Separating the Body Clock from the Ward Clock in Short ICU Stays

Patients in intensive care live on two clocks at once. One is the ward clock: nursing rounds, feeding, medication and light schedules that repeat every 24.0 hours. The other is the body clock: the patient's own circadian rhythm, which illness, sedation and sepsis can push to run fast or slow, for example to 21 or 27 hours. Both clocks leave a daily oscillation in heart rate, temperature, blood pressure and respiratory rate. A typical stay lasts only 48–72 hours, and over a window that short, ordinary Fourier analysis cannot tell a 24-hour rhythm from a 27-hour one. The two show up as a single blurred peak. So it is not known whether "the patient's rhythm has drifted away from the ward's rhythm" can be measured in the ICU, or whether it predicts how the patient will do.

We propose to measure it. We treat a patient's hourly vital signs as several noisy views of the same few rhythms. Each vital sign shares the rhythms' periods but has its own strength and timing. We then use low-rank Hankel recovery (explained below) to separate rhythms closer together than the classical resolution limit, while also removing the isolated spikes caused by interventions and sensor artifacts. The note asks two questions, in order. First, can this separation be done reliably on real ICU data, checked against long stays where the answer is known? Second, does the resulting "rhythm drift" carry information about mortality and sepsis that existing features and deep sequence models do not?

Core Claim

The classical resolution limit says that two frequencies closer than about $1/T$ cannot be separated from a record of length $T$. For $T = 48$ hours, $1/T \approx 0.021$ cycles per hour. The gap between a 24-hour and a 27-hour rhythm is about $0.0046$ cycles per hour, roughly a fifth of that limit.

Two results suggest the limit can be beaten here.

ICU vitals fit this setting. Heart rate, temperature, mean arterial pressure, respiratory rate and oxygen saturation are five or more simultaneous views of the same daily rhythms.

The claim is therefore specific. A multichannel Hankel estimator with a spike-removal term can recover a two-line rhythm structure (a 24.0-hour ward line plus a free-running endogenous line) from 48–72 hours of hourly vitals. The gap between the two lines is a new, interpretable biomarker. To our knowledge, no paper in the literature we reviewed has attempted this.

What Exists and What Differs

The methods we build on were all validated away from clinical data.

- Separation-Free Super-Resolution works on synthetic sinusoids. It reports relative error near $10^{-9}$ at separation 0.0001, but only without noise. It also warns that additive noise makes the Hankel null space ill-conditioned. That warning constrains our design: we cannot rely on the noiseless result, and must gain stability from multiple channels. - Vandermonde Factorization of Hankel Matrix for Complex Exponential Signal Recovery supplies the treatment of missing samples. It factorizes the Hankel matrix to complete non-uniformly sampled, damped exponentials in NMR spectroscopy at 22% sampling. We reuse this idea to handle missing hours without interpolating them.

  • Demixing Sines and Spikes Using Multiple Measurement Vectors supplies the spike model. It separates spectral lines from sparse, large outliers shared across snapshots, which matches a vasopressor bolus or a transducer flush. It was tested only on synthetic random-phase sinusoids, and we replace its atomic-norm formulation with a Hankel one. - Block Hankel Tensor ARIMA also builds Hankel structures from many short series. It uses them to forecast and never estimates frequencies. We use the structure to estimate frequencies and do not forecast.

On the clinical side, the motivation comes from work showing that trajectories beat snapshots. The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness and Association between albumin changes and prognosis in older patients with acute myocardial infarction both find that single admission values miss physiological transitions. These papers track levels and slopes, not rhythms.

The continuous-time deep models are the main competitors for prediction. Neural Controlled Differential Equations for Irregular Time Series and ContiFormer handle irregular sampling well. However, Neural CDEs build their input path with cubic splines, and spline interpolation acts as an uncontrolled low-pass filter that distorts oscillatory structure (a concern discussed around Cubic NK-SVD). We treat that distortion as something to measure, not assume.

Finally, Fourier-based neural forecasters such as FITS explicitly break down on short look-back windows. That is exactly the regime here, and it is why a discrete-FFT neural model is not our starting point.

Research Program

Data. For each patient we take the first $T$ hours of hourly-aggregated vitals, using the MIMIC-Extract pipeline (MIMIC-Extract). We remove a slow trend per channel and standardize.

Model. Let $Y \in \mathbb{R}^{C \times T}$ hold $C$ vital-sign channels, with observed entries indexed by $\Omega$. We stack each channel's Hankel matrix side by side into a block Hankel matrix $\mathcal{H}(X)$. If the channels share $R$ rhythms, its rank is $2R$, because each real sinusoid contributes two complex exponentials. We solve

$$ \min_{X,\,S}\ \|\mathcal{H}(X)\|_* + \lambda \sum_{t} \|S_{:,t}\|_2 + \frac{\mu}{2}\,\|P_\Omega(Y - X - S)\|_F^2 , $$

where:

  • $X$ is the rhythmic part;
  • $S$ is a spike matrix whose columns (hours) are mostly zero. The group penalty encodes that an intervention perturbs several vitals in the same hour;
  • $\|\cdot\|_*$ is the nuclear norm, the sum of singular values, which encourages low rank;
  • $P_\Omega$ keeps only observed entries, so missing hours are completed rather than interpolated;
  • $\lambda$ and $\mu$ trade off spike sparsity and data fit.

Cost. At $T = 72$ and $C = 6$, the block Hankel matrix is about $36 \times 222$. Its singular value decomposition takes microseconds. The computational burden that limits these methods on long signals is therefore irrelevant at ICU scale, and a full cohort runs on a laptop.

From matrix to frequencies. We estimate the frequencies from the recovered low-rank $X$ using ESPRIT, and select the model order by the singular-value gap. We then refine frequencies, damping and amplitudes by nonlinear least squares, starting from the ESPRIT estimates.

Clinical features. Per patient we record:

  • whether a second near-daily line is present;
  • its period, $P_{\text{endo}}$;
  • the drift $|1/P_{\text{endo}} - 1/24|$;
  • per-channel amplitudes and phases;
  • spike density.

Analysis in three stages.

  1. Semi-synthetic identifiability. We inject two-line rhythms with known periods into real residual noise. The noise comes from patients' de-rhythmed vitals with real missingness masks. We map error and detection power as functions of separation (in units of $1/T$), $T$, and number of channels, and compare against the Cramér–Rao bound, the best variance any unbiased estimator can achieve.
  2. Long-stay consistency. We take MIMIC-III stays of 8–10 days, the longest MIMIC-Extract keeps. Over 240 hours the resolution limit is about 0.0042 cycles per hour, so a 24 vs. 27-hour pair is roughly resolvable with classical methods. We estimate the two lines on the full stay, then ask whether each estimator recovers the same answer from only the first 48 or 72 hours. This is the only real-data reference available. We acknowledge that rhythms change over ten days, so agreement is an imperfect but informative check.
  3. Clinical value. We test whether the rhythm features improve prediction beyond strong baselines.

Baselines

Naive reference. For frequency recovery, the naive estimator is the peak of a zero-padded periodogram. Zero-padding only interpolates the spectrum on a finer grid; it does not add resolution, as noted in Data-driven Estimation of Sinusoid Frequencies. We expect it to report one merged line, and that expectation is itself worth confirming on real data. For prediction, the naive reference is a gradient-boosted model on the admission and summary features of MIMIC-Extract with no rhythm information.

Single-line and two-line cosinor. The strongest simple spectral competitor is cosinor regression, the standard tool in chronobiology. It fits one sinusoid of free period by nonlinear least squares. For a single line at decent signal-to-noise ratio, this is already precise well below $1/T$. We therefore also run a two-sinusoid nonlinear least-squares fit, with the model order chosen by the Bayesian information criterion and many random restarts. To keep the comparison fair, it gets the same detrended inputs, the same masks, and a restart budget matched in compute to our method. Beating it means one of two things: higher detection power for the second line at separations of 0.2–0.5 of $1/T$ at equal false-positive rate on null injections, or lower frequency error closer to the Cramér–Rao bound.

Classical subspace methods. Per-channel ESPRIT, and multichannel ESPRIT on the sample covariance, both use model order from minimum description length. They test whether the convex Hankel step adds anything over subspace methods, which become unstable below the separation limit under noise (Stability and Super-resolution of MUSIC and ESPRIT).

Prediction baselines. The strongest learned competitors are Neural CDE and ContiFormer trained on the same hourly vitals and windows. Neural CDE reports an AUC of 0.880 on its PhysioNet sepsis task (Neural Controlled Differential Equations for Irregular Time Series). On MIMIC-III, MIMIC-Extract reports mortality AUROCs between 0.85 and 0.90, with AUPRC often below 0.50. On the PhysioNet 2019 challenge, winning utility scores were between 0.33 and 0.36 (Early Prediction of Sepsis from Clinical Data). We will rerun these models ourselves under one split and one tuning budget instead of comparing to published numbers, because their windows and cohorts differ from ours.

Beating these baselines means that adding the rhythm features to the deep model, or to gradient boosting, raises AUPRC with a bootstrap confidence interval excluding zero. We cannot claim the features improve on a deep model that already sees the raw series unless that gain appears.

For frequency recovery on real ICU vitals, we found no reported result on Lacuna for any method.

Ablations. Each ablation isolates one ingredient:

  • single-channel instead of multichannel Hankel, which tests the multiple-snapshot gain;
  • no spike term;
  • cubic-spline-filled input instead of masked completion, which directly measures how much the spline preprocessing used by Neural CDE distorts rhythm estimates;
  • a "24.0-hour line only" model, which tests whether the second line is real or an artifact of fitting.

Why not just…?

Why not just fit a cosinor with a free period? For one line it works, and we use it as the single-line baseline. The question here is whether two lines are present, and a free-period single sinusoid fit to a 24.0 + 27-hour mixture returns a meaningless average period. Two-sinusoid least squares is the real threat. It is close to maximum likelihood, and with good starting points it can approach the Cramér–Rao bound. Our expected advantages are narrower:

  • the Hankel step gives model order and starting points without restarts;
  • it shares frequencies across channels;
  • it removes spikes that would otherwise drag a least-squares fit.

The semi-synthetic power curves settle this. If two-line least squares with BIC matches our detection power and error, we will report that the simple estimator suffices. In that case the contribution becomes the long-stay validation and the clinical finding, not the estimator.

Why not just give hourly vitals to a deep sequence model and let it learn the rhythm? Neural CDE and ContiFormer may already absorb daily structure implicitly. The incremental-value test answers this directly: rhythm features are added to the deep model, not compared against it alone. There are two further reasons for explicit features. First, the deep models' spline paths may suppress oscillations, and the spline ablation will show whether they do. Second, a period is a unitless, level-free quantity. It may transfer across hospitals better than raw values, whose distributions shift sharply between centers (Early Prediction of Sepsis from Clinical Data). We test this by training on PhysioNet hospital A and evaluating on hospital B.

Why not just use day–night differences or lag-24 autocorrelation? These capture rhythm amplitude cheaply, and they enter every predictive comparison as extra baseline features. They cannot express drift away from 24 hours, which is the variable we claim is new. If drift adds nothing beyond amplitude, then the case for super-resolution in this setting fails, and we will say so.

Decisive Checks

The project succeeds or fails on three results.

  1. Identifiability. On semi-synthetic injections with real noise, the multichannel estimator must detect a second line at a separation of $0.25/T$ with $T = 72$ hours, at a substantially higher rate than two-line least squares and ESPRIT, at a matched 5% false-positive rate.
  2. Long-stay agreement. On 8–10 day stays, the drift estimated from the first 72 hours must agree with the full-stay estimate better than all baselines do. If no estimator achieves meaningful agreement, the biomarker is not measurable and the clinical analysis stops there.
  3. Clinical value. Drift must improve mortality AUPRC on MIMIC-III and sepsis utility on PhysioNet 2019 beyond amplitude features and beyond Neural CDE and ContiFormer. It must also hold up under cross-hospital transfer.

A clean negative at check 1 or 2 is still useful: it would establish that ICU stays are too short and too noisy for rhythm-drift claims, which bounds what any rhythm-based ICU biomarker can deliver.