removing Fluid Dilution from ICU Lab Trajectories

When an intensive care patient receives several liters of intravenous fluid, the concentration of nearly everything measured in their blood falls, including albumin, hemoglobin, and total protein. Nothing about the underlying biology has to change for this to happen. Clinical trajectory models read these drops as organ failure or recovery anyway. This proposal borrows a move from geochemistry, where weathering is treated as a systematic and invertible distortion rather than as noise. We model each patient's serial lab values as a shared, time-varying "dilution factor" layered on top of analyte-specific biological trajectories. We then recover the undiluted trajectories and ask whether they predict mortality better than the raw ones.

The idea fits in two sentences. Fluid loading and capillary leak scale all co-measured plasma analytes by roughly the same multiplicative factor at each moment, while real biology moves them independently. So a joint model across several analytes can estimate that shared factor and subtract it, much as geochemists use immobile elements to reconstruct what a rock looked like before alteration. If this works, it turns "albumin dropped 4 g/L" from an ambiguous number into a separable statement about plasma volume versus hepatic synthesis and leak. It would also give a principled way to build the longitudinal, corrected versions of indices such as the albumin-corrected anion gap, which is currently computed only once.

Core Claim

The work on ancient glass shows that post-depositional alteration can be inverted when it follows deterministic rules. Two of those papers do this cross-sectionally. One uses logistic regression and discriminant analysis to classify weathered glass by type with 100% accuracy across 67 relics (Analysis of the identification results of the ancient glass products based on the progressive regression model). The other uses a back-propagation network to predict pre-weathering composition with error below 0.1 (Composition analysis and identification model of ancient glass products based on Spearman correlation coefficient and BP neural network). These papers motivate our claim that systematic alteration can be removed. They also hint at its main pitfall: the same alteration process moves silica in opposite directions in different glass matrices. Neither has a time axis or a mass-balance constraint.

Our claim is that hemodilution in plasma is the clinical analogue, and a simpler one, because its main effect is a common multiplicative factor across analytes. We write the observed log-concentration of analyte $k$ for patient $i$ at time $t$ as

$$\log y_{ikt} = m_{ik}(t) + d_i(t) + \epsilon_{ikt},$$

where $m_{ik}(t)$ is the undiluted biological trajectory of analyte $k$, $d_i(t)$ is a log-dilution factor shared by all analytes in that patient, and $\epsilon_{ikt}$ is measurement noise.

The model is identifiable only if something anchors $d_i(t)$. We use two anchors, playing the role that immobile elements play in geochemistry (Yin et al.; Liu). The first is charted fluid intake and output, which gives a mass-balance prior on how $d_i$ should move. The second is red-cell mass, which changes only through charted transfusion or bleeding events, so hemoglobin's biological trajectory is tightly constrained between those events.

The specific novelty is the joint, longitudinal, multi-analyte decomposition with a physical prior. Current practice offers two alternatives: a single static correction formula, or a net-curve trajectory model that ignores dilution.

Relation to Prior Work

Three clinical papers motivate the gap. The first ties the change in albumin during hospitalization for acute myocardial infarction to mortality through an S-shaped curve with inflection points at −3.27 and 0.92 g/L (Association between albumin changes and prognosis in older patients with acute myocardial infarction). It notes that resuscitation fluid and capillary leak drive albumin, but it never separates them, and we do. The second applies a static albumin correction to the anion gap at ICU admission and reaches a Random Forest AUC of 0.73 for 28-day mortality in acute respiratory distress syndrome (Albumin-corrected anion gap as a predictor of 28-day mortality in acute respiratory distress syndrome: A machine learning-based retrospective study). It has no trajectory, and we make the correction longitudinal and dilution-aware. The third finds that the neutrophil-to-lymphocyte ratio predicts ICU mortality and closes by calling for serial trends instead of snapshots (The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study).

Two papers supply the method. TimeFRAME is a Bayesian model in Stan that separates source mixing from consumption-driven isotopic fractionation in gas time series. It uses Gaussian-process and Dirichlet–Gaussian-process priors to stop parameters oscillating between time steps (Technical note: A Bayesian mixing model to unravel isotopic data and quantify trace gas production and consumption pathways for time series data – Time-resolved FRactionation And Mixing Evaluation (TimeFRAME)). We reuse its idea of temporally correlated latent alteration terms. Our setting differs in that the alteration term is shared across many analytes rather than inferred from one signal, which sidesteps TimeFRAME's acknowledged difficulty in identifying the consumed fraction. The Bayesian joint model of intensive biomarker variance is the second method source (A Bayesian Approach to Modeling Variance of Intensive Longitudinal Biomarker Data as a Predictor of Health Outcomes). In simulation under heavy measurement error it kept 94.5% parameter coverage where a two-stage approach fell to 11.5%. That result tells us to propagate uncertainty about $d_i(t)$ into the outcome model rather than plugging in point estimates. That work is univariate and assumes stationary variance, whereas our latent factor is multivariate and driven by non-stationary fluid events.

One paper constrains the design: the HiRID benchmark found that gradient-boosted trees on hand-built window features beat recurrent, convolutional, and attention models (HiRID-ICU-Benchmark - A Comprehensive Machine Learning Benchmark on High-resolution ICU Data). We therefore do not propose a new end-to-end sequence network. The corrected trajectories become features for the strongest existing predictor.

Research Program

We build the cohort from MIMIC-IV, the dataset used in the ACAG study, which contains lab values together with charted fluid inputs, outputs, and transfusions. The analytes are albumin, hemoglobin, hematocrit, total protein where available, and the electrolytes needed for the anion gap. Each $m_{ik}(t)$ gets a Gaussian-process prior with analyte-specific length scales. The shared factor $d_i(t)$ gets a Gaussian-process prior whose mean is driven by a one-compartment volume model, with net charted fluid balance as its input. A per-patient leak coefficient lets albumin, but not hemoglobin, escape into tissue faster when capillaries leak; this is the analogue of matrix-dependent leaching. Transfusion and bleeding enter as jumps in the hemoglobin trajectory.

Inference runs in Stan following TimeFRAME's setup. If per-patient sampling proves too slow at MIMIC scale, we fall back to a variational approximation, validated against full sampling on a subset.

The outputs are, for each patient, corrected trajectories of albumin and of the albumin-corrected anion gap with posterior uncertainty, plus a dilution trajectory. These feed the outcome models described below.

Baselines

The naive reference point is the admission snapshot. For the ARDS task, the reported comparison is admission ACAG with a Random Forest, which achieved AUC 0.73 on the validation set, with other learners at 0.69 to 0.72 (Albumin-corrected anion gap as a predictor of 28-day mortality in acute respiratory distress syndrome: A machine learning-based retrospective study). We re-run it on our cohort definition so that it shares our splits.

The second baseline is the raw-trajectory approach from the albumin study: the change between admission and hospitalization-average albumin, entered through restricted cubic splines. We reproduce it on a MIMIC-IV acute myocardial infarction subset. No reported result found on Lacuna for this approach on MIMIC-IV, because the original cohort is a private single-center one with 288 analyzed patients.

The strongest competitor is LightGBM with engineered rolling window features over raw labs. On HiRID this kind of model reached AUPRC 38.9 for circulatory failure and beat every deep sequence model tested (HiRID-ICU-Benchmark - A Comprehensive Machine Learning Benchmark on High-resolution ICU Data). In our setting we give it raw lab windows plus charted fluid-balance and transfusion windows. It therefore sees exactly the information our model uses, only without the structural decomposition. All learners share the same patients, splits, and Bayesian-optimization hyperparameter budget (equal number of trials).

Beating this baseline means one of two things. Either LightGBM on corrected trajectories achieves higher held-out AUROC and AUPRC than LightGBM on raw labs plus fluid features, with bootstrap intervals that exclude zero difference, or it matches that performance while giving trajectories that pass the physiological checks below.

Two ablations isolate the key ingredient. The first removes the fluid-balance prior, so dilution is inferred from analyte co-movement alone. The second removes the multi-analyte sharing, so each analyte is corrected only by its own fluid-driven volume model.

Why not just…?

Why not just use ratios of co-measured analytes, such as albumin divided by hemoglobin? A ratio cancels any factor the two analytes share, with no model needed, and this is the most serious threat. It breaks wherever the denominator moves for biological reasons, as with bleeding, transfusion, or hemolysis. These are common in exactly the patients for whom dilution matters. The glass literature shows the same failure: one alteration process drives silica up in one matrix and down in another (Component analysis of ancient glass products based on hierarchical analysis clustering algorithm). We include the ratio as a baseline feature set and stratify results by transfusion exposure. If the ratio matches our model even in the transfused stratum, we will report that the latent model adds nothing.

Why not just add fluid balance as a feature to LightGBM? This may match us on raw prediction, because trees can learn interactions. That is why it is our strongest baseline rather than a straw man. The difference is that it yields no corrected trajectory, so it cannot answer questions such as whether an albumin drop reflects leak. Our claim stands or falls on two things: whether corrected features add predictive value on top of this baseline, and whether the corrected trajectories pass the bolus check below. If they add nothing on either count, the decomposition is only interpretive and we will say so.

Why not just let a deep sequence model learn dilution from raw data? On both available ICU benchmarks, recurrent and attention models failed to beat tree models: on HiRID at 5-minute resolution, and on MIMIC-Extract hourly intervention tasks, where Random Forests matched recurrent networks at Macro AUROC 0.90 to 0.94 (MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III). We include a GRU on raw inputs for completeness, but nothing in the file suggests it will be the binding competitor.

Why not just use the static admission correction? Admission values cannot tell a falling patient from a rising one. The albumin study's S-curve shows that the direction and size of change carry mortality risk, with an odds ratio of 2.12 between its inflection points (Association between albumin changes and prognosis in older patients with acute myocardial infarction).

Proposed Experiments

The first check is synthetic recovery. We take real albumin and hemoglobin trajectories from patients who received little fluid, inject a known dilution signal built from other patients' fluid charts, and measure how accurately each method recovers the original trajectory and how well calibrated its uncertainty is. Following the joint-model study, we report parameter coverage alongside error.

The second check, and the decisive physiological one, is a bolus natural experiment. In patients with labs drawn shortly before and after a charted crystalloid bolus, the inferred dilution factor should explain the drop in hemoglobin and albumin. The corrected albumin trajectory should then show no discontinuity at the bolus that cannot be explained by noise. Raw trajectories and the ratio baseline serve as comparisons. If the corrected trajectory still jumps at boluses, the dilution model is wrong, and we stop before running the outcome tasks.

The third check is prognostic. For 28-day mortality in the MIMIC-IV ARDS cohort, we compare corrected longitudinal ACAG against the admission-ACAG reference (AUC 0.73 reported) and against the strongest LightGBM baseline. In a MIMIC-IV acute myocardial infarction subset, we test whether the S-shaped relationship between albumin change and mortality sharpens or moves once dilution is removed. A finding that the inflection points shift toward zero would suggest that part of the reported "decompensation" signal is fluid.

As a secondary, exploratory transfer, we fit the model to HiRID patients with charted fluids and test whether corrected features improve on circulatory-failure AUPRC 38.9. We treat this as exploratory because the fluid charting in HiRID is not described in the file.