Many clinical measurements rise and fall on a roughly 24-hour cycle: blood pressure, heart rate, inflammatory markers, lymphocyte counts. When we track a patient over time and ask "is this getting worse?", that daily swing can look like decline or recovery. The usual fixes are to subtract a fixed daily curve or to measure at the same clock time every day. Both assume the patient's internal clock is stable. In the patients we most want to monitor, it often is not. Critical illness, stress, late eating and sleep loss shift the timing and flatten the size of these rhythms. A same-time-every-day measurement then samples a different point of a moving wave each day, and the drift in clock timing shows up as a false trend.
We propose a tracker that treats the daily rhythm's timing and size as quantities that drift and are estimated, kept separate from the slow disease trend. We pair it with a sampling policy in which the analysis side, not the sensor, decides when the next measurement is worth taking. That policy borrows from receiver-driven network transport, where the destination hands out permission "credits" instead of letting senders push data blindly. In our version, a fixed measurement budget shared across a ward of patients is split by issuing credits to the patients and time points where a new reading would most reduce uncertainty about the disease trend. The tracker is tested on synthetic data with known ground truth and on intensive-care vital signs. We also run one cellular-scale check: whether the inflammatory-to-scarring fibroblast transition seen in a human wound-healing atlas is confounded by the unrecorded time of day at which tissue was sampled.
Core Claim
The claim has two parts. First, when the daily rhythm's timing drifts, any trend estimator that assumes a fixed rhythm, or that samples at a fixed clock time, makes trend errors that grow with the drift rate. A model that estimates timing and size as slowly varying hidden states removes most of that error. Second, under a fixed sampling budget, choosing when to measure matters as much as the model. Spending measurements to pin down the rhythm's current timing, only when that timing is uncertain, beats both uniform and fixed-clock-time schedules at estimating how fast the patient is changing.
Written out, a biomarker reading at time $t$ is modeled as
$$y(t) = s(t) + A(t)\cos\!\left(\frac{2\pi t}{\tau} - \phi(t)\right) + \varepsilon(t),$$
where $s(t)$ is the slow disease trend whose rate of change $\dot s(t)$ is the quantity clinicians care about, $A(t)$ is the rhythm's size, $\phi(t)$ is its timing (phase), $\tau$ is a period near 24 hours, and $\varepsilon(t)$ is measurement noise. $A$ and $\phi$ follow slow random walks, so a flattened or shifted rhythm is something the model expects rather than an anomaly. The model is a state-space model, meaning hidden quantities that evolve over time and are updated by a filter as each reading arrives. Its posterior variance on $\dot s$ drives the sampling policy.
Where This Sits in Prior Work
Two lines of work motivate the gap. Longitudinal biomarker studies show that trajectories beat snapshots. In-hospital albumin trajectories in myocardial infarction predict mortality better than admission values (Zhang et al.). Long-run hsCRP trajectories separate risk among people with identical baselines (Pan et al.). DiPro separates static anatomy from evolving pathology across ICU vitals and chest X-rays (DiPro). None of these models a daily oscillation, so any rhythm the data contain is either absorbed into the trend or left to a network to learn implicitly. Circadian research shows the rhythm is not fixed. Peripheral tissues run their own clocks (Pulimeno et al.; Petrenko et al.). Meal timing, stress and sleep loss shift those clocks away from the brain's master clock (Lu et al.; corneal clock under restraint stress; lacrimal gland study). In UK Biobank accelerometry, routine consistency predicts mental-health outcomes beyond sleep duration (Moebus and Holz), which suggests that day-to-day rhythm instability is itself informative.
The closest methodological ancestor is MOSAIC. It fits linear, exponential and damped-harmonic (ECHO) models jointly across omics layers, sharing a period across modalities so that noisy channels can recover rhythms. MOSAIC supplies the idea of fitting rhythm and non-rhythm components together, and we reuse it. It differs from our proposal in three ways. It fits a fixed-parameter curve to a whole time series rather than tracking a drifting phase online. It was run on Neurospora and synthetic data, not on clinical streams. It takes the sampling schedule as given.
The sampling side reuses an idea from receiver-driven transport. In ExpressPass, receivers issue small credit packets and senders transmit only on credit (ExpressPass). In Homa, receivers issue grants ordered by remaining message size (Homa). AMUSE shows that a receiver can throttle senders it does not control by rewriting the advertised window (AMUSE). We take only the allocation principle from these systems: the party that knows what it needs decides who sends next, and a shared capacity is divided by explicit grants. Their datacenter results, such as microsecond tail latencies and convergence within 3 round trips, do not transfer to clinical sampling, and we make no claim that they do. The scarce resource here is nurse-drawn samples, patient burden and telemetry slots, not switch buffers.
The fibroblast check is motivated by the cross-tissue atlas (fibroblast atlas). It found F6 inflammatory myofibroblasts peaking around day 7 after wounding and being replaced by F7 contractile myofibroblasts by day 30. A separate review ties loss of BMAL1 rhythm to TGF-β-driven fibrosis (Chen et al.), but treats fibrosis as one undifferentiated endpoint. Neither asks whether biopsy time-of-day confounds the subtype proportions on which these trajectories rest.
Research Program
We will build three things. The first is the tracker: a filter over $(s, \dot s, A, \phi)$ with a near-24-hour period. A particle or unscented filter is needed because phase enters the model nonlinearly. Clock-time covariates such as meals, lights and sedation can optionally nudge $\phi$.
The second is the credit scheduler. At each scheduling interval, a ward-level budget of $B$ measurements is divided among patients. Each patient's bid is the expected reduction in posterior variance of $\dot s$ from one reading at each candidate time, weighted by the posterior probability that $\dot s$ crosses a clinical threshold. Credits go to the highest bids, with a minimum per-patient floor to prevent starvation. This is the analog of the max-min fairness concern raised for credit schemes at scale.
The third is an analysis of when phase-awareness matters. We will characterize trend error as a function of drift rate in $\phi$ and amplitude loss in $A$, and identify the regime where simpler methods are adequate. That regime is a deliverable, not something to hide.
Baselines
The naive reference is a raw-trend estimate: a local linear regression of the biomarker on time with no rhythm term, using all available readings. Every method must beat it, and it also shows how much daily rhythm inflates false trend alarms.
The strongest simple competitor is cosinor detrending: fit a fixed 24-hour sine with constant amplitude and phase per patient, subtract it, and estimate the trend from the residual. A MOSAIC-style joint fit with a shared period and a linear or exponential trend term is the more capable version of this, and we include it as well (MOSAIC). No reported result was found on Lacuna for cosinor or MOSAIC on ICU vital-sign trend estimation, so we will run both ourselves on identical patient windows.
The strongest existing learned method is DiPro on MIMIC-IV/MIMIC-CXR. The file reports an AUPRC of 0.742 for in-hospital mortality and 0.440 length-of-stay accuracy (DiPro). We will compare against it on the same splits by adding our phase-separated trend features as inputs, with an equal tuning budget for both variants. Exceeding 0.742 AUPRC on mortality, with a confidence interval from repeated seeds that excludes the baseline, would count as beating it. Our main metrics are trend error and false-alarm rate, which DiPro does not report, so on those metrics DiPro's learned representation is probed by a linear readout of $\dot s$.
For sampling, the baselines are uniform-interval sampling, fixed-clock-time sampling (the clinical default), and a single-patient uncertainty-driven schedule without ward-level partitioning. All receive the same total budget $B$. The key ablations are the full model with phase frozen after day one, which isolates the drifting-phase ingredient, and the full model with uniform sampling, which isolates the scheduler.
Why not just…?
Why not just subtract a fixed daily curve (cosinor) and track the residual? This works when the rhythm is stable, and we expect it to tie our method in that regime. The circadian literature in the file suggests stability is exactly what fails in ill, stressed or late-eating patients (Lu et al.; Chen et al.). A fixed phase then leaves a residual oscillation that beats against the true phase and produces spurious slopes. The drift-sweep experiment below measures the drift rate at which cosinor stops being adequate. If ICU patients' estimated drift falls below that threshold, the honest conclusion is that cosinor suffices, and we will report it.
Why not just measure at the same clock time every day? This is the most serious threat because it costs nothing. With a stable phase it cancels the rhythm exactly. With phase drifting at $\dot\phi$ radians per day, a fixed-time reading picks up a spurious daily change of roughly $A\,\dot\phi\,\sin(\cdot)$, which is first-order in the drift. A 1–2 hour shift over a week therefore produces a slope comparable in size to many clinically meaningful trends. The comparison under equal budget on synthetic and subsampled MIMIC data settles this directly. If fixed-time sampling matches us on real ICU data, the scheduler contribution is void.
Why not just let a sequence model like DiPro learn the rhythm implicitly? A large model can in principle represent a drifting oscillation. However, DiPro's disentanglement separates static anatomy from dynamic change and has no oscillatory prior (DiPro), and its outputs are not calibrated estimates of $\dot s$ that a scheduler can act on. The feature-augmentation test shows whether explicit phase separation adds information beyond what DiPro already captures.
Why not just use generic Bayesian experimental design or standard priority queues instead of credits? The single-patient uncertainty-driven schedule is generic experimental design, and it appears as a baseline. The credit framing adds only cross-patient partitioning of a shared budget. Static priority classes, the DiffServ-style alternative, cannot reorder patients as their risk changes, which mirrors the preemption problem Homa reports for static priorities (Homa). If ward-level partitioning gives no gain over independent per-patient schedules, we drop it.
Proposed Experiments
The synthetic study generates trajectories from the model above, with known $\dot s$, phase drift rates from zero to several hours per week, amplitude loss from none to complete flattening, and sampling budgets of 2, 4 and 6 readings per day. These budgets echo MOSAIC's 2–6 hour sampling grid. Metrics are mean absolute error on $\dot s$ and the false-alarm rate for threshold crossing at matched sensitivity. The decisive check is whether the gap between our method and the cosinor and fixed-clock-time baselines is near zero at zero drift and grows with drift. If it does not grow, the mechanism is wrong.
The ICU study uses hourly heart rate and blood pressure from MIMIC-IV. Full-resolution data serve as the reference trend. We then subsample to a ward budget and compare schedules on trend error, deterioration-alarm precision, and lead time before documented deterioration. Mortality AUPRC is evaluated against the DiPro figure of 0.742 as described above. We also report the distribution of estimated phase drift across patients, since that number determines whether the approach matters in practice.
The cellular check reanalyzes the 58,823-cell wound-healing scRNA-seq dataset and the broader atlas (fibroblast atlas). We infer each sample's circadian phase from core clock genes (BMAL1, PER1-3, CRY1-2) within fibroblasts, test whether F6 and F7 subsets differ in clock-gene phase, and test whether F6 and F7 proportions vary with inferred sample phase after conditioning on wound day. A null result supports the atlas trajectories as clock-robust. A strong phase association would mean that snapshot atlases mix time-of-day effects into state transitions. That finding would motivate time-stamped sampling in future longitudinal tissue studies and would be a direct cellular instance of the same confound the tracker addresses.
References
- 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
- Multimodal Disease Progression Modeling via Spatiotemporal Disentanglement and Multiscale Alignment
- Autonomous and self-sustained circadian oscillators displayed in human islet cells
- Pancreatic α- and β-cellular clocks have distinct molecular properties and impact on islet hormone secretion and gene expression
- A late eating midpoint is associated with increased risk of diabetic kidney disease: a cross-sectional study based on NHANES 2013–2020
- Chronic Restraint Stress Misaligns Corneal Clock via β-Adrenergic and Glucocorticoid Signaling, Altering Epithelial, Neural, Immune, Metabolic States
- Circadian disruption and ROS-NLRP3 signaling mediate sleep deprivation-enhanced silica nanoparticle toxicity in lacrimal glands
- Beyond the hours slept: inconsistent sleep routines threaten mental health in 100,000 UK Biobank participants
- MOSAIC: A Joint Modeling Methodology for Combined Circadian and Non-Circadian Analysis of Multi-Omics Data
- Credit-Scheduled Delay-Bounded Congestion Control for Datacenters
- Homa: a receiver-driven low-latency transport protocol using network priorities
- AMUSE: Empowering users for cost-aware offloading with throughput-delay tradeoffs
- A single-cell and spatial genomics atlas of human skin fibroblasts reveals shared disease-related fibroblast subtypes across tissues
- Circadian clocks and their role in kidney and eye diseases across organ systems