Self-Referenced Labs: Separating Patient Baselines from Disease Signal in Fair Multimodal Clinical Prediction

Routine blood tests such as the complete blood count and the metabolic panel are the most widely available clinical data in the world. Ratios built from them, like the neutrophil-to-lymphocyte ratio (NLR), predict mortality and disease severity across many conditions. These tests carry a hidden problem, though: healthy people from different populations sit at systematically different baseline values. A model that reads an absolute neutrophil count, or a fixed NLR cutoff, can therefore mistake a population's normal for a patient's health. When such lab streams are fused with imaging, the result may be a model that misses disease more often in some groups. A recent benchmark observed exactly this kind of outcome: fusing ICU health records with chest X-rays widened racial gaps in sensitivity. That benchmark did not identify which input caused the widening.

This note proposes to find out, and to test one cheap fix. The fix is to feed the model each lab value as a deviation from that same patient's own earlier measurements, rather than as an absolute number. The question is whether that removes much of the group gap without using race labels and without losing accuracy. The work has two parts. The first assigns a multimodal model's group sensitivity gap to its lab, imaging, and interaction components. The second tests whether "self-referenced" labs shrink the part of the gap carried by labs. This matters most in the realistic setting where imaging is missing for most patients and the labs carry the prediction.

Core Claim

When Does Multimodal Learning Help in Healthcare? A Benchmark on EHR and Chest X-Ray Fusion (CareBench) benchmarks 14 fusion methods on MIMIC-IV and MIMIC-CXR. It finds three things. Fusion raises phenotyping AUPRC from 0.472 to 0.495 when both modalities are present. Fusion widens racial disparities, mainly through unequal true positive rates. And with about 75% of X-rays missing, a model trained only on health records often wins. This paper motivates the proposal: it supplies the phenomenon but not the mechanism.

The working hypothesis is specific. A meaningful share of the true-positive-rate gap comes from the lab stream, not the image stream, because absolute lab values encode population-specific baselines. The research file flags systematic differences in reference ranges for leukocyte subsets and albumin as an untested threat to transfer.

The proposed representation changes each lab feature from an absolute value $x_{i,k}(t)$ to

$$z_{i,k}(t) = \frac{x_{i,k}(t) - m_{i,k}}{s_k},$$

where $i$ indexes the patient, $k$ the lab test, and $t$ the time. Here $m_{i,k}$ is the median of patient $i$'s measurements of test $k$ taken before the prediction window, and $s_k$ is the pooled within-patient standard deviation of test $k$. When a patient has no prior history, $m_{i,k}$ falls back to an age-and-sex reference median estimated on the training set. Race is never used.

The representation removes between-person baseline differences by construction while keeping acute change, which is what an inflamed or deteriorating patient shows. It also addresses a second known weakness of the biomarker literature: reliance on single snapshots. The NHANES COPD study, The neutrophil-lymphocyte ratio predicts all-cause and cardiovascular mortality among United States adults with COPD, derives one population-wide cutoff (NLR ≥ 2.56) from one draw.

The proposal differs from prior work in two ways. CareBench and Benchmarking Multimodal Clinical Foundation Models to Reveal Significant Demographic Disparities audit disparities but do not attribute them to a modality. Fusion architectures such as DrFuse and MedFuse change how streams are combined and never measure fairness. Self-referencing changes what the lab stream says before fusion happens, and it can be dropped into any of those architectures.

Research Program

The first component is a gap attribution study on CareBench's matched cohort (7,149 stays) and base cohort (26,947 stays). For in-hospital mortality and the 25 phenotypes, it trains three models on identical splits: labs only, X-ray only, and fused. The fused model uses MedFuse's sequential fusion module, since MedFuse handles partially paired data.

Three analyses follow. The first compares group-wise true positive rates at a shared operating point across the three models. The second is a baseline-shift test. In each group's held-out data, lab values are moved by the difference between that group's training-set median and the overall training-set median, computed among patients negative for the target. The test then measures how much the fused model's gap closes. If baselines drive the gap, this intervention should close much of it. If it does not, the hypothesis is weakened before any method is built.

The third analysis applies the shortcut test from Detecting shortcut learning for fair medical AI using shortcut testing (ShorT). ShorT varies a gradient-reversal weight on a demographic head attached to an image encoder and checks whether encoding correlates with unfairness. This work reuses that method on the lab encoder and the X-ray encoder separately, extending a test that has so far only been run on single-modality image models.

The second component swaps in self-referenced labs and repeats everything. The study reports results separately for modality-distributed conditions (heart failure, COPD) and localized ones (sepsis), because CareBench shows that fusion helps only the former. It also reports results separately for patients with and without prior lab history, because access to care may itself differ by group.

That last point is the main confound, and the program treats it as a measured quantity rather than an assumption. It reports the history-availability rate by group. It also checks whether the fallback path, which uses only age-and-sex reference medians, reintroduces the gap.

The third component is a cross-institution check on INSPECT. INSPECT pairs CT pulmonary angiograms with health records and includes mortality endpoints. The lab encoder trained on MIMIC is applied to INSPECT patients' routine labs, combined with the frozen image embeddings used in the foundation-model benchmark, and evaluated on mortality and pulmonary embolism. Race and age gaps are measured as underdiagnosis rate, matching that paper's metric.

Baselines

The naive reference point is the model with no change at all: MedFuse fused on absolute lab values, plus the labs-only model. CareBench reports that the labs-only model is often the stronger of the two under realistic missingness. On the partially paired MIMIC test set, MedFuse reports in-hospital mortality AUROC 0.874 and AUPRC 0.594, and phenotyping AUROC 0.768 and AUPRC 0.429. Reproducing these numbers on our splits is a precondition for comparison. For the racial true-positive-rate gap itself, no reported result was found on Lacuna for CareBench's numerical gap values, so this proposal will measure it.

The strongest accuracy competitor is DrFuse. It splits each patient's features into record-specific, image-specific, and shared parts, and weights modalities per disease. It reports an 8% relative PRAUC gain under missing modalities. It is the strongest competitor because, if modal inconsistency rather than lab baselines explains the gap, a better fusion design alone might narrow it. DrFuse will be run on absolute labs and on self-referenced labs. That 2×2 design separates the effect of the representation from the effect of the architecture.

The strongest fairness competitor is adversarial debiasing, in which a head tries to predict the demographic attribute and the encoder is trained to defeat it. Benchmarking Multimodal Clinical Foundation Models to Reveal Significant Demographic Disparities reports that this cut the age underdiagnosis gap from 0.333 to 0.069 (79%) at a cost of 0.011 AUROC on the 632M-parameter MedImageInsight. The same method failed on the 86M-parameter BiomedCLIP. Our lab encoders are small, so this capacity limit is a reason to expect self-referencing to compete.

All methods share splits, encoders, and a fixed hyperparameter search of equal size, with thresholds chosen on validation data.

Two ablations isolate the key ingredient. The first is "population-normalized" labs, which standardize each test against the whole training cohort. These change scale but not baseline, so they separate a plain normalization effect from the effect of removing personal baselines. The second is self-referencing with history artificially removed, which tests whether the benefit comes from the patient's own baseline or merely from added temporal information.

Beating the baselines means that, at matched overall AUROC (within 0.01 of the best baseline), self-referencing reduces the worst-group true-positive-rate gap more than adversarial debiasing does. It must also do so in the 75%-missing base cohort, not only in the matched cohort.

Why not just…?

Why not just add race to the model or use race-specific lab reference ranges? This is a real threat, because it targets the same baseline mechanism directly. It will be run as a baseline that conditions lab normalization on recorded race. Its weaknesses are that it needs a race label at deployment, it treats a social category as a proxy for a biological baseline, and it fails for the many patients poorly described by any category. Self-referencing uses the patient's own data instead. If race-conditioned normalization closes the gap and self-referencing does not, we will report that plainly.

Why not just set a separate decision threshold per group after training? Post-hoc threshold adjustment can equalize true positive rates by construction. It is the cheapest serious alternative and will be included. However, it buys equal sensitivity by raising false positives in the lower-scoring group, and it requires group labels at deployment. The comparison is therefore between false positive rate and AUROC at equalized sensitivity. If self-referencing reaches equal sensitivity with fewer false positives, the representation has carried real signal that thresholding cannot recover.

Why not just drop imaging and use a gradient-boosted model on lab ratios? Development and validation of an artificial intelligence model for predicting de novo distant bone metastasis in breast cancer reports an external AUC of 0.908 from routine labs with LightGBM, and CareBench finds that record-only models often win under missingness. A LightGBM model on absolute and on self-referenced labs will be included. The concern about baselines applies just as much to a lab-only model, so if self-referencing helps there too, the finding is about the labs rather than about fusion. CareBench's gains for heart failure and COPD indicate that imaging still matters for modality-distributed conditions, and the per-condition split will show where.

Why not just hide demographic variables from the model? ShorT shows that models reconstruct demographics from pixels without being given them, and Fairer AI in ophthalmology via implicit fairness learning shows that grounding predictions in pathology beats blinding. Blinding cannot remove baseline information carried in absolute lab values. A demographic probe on lab embeddings, compared between absolute and self-referenced inputs, will measure how much of that information remains.

Decisive Checks

Three results would decide the question. First, the baseline-shift test must close a substantial share of the fused model's true-positive-rate gap. If it closes little, the mechanism is wrong, and the project reduces to a negative result about where fusion bias lives.

Second, self-referenced labs must narrow the worst-group gap at matched AUROC on CareBench's base cohort, beating adversarial debiasing and showing a better false-positive tradeoff than per-group thresholds.

Third, the benefit must hold for patients without prior history (the fallback path) and must transfer to INSPECT with a smaller race underdiagnosis gap than absolute labs.

A cheap check on NHANES runs alongside these. It asks whether the published NLR cutoff of 2.56 has different sensitivity for mortality across racial groups in that cohort. No reported result was found on Lacuna for this, and the analysis needs no deep learning to run.