How much a patient's vital signs swing is a strong predictor of death. A wide daily temperature range, fluctuating sodium, and unstable blood pressure all predict mortality independently of the average value. The catch is that the swing a hospital records depends on how often someone measures it. If a nurse charts temperature 24 times a day, the observed maximum minus minimum will be wider than if they chart it 4 times, even when the patient's physiology is identical. How often patients are charted differs across hospitals and, within one hospital, across patient groups who receive different intensities of care. So a variability feature learned at one site silently mixes biology with charting habits. When the model moves to a small new hospital, or is applied to an under-monitored subgroup, it misreads sparse charting as physiological stability and under-calls risk.
We propose to measure this confound directly and to remove it. Each patient's vital-sign series will be passed through a small state-space model that handles irregular time gaps. The model returns an estimate of the underlying volatility, split into slow drift and fast jitter, together with its uncertainty, and that estimate does not depend on how many measurements were taken. Its prior is fitted once on a large source ICU, so a new hospital needs no labels to compute the features, and only a few dozen labeled patients to recalibrate. The key experimental lever is HiRID, which records vitals every two minutes. By thinning HiRID to the charting schedules of other hospitals, we get the same patients observed at different cadences, which serves as a ground truth that no prior variability study has used.
Core Claim
Our central claim is empirical and testable. A substantial part of the cross-site and cross-group degradation of variability-based ICU risk models comes from measurement intensity rather than from differences in patient biology. An estimator that separates the two will close most of that gap in small target cohorts, and it will narrow subgroup disparities in missed deterioration. The supporting reasoning is elementary. For roughly independent readings, the expected range of $n$ samples grows like $\sqrt{2\log n}$, which is the standard extreme-value growth of Gaussian maxima. The daily temperature gap that Soltani et al. found to predict mortality linearly across MIMIC-IV and eICU is therefore mechanically inflated at sites and in patients that are charted more often. Forward-fill imputation creates flat segments that shrink standard deviations in the opposite direction.
Adapt the Data, Not the Model motivates this work. It showed that recording cadence and forward-fill habits, not marginal value shifts, drive much of the MIMIC-IV to eICU and HiRID degradation. Its fix, however, is a learned adapter that needs target labels and does not isolate volatility. MIMIC-IF supplies the population angle. It documents care-intensity disparities in MIMIC-IV, such as shorter ventilation for Black and Hispanic patients at comparable presentations, and a negative correlation between a group's mortality rate and model AUC. Unequal care intensity plausibly means unequal charting, which would bias variability features by group. To our knowledge, no work in the file tests this link.
Research Program
The estimator borrows its structure from Yu et al.. That study decomposed heart-rate variability into slow trajectory wiggliness and fast residual jitter within one Bayesian joint model. It showed that plugging empirical standard deviations into a downstream regression collapses confidence-interval coverage to 11.5%, against 94.5% for joint estimation. It also showed that the two scales can move in opposite directions under stress. Its MCMC fitting struggled with small groups, and it was designed for regularly sampled laboratory data from 165 Black women.
We keep the decomposition but make it fast and cadence-aware. For patient $i$ and signal $s$, reading $y_{ik}$ at time $t_{ik}$ is modeled as
$$y_{ik} = \mu_i(t_{ik}) + \varepsilon_{ik}, \quad \varepsilon_{ik}\sim\mathcal N(0,\sigma_{e,i}^2),$$
$$\mu_i(t_{ik}) = \mu_i(t_{i,k-1}) + \eta_{ik}, \quad \eta_{ik}\sim\mathcal N\!\left(0,\ \sigma_{b,i}^2\,(t_{ik}-t_{i,k-1})\right).$$
Here $\mu_i$ is the latent physiological level, $\sigma_{b,i}^2$ is drift variance per unit time (slow volatility), and $\sigma_{e,i}^2$ is fast jitter plus measurement noise. Because the drift variance scales with the actual time gap, a Kalman filter handles irregular charting exactly, and both variance parameters are defined per unit time rather than per reading. Each patient's two variances get a log-normal prior whose hyperparameters are estimated on the source hospital by empirical Bayes. Patients with few readings are therefore shrunk toward the population, not reported as falsely stable. The per-patient posterior means and standard deviations of $\log\sigma_b$ and $\log\sigma_e$, computed over rolling windows for heart rate, blood pressure, respiratory rate, and temperature, replace raw standard deviation and max-minus-min features in an otherwise standard gradient-boosted tree. This is inexpensive enough for bedside use, unlike LLM serialization.
The program has three parts. First, a measurement study quantifies how much standard variability features shift under controlled thinning of HiRID and how much of their MIMIC-IV subgroup differences is explained by measurement counts. Second, the cadence-invariant features are validated: their distributions should be stable under thinning while their prognostic value is retained. Third, a transfer and fairness evaluation trains on a large source, adapts with 16 to 200 labeled target patients, and audits subgroup performance at the target.
Baselines
The naive reference applies the source-trained LightGBM with the standard HiRID-benchmark summary features (minimum, maximum, mean, measurement density over lookback windows) directly to the target with no adaptation. The same features are also trained on the target cohort alone at each small sample size. This mirrors the target-only Cox model that Bellot et al. report at a C-index of 0.500 on the small Euro study.
The strongest in-domain competitor is that same LightGBM pipeline, because HiRID-ICU-Benchmark reports it beating deep sequence models, with an AUPRC of 38.9 for circulatory failure. We will reproduce the benchmark's preprocessing and use an identical hyperparameter search budget for every feature set. Our method must match that in-domain AUPRC within noise; any loss there is a cost to be reported.
The strongest transfer competitor is INPUTADAPTER. It reports outperforming or tying every baseline, including end-to-end fine-tuning, on MIMIC-IV to eICU, with a +14.1 point AUCPR gain on acute kidney injury. We will run it with the same frozen LightGBM-equivalent predictor where its architecture permits, or else a matched GRU predictor, on the same target labeled subsets. Beating it means higher target AUPRC at 16 to 64 labeled patients, and a smaller worst-subgroup gap at any budget. We do not expect to beat it at large target sizes.
Record2Vec is the few-shot transfer reference, since it approaches a 36,000-sample model with 16 target samples. For numerical target scores on HiRID or eICU splits, no reported result was found on Lacuna, so we will rerun it with MedGemma on identical splits. TSB instance reweighting (C-index 0.677 on Euro) is run on the same tabular features to test whether reweighting source patients already fixes the shift.
Two ablations isolate the key ingredient. The first removes the time-gap scaling, treating readings as equally spaced, to test whether cadence-awareness matters. The second removes source-fitted shrinkage, using per-patient maximum-likelihood variances, to test whether borrowing strength matters when the target is small.
Why not just…?
Why not just add measurement count as a feature to LightGBM? This is the most serious threat, and the benchmark features already include density. It lets the source model learn how the observed range relates to count at the source. But count also encodes care intensity, which predicts outcome through clinician concern, and that relationship is site- and group-specific according to the care disparities MIMIC-IF documents. The model will therefore partly learn "frequently charted means sick," a shortcut that should break across hospitals and groups. The thinning experiment settles the question. If count-augmented LightGBM is as stable as our features when HiRID is thinned, and matches them on eICU and in subgroup audits, the estimator adds nothing, and we will report that.
Why not just downsample every site to a common, sparsest cadence before computing features? This removes the expected-range bias by construction and is a genuine competitor for bias. It discards most readings, though. With few target patients, the resulting variability estimates are noisy, and cadences are irregular within a patient, so no single grid is truly common. Our estimator should match its invariance while having lower variance. We will include it as a baseline and compare both bias under thinning and small-sample AUPRC.
Why not just use INPUTADAPTER or fine-tuning, which already learn cadence corrections? Both need target labels and enough of them to fit an adapter network. Neither exposes whether the correction removed a measurement artifact or a real biological difference, and INPUTADAPTER was not audited for subgroup equity. Our label-budget curves test the small-sample claim directly. The subgroup audit tests whether a learned correction happens to fix group-level measurement disparities; it may, and if so that is a finding in its favor.
Proposed Experiments
The first decisive check uses HiRID's two-minute data. We train each model on full-resolution HiRID, then evaluate on the same held-out patients after thinning their vitals to cadences sampled from MIMIC-IV and eICU charting patterns, with forward-fill applied as those systems do. Patients and outcomes are fixed, so any drop in AUPRC for circulatory and respiratory failure, and any drift in feature distributions measured by Wasserstein distance per feature, is attributable to cadence alone. We predict that the naive max-minus-min and standard-deviation features drift in the direction the $\sqrt{2\log n}$ argument implies, and that our posterior volatility features stay stable.
The second check is an audit in MIMIC-IV, where ethnicity, sex, age, and insurance are recorded. We will test whether readings per day differ across groups after adjusting for acuity score. We will then test whether the temperature gap and blood-pressure standard deviation differ by group among patients with the same outcome, and whether those differences shrink under our estimator. The outcome measure is the group gap in false-negative rate at a fixed alert rate for mortality and deterioration. The worst-group AUC will also be checked against the risk-performance inversion that MIMIC-IF reports.
The third check is a transfer evaluation. Models are trained on MIMIC-IV and transferred to eICU and HiRID with 0, 16, 32, 64, 128, and 200 labeled target patients, repeated over 20 random draws. We report target AUPRC, calibration slope, and worst-subgroup gap against all baselines above. Repeated draws matter because at 16 patients the differences between methods can be smaller than the variation between draws.
The proposal succeeds if three results hold together. Thinning should degrade naive features but not ours. Measurement intensity should explain a measurable share of the subgroup variability differences. And at 64 or fewer target patients, our features should beat both count-augmented LightGBM and common-cadence downsampling on worst-subgroup error without losing in-domain AUPRC. If count-augmentation matches us everywhere, the contribution reduces to the measurement study, which is still worth publishing as a warning about variability features.
References
- Association of body temperature and mortality in critically ill patients: an observational study using two large databases
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- A Bayesian Approach to Modeling Variance of Intensive Longitudinal Biomarker Data as a Predictor of Health Outcomes
- Boosting Transfer Learning with Survival Data from Heterogeneous Domains
- HiRID-ICU-Benchmark - A Comprehensive Machine Learning Benchmark on High-resolution ICU Data