Clinical early-warning models learn the numbers of the hospital they were trained in: its typical heart-rate baselines, how often it draws labs, and how long results take to be logged. When moved to a new hospital they degrade. Sensory neuroscience has studied a structurally similar problem. When a smell gets stronger or a visual scene gets brighter, every neuron fires earlier and harder, yet the brain still recognizes what it is sensing. It does this by reading which neurons fire before which others, and by how much, rather than how loudly any one neuron fires. We propose to represent a patient's multi-biomarker trajectory the same way. The representation is the set of relative lags between when each signal starts to deviate from that patient's own baseline: heart rate, respiratory rate, temperature, lactate, white cells. It is built to be insensitive to the hospital-specific shifts that break current models.
The specific new ingredient is how the lags are estimated. Clinical timestamps are noisy, because charting is delayed and draws are irregular, so a naive "time of first threshold crossing" is fragile. We instead estimate each pairwise lag as the center of mass of a cross-correlogram, which is the histogram of time offsets between deviation events in two channels. This is the statistic used to show that cortical spiking sequences are stereotyped. We show it inherits a jitter-robustness property from point-process theory: the integral of the correlogram is unchanged by independent timestamp noise, and its center of mass is unchanged by zero-mean noise. The test is whether this representation narrows the cross-hospital gap in sepsis prediction, and whether it keeps prognostic signal in septic patients, where static ratios such as the neutrophil-to-lymphocyte ratio fail.
Core Claim
Two lines of neuroscience motivate the representation. In a 50,000-neuron olfactory bulb model, higher odor concentration advanced all spike latencies while preserving their relative pattern, so identity could be decoded independent of intensity (Psychophysical properties of odor processing...). On recorded salamander retina, a timing-sensitive readout detected luminance boundaries at roughly 4% error, against roughly 15% for a rate-based perceptron, because global contrast shifted all spikes together (Computing Complex Visual Features with Retinal Spike Times). Over days and weeks, millisecond timing stayed stable while firing rates drifted (Temporal coding carries more stable cortical visual representations than firing rate over time). Burst timing also stayed rigid while burst magnitude varied, with drift falling orthogonal to the coding manifold (Stable representation of a naturalistic movie emerges from episodic activity with gain variability).
These papers motivate the idea that "relative timing survives gain and baseline shifts." They do not establish it for physiology, and we treat that transfer as the hypothesis under test, not as evidence.
The method comes from two other papers. Gating of Sensory Input by Spontaneous Cortical Activity quantified within-packet firing order with a cross-correlogram center-of-mass statistic, and found it correlated at about r ≈ 0.75 between spontaneous and evoked activity. Cumulants of Hawkes Processes are Robust to Observation Noise proved that integrated cumulants of multivariate point processes are invariant to independent random translation of event times, while maximum-likelihood fits collapse. That proof was validated only on synthetic and financial data.
The center-of-mass lag is defined as follows. Let $C_{ij}(\tau)$ be the cross-covariance density between deviation events in channel $i$ and channel $j$ at lag $\tau$, estimated within a window of a single patient's stay. The lag feature is
$$ \ell_{ij} = \frac{\int \tau \, C_{ij}(\tau)\, d\tau}{\int C_{ij}(\tau)\, d\tau}. $$
If each event in channel $i$ is displaced by independent noise with mean $d_i$, then $C_{ij}$ is convolved with the distribution of the difference of two noises. Convolution preserves the integral, which is the Trouleau et al. result at second order. It also shifts the first moment by exactly $d_j - d_i$. The consequences are:
- Zero-mean charting jitter leaves $\ell_{ij}$ unbiased.
- A hospital's systematic lab-reporting delay appears as a constant per-channel offset.
- That offset can be removed by centering each hospital's lag features on its own unlabeled patients.
Deviation events are defined relative to each patient's own running baseline, so absolute level and per-hospital calibration offsets drop out before lags are computed.
This differs from the closest existing work in three ways. Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments compresses irregular ADNI visit histories into fixed vectors aligned to a population trend; it uses values, not inter-channel onset order. Does Temporal Encoding Matter? found that elapsed-time encodings added nothing over event order for GRUs on MIMIC-IV code streams. That result constrains us: it warns that adding "time" naively can hurt, with sinusoidal encodings dropping next-event precision from 38.7% to about 20%. But it encoded time since the previous event on administrative codes, not lags between physiological channels. The PhysioNet 2019 challenge documented that vitals lead labs and that hospital distribution shift hurts generalization, yet most entries used hourly bins with forward fill (Early Prediction of Sepsis from Clinical Data).
Research Program
Event extraction. For each of the challenge's 40 variables, we detect deviation events as crossings of a patient-specific baseline. The baseline is a robust running median, with a threshold in units of that patient's own median absolute deviation. Both the threshold and the baseline window are tuned on training-site data only.
Lag estimation. In sliding windows of 12 to 24 hours we estimate $C_{ij}$ with a kernel estimator. We keep $\ell_{ij}$ together with its integral as a reliability weight, because sparse channels give few event pairs. Pairs with too few events are marked missing rather than imputed.
Hospital centering. We center lags per hospital using unlabeled patients from that hospital, which needs no target labels.
Models. The lag vector, its weights, and the event counts feed two downstream models: a gradient-boosted classifier and a small GRU over successive windows. This separates the value of the representation from model capacity.
Synthetic validation. Before any clinical claim, we validate the estimator on simulated multivariate Hawkes data. We inject three kinds of distortion: zero-mean jitter, constant per-channel delays, and multiplicative gain. We check that $\ell_{ij}$ behaves as the convolution argument predicts.
Baselines
Naive reference. The do-nothing reference is the challenge's rule-based or static-score approach: an admission-time severity score applied unchanged at the new site. This is motivated by the finding that admission scores decay over the stay (Optimal intensive care outcome prediction over time using machine learning).
Strongest existing method. The strongest competitor is the hourly-binned, forward-filled sequence model typical of top PhysioNet 2019 entries. Top teams reached utility scores of 0.33 to 0.36 (Early Prediction of Sepsis from Clinical Data). We reimplement it as a GRU and as gradient boosting on binned values plus delta features. Every model gets the same splits and the same hyperparameter search budget, measured in trials and GPU-hours.
One caveat on that number: the 0.33–0.36 figure was scored on the hidden test sets. We could not find a reported Lacuna result for the standard models on a public cross-site split. We therefore rerun all baselines ourselves rather than compare to that number directly.
Order-only baseline. A GRU over the order of deviation events with no lag values is the "order is enough" baseline of Marco et al. For MIMIC-IV mortality, the reference point is the hybrid LSTM at AUROC 0.87 in An Extensive Data Processing Pipeline for MIMIC-IV.
Ablations. Four ablations isolate the key ingredient:
- Integrated correlogram mass only, which is jitter-invariant but has no lag.
- Naive first-crossing lags instead of center-of-mass lags.
- Lags without hospital centering.
- A shuffle control that permutes lag vectors across patients with the same event counts.
What counts as winning. Beating the strongest baseline means a smaller drop in utility and AUROC from in-site to held-out-site, at equal or better in-site performance. Bootstrap confidence intervals over patients must exclude zero.
Why not just…?
Why not just standardize each variable per hospital? This is the most serious threat. Per-site z-scoring or quantile normalization directly removes the cumulative-distribution shifts Reyna et al. documented, and it is trivial to implement. We run it as a baseline and also combine it with the strongest model. Normalization cannot correct differences in sampling frequency or reporting delay; lags centered on unlabeled data can. If per-site normalized binned models close the cross-site gap as well as lag features do, we will report that as the answer. In that case the timing hypothesis adds nothing.
Why not just rely on event order, as the GRU foundation-model study suggests? Does Temporal Encoding Matter? is a real constraint. Its authors limit the conclusion to dense ICU code streams, but order alone may still carry most of the signal here. The order-only GRU over the same deviation events tests this exactly. If it matches the lag model, the magnitude of relative latency is not useful clinically.
Why not just add rate-of-change and variability features? Trajectories and volatility already beat snapshots for albumin (Association between albumin changes and prognosis...) and sodium (Sodium fluctuation as a parameter in predicting mortality). Delta and variability features are included in the strongest baseline. They are per-channel and scale with gain and sampling rate, whereas lags are cross-channel. The ablations show whether cross-channel timing adds anything beyond them.
Proposed Experiments
Cross-site sepsis prediction. The primary experiment uses the public PhysioNet 2019 data. We train on Hospital A and test on B, then the reverse, reporting challenge utility and AUROC. As a third, more distant site, we map MIMIC-IV ICU stays to the same 40 variables and Sepsis-3 labels.
Synthetic robustness check. We inject controlled timestamp jitter and per-channel delays into the held-out site's records. The prediction is that center-of-mass lag models degrade much less than naive-lag and binned models, and that the remaining degradation tracks the per-channel delay until centering is applied.
Sepsis-paradox test. On MIMIC-IV, we ask whether neutrophil, lymphocyte, and inflammatory-marker lags predict 28-day mortality inside the sepsis subgroup, where admission NLR lost significance. The reference point is that adding NLR moved SAPS I AUC only from 0.746 to 0.757 overall (The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness). Labs are sparse, so this test may be underpowered. We will report the number of patients with estimable lags.
Decisive outcome. The idea succeeds if three conditions hold together: the lag model has a smaller cross-site gap than per-site-normalized binned models, it beats order-only and shuffled-lag controls, and it retains signal in sepsis. The idea fails if normalization alone matches it.
References
- Psychophysical properties of odor processing can be quantitatively described by relative action potential latency patterns in mitral and tufted cells
- Computing Complex Visual Features with Retinal Spike Times
- Temporal coding carries more stable cortical visual representations than firing rate over time
- Stable representation of a naturalistic movie emerges from episodic activity with gain variability
- Gating of Sensory Input by Spontaneous Cortical Activity
- Cumulants of Hawkes Processes are Robust to Observation Noise
- Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments
- Does Temporal Encoding Matter? Evidence from GRU-based Clinical Foundation Models
- Early Prediction of Sepsis from Clinical Data: the PhysioNet/Computing in Cardiology Challenge 2019
- Optimal intensive care outcome prediction over time using machine learning
- An Extensive Data Processing Pipeline for MIMIC-IV
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Sodium fluctuation as a parameter in predicting mortality in general hospitalized patients
- The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study