Hospital records do not sample patients on a schedule. Blood is drawn and vitals are charted more often when a patient looks sick. Most clinical prediction pipelines first "fill in" the gaps by drawing a line or smooth curve between observations, then fit a model. Filling a gap this way acts as a low-pass filter: it keeps slow trends and removes fast fluctuations. How much it removes depends on the gap length. This proposal starts from one consequence of that fact. Because gap length in clinical data is set by clinical concern, the filter is not the same for every patient. It is sharp for the sickest patients and heavily blurring for stable ones. Interpolation therefore does two things at once. It erases real fast dynamics from under-observed patients. It also turns "how often was this patient measured" into apparent physiology, because features such as variability, slope, and peak count end up partly measuring sampling density.
We propose to measure this coupling directly and to separate the prognostic signal in biomarker trajectories into three parts: the part carried by the values, the part carried by the timing of measurements, and the part destroyed by interpolation. The central tool is a "mask transplant" experiment. We take densely recorded vital signs, where the true signal is known. We impose on them the actual lab-draw timestamps of the same patient or of a randomly chosen other patient, then interpolate and predict outcomes. If distortion and its effect on mortality prediction differ between a patient's own mask and a random mask, the field's standard practice of evaluating imputation under random missingness understates the problem. It also means that some reported trajectory "biomarkers" are partly sampling artifacts.
Core Claim
For linear interpolation across a gap of length $g$, a sinusoidal component at frequency $f$ is retained with approximate gain
$$H(f; g) \approx \left(\frac{\sin(\pi f g)}{\pi f g}\right)^2,$$
which is the frequency response of the triangular kernel that linear interpolation applies. The effective cutoff is therefore roughly $1/g$ for each patient, set by that patient's local gap lengths. In clinical data $g$ is correlated with severity, so the cutoff is correlated with the outcome. Our claim has two parts, and each can be falsified.
First, under acuity-linked sampling, interpolation-based pipelines lose measurably more prognostic information than the same pipelines under random sampling at matched missingness. The loss is concentrated in low-acuity patients.
Second, a meaningful share of the apparent prognostic value of trajectory-variability features after interpolation can be recovered by a model that sees only the measurement timestamps and never the values.
The remedy we test is lightweight and follows from the diagnosis. We give the model each series' observation-based spectrum, estimated by the Lomb–Scargle periodogram. The Lomb–Scargle periodogram fits sinusoids directly to irregularly timed points, with no gap-filling. Alongside it we give the model its spectral window, which is computable from the timestamps alone and says which frequencies that patient's sampling can resolve. This separates "what the data show" from "what the schedule allowed us to see."
Relation to Prior Work
Several papers establish the low-pass mechanism, and they motivate this work. Fraternale et al. showed that linear interpolation across Voyager 2 tracking gaps damps high-frequency power and biases spectral slopes. Jang et al. showed that interpolation adds blurring and phase delays that hurt multivariate forecasting. Their ChannelTokenFormer avoids interpolation through period-driven patching, but it was tested on industrial and environmental sensors rather than clinical panels. Demirel et al. and Wickramasinghe et al. document phase cancellation from linear interpolation and spurious "ghost" oscillations from splines, respectively. All of these treat missingness as unrelated to the quantity being predicted. Our contribution is to study the case where it is not.
The closest physiological work is Morelli et al. together with the ultra-short HRV study. Both simulate burst missingness on heart-rate-variability (HRV) data from healthy subjects. They find that interpolating interval durations behaves like a low-pass filter, and that no interpolation at all is best for fast HRV features at 30–50% missingness. We reuse their burst-missingness protocol as a sanity check on nsr2db. We depart from them in two ways: we use real clinical masks rather than synthetic ones, and we measure effects on outcome prediction rather than on feature error.
On the method side, LSCD supplies the differentiable Lomb–Scargle layer. LSCD conditions a diffusion imputer on the uninterpolated spectrum and reports gains on PhysioNet 2012. It does not ask whether better spectra improve clinical prediction, and it does not model how sampling times relate to acuity. Trajectory Flow Matching models irregular ICU trajectories in continuous time and notes that it handles periodic dynamics poorly. IMMFM replaces straight-line flow paths with piecewise-quadratic ones to avoid jumps in velocity. Its paths are themselves a smoother, and our mask-transplant protocol can measure their frequency response.
We deliberately do not build the obvious hybrid of Lomb–Scargle priors inside flow matching. With the five to fifteen lab draws a typical ICU stay provides, the periodogram is badly conditioned. The spectral window makes that visible, and we treat it as information rather than something to hide.
The clinical literature provides the application regime and one pointed puzzle. Salciccioli et al. found that admission neutrophil-to-lymphocyte ratio (NLR) predicts 28-day ICU mortality (adjusted OR 1.71) but loses predictive power in sepsis. Zhang et al. and Pan et al. show that trajectories of albumin and hsCRP carry risk that snapshots miss. None of these studies separates trajectory signal from the density of measurement. Sepsis patients are both the most densely sampled and the most dysregulated, so the NLR paradox is a natural place to check.
Research Program
The first component is the mask-transplant benchmark. We will use the periodic vital-sign streams in eICU and MIMIC as dense reference signals: heart rate, mean arterial pressure, respiratory rate, and oxygen saturation. For each ICU stay we extract the timestamps of that patient's complete blood count and chemistry draws, and we subsample the dense vitals at exactly those times. There are three masking conditions: the patient's own lab-draw mask, a mask transplanted from a random patient with matched total count, and uniform random masking at matched rate. Each subsampled series is then processed in several ways: forward-fill, linear interpolation, cubic spline, a Gaussian-process fit, CSDI and LSCD imputation, and IMMFM-style quadratic paths. Because the dense signal is known, we can compute per-patient band-wise power error, spectral slope error, retention of acute events (hypotensive dips, tachycardic runs), and the empirical transfer function. We can then compare that transfer function against the prediction of the gain formula above using each patient's gap lengths.
The second component is a decomposition of prognostic value. For in-hospital mortality we train matched models on four inputs. The dense-truth model sees the full vitals and serves as an oracle. The interpolated-sparse model sees each gap-filled pipeline's output. The raw-points model sees the observed values and times with no filling. The timestamps-only model sees only when measurements happened. The difference between oracle and interpolated-sparse is the information destroyed by interpolation. The performance of the timestamps-only model sets an upper bound on how much of any trajectory feature could be sampling artifact. We report all of these stratified by acuity (SAPS or APACHE tercile), because the claim predicts that the loss concentrates in low-acuity strata.
The third component tests the remedy on real labs, where no dense truth exists. Here we compare interpolated trajectories, raw points, and raw points augmented with the Lomb–Scargle spectrum plus spectral window. The benchmarks are PhysioNet 2012 mortality and the MIMIC-IV cw-24 benchmark. As a clinical case study we revisit NLR in sepsis: does a trajectory representation restore NLR's prognostic value in the sepsis subgroup, and does that restoration survive once we control for draw frequency? We add a synthetic check for labs. We insert albumin-like drops of known depth and duration, following the −3.27 g/L inflection reported by Zhang et al., into smooth real lab trajectories. We then measure how often each pipeline still detects them under own-patient versus random masks.
Baselines
The naive reference is the admission snapshot combined with forward-fill: last observation carried forward, fed to gradient-boosted trees. This is what most clinical risk scores effectively do. Salciccioli et al. report a SAPS I AUC of 0.746 on MIMIC-II, rising to 0.757 with NLR. That is the scale of gain a snapshot biomarker buys. Any trajectory method that does not clearly exceed a snapshot model on the same cohort has not justified itself.
The standard practice baseline is linear interpolation onto an hourly grid, fed to a transformer. We give it the same encoder, the same hyperparameter search budget of 40 trials per method, and the same training compute as every other arm. For the imputation step, the strongest competitor is LSCD, which reports MAE 0.211 against CSDI's 0.219 on PhysioNet 2012 at 10% missingness. We will reproduce that number before using LSCD's output downstream, and we will account for its reported ~45% training overhead in compute-matched comparisons. For continuous-time trajectory modeling, the strongest competitor is Trajectory Flow Matching. It reports an 83% error reduction on eICU sepsis and a 15–20% reduction over NeuralODE and LatentODE elsewhere, but the file gives only relative figures, not absolute ones, so we will rerun it on our splits. For mortality on MIMIC-IV cw-24, the strongest reported number is ALFIA's AUPRC of 0.585 (Adaptive Transformer Layer Fusion), against 0.577 for an AutoGluon ensemble. We will report our models on that exact benchmark. We will also include GRU-D, which ingests missingness masks and time since the last measurement; the cortical-thickness diffusion paper reports it at MAE 0.117 in a different setting.
The key ablation is the timestamps-only model, together with a variant of the remedy that has the spectral window but no values. These isolate how much of any gain comes from sampling information rather than physiology. Beating the baselines means three things. The remedy must match or exceed the best raw-points model in AUPRC with non-overlapping bootstrap intervals. Its gain must be larger in low-acuity strata. And it must retain its advantage after we regress out the timestamps-only prediction.
Why not just…?
Why not just add missingness indicators and time-since-last-measurement, as GRU-D does? These features encode sampling density. That is exactly why they are an important baseline and not a solution to the question posed here. GRU-D mixes the timing signal with the value signal, so a gain from it cannot tell us whether physiology or the clinician's worry is doing the predicting. The timestamps-only ablation and the acuity-stratified analysis separate the two. If GRU-D matches the oracle in the mask-transplant experiment, we will report that simple timing features are enough.
Why not just feed raw irregular points to a continuous-time model and skip interpolation entirely? This is the real threat. ChannelTokenFormer, TFM, and Morelli et al.'s "no interpolation" result all point this way. If the raw-points model closes the oracle gap under own-patient masks, the Lomb–Scargle remedy adds nothing and we will drop it. The contribution would then be the diagnosis: showing that the interpolated pipelines still widely used are biased specifically by acuity, and that trajectory biomarker studies should be re-examined. The mask-transplant experiment settles this directly.
Why not accept that lab biomarkers change slowly, so spectral distortion is irrelevant? For many labs this may be true. That is why the benchmark uses vitals, where dense truth exists, and why the lab component measures detection of acute inflections rather than spectra. If inserted albumin-like drops survive every pipeline equally well under own-patient masks, we will conclude that the distortion matters for vitals and not for labs, and say so.
Why not build Lomb–Scargle-conditioned flow matching, the natural hybrid of LSCD and TFM? With a handful of lab draws the periodogram cannot resolve frequencies reliably, and a learned spectral prior would mostly encode the sampling pattern. That is the confound this proposal is designed to expose. We will run the hybrid only if the spectral-window analysis shows resolvable bands for a substantial fraction of patients.
Decisive Checks
The first check is the own-mask versus random-mask contrast on eICU vitals. The claim fails if interpolation-induced AUPRC loss is statistically indistinguishable between the two conditions at matched missingness, or if the loss is not larger in the low-acuity tercile. The second check is whether the empirical per-patient transfer functions follow the gap-length gain formula. If they do, a patient's gap lengths alone would let us flag, in advance, which patients' trajectory features cannot be trusted. The third check is the sepsis NLR subgroup. If trajectory NLR regains significance in sepsis but loses it after conditioning on the timestamps-only score, the recovered signal was sampling density rather than immune dynamics. That result would bear on the whole trajectory-biomarker literature. As a calibration step before any clinical claim, we will confirm that our pipeline reproduces the qualitative nsr2db finding of Morelli et al.: no interpolation is best for RMSSD at 30–50% missingness.
References
- Spectra and correlations in the solar wind from Voyager 2 around 5 AU
- Towards Robust Real-World Multivariate Time Series Forecasting: A Unified Framework for Dependency, Asynchrony, and Missingness
- Finding Order in Chaos: A Novel Data Augmentation Method for Time Series in Contrastive Learning
- Recognising Activities in Real Time Using Body Worn Passive Sensors With Sparse Data Streams: To Interpolate or Not To Interpolate?
- Analysis of the Impact of Interpolation Methods of Missing RR-intervals Caused by Motion Artifacts on HRV Features Estimations
- Error Estimation of Ultra-Short Heart Rate Variability Parameters: Effect of Missing Data Caused by Motion Artifacts
- LSCD: Lomb--Scargle Conditioned Diffusion for Time series Imputation
- Trajectory Flow Matching with Applications to Clinical Time Series Modelling
- Longitudinal Flow Matching for Trajectory Modeling
- The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study
- Predicting ICU In-Hospital Mortality Using Adaptive Transformer Layer Fusion
- Conditional Score-Based Diffusion Model for Cortical Thickness Trajectory Prediction