When a doctor orders a blood-gas test at 3 a.m., the lab value that comes back is only part of the information. The decision to order the test says something about the patient too: someone thought the patient was getting worse. The long quiet stretch before it, when nobody ordered anything, is also informative. Most models for irregularly sampled time series use the values and ignore the timing. Some go further and assume the timing carries no information about the hidden state, as in the consistency proof for Neural Jump ODEs. This proposal builds a recursive state estimator that treats when each channel is observed as a second, noisy measurement of the hidden state. A recursive state estimator is a filter, like the Kalman filter, that keeps a running mean and uncertainty for the hidden state and updates them each time data arrives. In our estimator, both the arrival of an observation and a period with no arrivals move the belief, alongside the measured values.
The estimator itself is kept deliberately simple, following a warning from the Kalman filtering literature. Optimization or Architecture: How to Hack Kalman Filtering showed that a linear Kalman filter whose noise covariances are trained by backpropagation on the task loss beats neural filters it was previously thought to lose to. It reported 44% lower MSE on Doppler radar, 18% on MOT20 and 10% on lidar. That result was established only on uniform time grids. We carry it into continuous time with arbitrary gaps and asynchronous channels, then add a timing model inside the filter recursion. The main question is whether modeling timing, rather than adding architectural capacity, is what closes the gap on clinical data.
Core Claim
A continuous-discrete Kalman filter is a filter whose hidden state evolves continuously in time between discrete measurement arrivals. We propose one with learned linear drift, learned diffusion and learned per-channel observation noise, run causally one scalar observation at a time. We claim that adding a state-dependent arrival-rate model for each channel, with the timing evidence folded into the filter's update step, gives three things. It gives lower forecast error than timing-blind filters when sampling is informative. It gives better-calibrated uncertainty than neural models that see only observation masks. It collapses back to the timing-blind filter when sampling is uninformative. We expect clinical data such as PhysioNet'12 and MIMIC to be in the first regime and climate data such as USHCN in the last. That contrast is the core test.
Mechanism
The latent state $x(t)$ follows a linear stochastic differential equation, meaning linear dynamics plus random noise, with drift matrix $A$ and diffusion $L L^\top$. Over any gap $\Delta t$, the mean moves by the matrix exponential $e^{A\Delta t}$. The accumulated process noise $Q(\Delta t)$ comes from the standard Van Loan block-exponential construction, so irregular intervals are handled exactly and no ODE solver is needed. When channel $c$ reports value $y$, we apply a scalar Kalman update with readout row $h_c$ and learned noise variance $r_c$. Because each channel's observation is processed on its own, any subset of channels can arrive at any time, and no matrix inversion is needed. This avoids the disconnection problem that GraFITi reports for fully asynchronous channels, because cross-channel coupling lives in the shared state covariance rather than in co-occurring timestamps. As in OKF, all covariances use Cholesky parameterization and all parameters are trained by backpropagating through the filter recursion on predictive negative log-likelihood.
The timing model gives each channel an arrival rate that depends on the state through a log-linear link:
$$\lambda_c(t) = \exp\big(w_c^\top x(t) + b_c\big),$$
where $w_c$ says which directions in the latent state make channel $c$ more likely to be measured. The timing likelihood of a gap from $t_{i-1}$ to $t_i$ that ends with an arrival on channel $c$ has two parts. One is the arrival term, $\lambda_c(t_i)$. The other is a survival term, $\exp(-\sum_{c'} \int_{t_{i-1}}^{t_i} \lambda_{c'}(s)\,ds)$, which is the probability that nothing arrived in between. Under a Gaussian belief with mean $m$ and covariance $P$, the expected rate has the closed form
$$\mathbb{E}[\lambda_c] = \exp\big(w_c^\top m + \tfrac12 w_c^\top P w_c\big),$$
so both terms can be folded into the Gaussian belief with a single Laplace (Newton) step. This is the same local-linearization logic an extended Kalman filter applies to a nonlinear measurement. A long silence pulls the mean toward "stable" directions. An arrival pulls it toward "concerning" directions before the value is even read. At zero cost the model can set $w_c = 0$ and become a plain continuous-discrete filter. That built-in fallback is what makes the uninformative-sampling control meaningful.
Relation to Prior Work
Three lines of work already put a Kalman-type recursion under irregular sampling, and each assumes timing is uninformative. Markovian Gaussian Process VAEs turn Matérn Gaussian-process priors into linear SDEs and run exact Kalman filtering and smoothing on encoder-produced pseudo-observations. We reuse its exact-discretization idea but drop the variational encoder, and we make timing an observation. SDE-VI learns latent distributions at observation times induced by linear time-varying SDEs, but it runs as batch inference with parallel scans rather than as an online filter. NJ-ODE jumps a deterministic neural state at arrivals, carries no covariance, and names state-dependent arrivals as outside its guarantees. That limitation is the gap this proposal targets.
Two papers do model timing, but neither folds it into online state estimation. Latent ODEs add a Poisson arrival rate driven by the latent state, but inside a batch generative decoder, so timing never updates a running belief. Continuous-discrete sensor scheduling treats arrival rates as control variables to be planned against a known model. We instead learn the rates from data as a description of how humans chose to sample, and use them as evidence.
On the neural side, IPN found that an observation-intensity channel alone carries substantial predictive signal. SLAN found pH most predictive of mortality despite being rarely sampled. We read both findings as motivation: timing matters. Our proposal differs in placing timing inside a probabilistic recursion instead of in input features.
The papers that constrain the design are OKF, which forbids crediting architecture before tuning the noise model, and zero-input compensation. The latter shows that neither zeroing nor holding missing measurements is universally better, which argues for modeling the gap explicitly rather than filling it.
Research Program
The work has three stages. First, build the continuous-discrete optimized Kalman filter (CD-OKF) on its own, and establish how far a well-trained linear filter already goes on irregular benchmarks. That answer is worth publishing even if the timing idea fails. Second, add the timing model and study identifiability. With enough latent dimensions, the model could explain timing with directions that do not affect values, so we will measure whether the learned $w_c$ align with value-relevant directions, for example by comparing timing-induced mean shifts to later observed values. Third, test a single nonlinear variant, a small neural readout linearized at the current mean in extended-Kalman-filter style. This checks whether linear latents are the bottleneck once timing is modeled. We keep it as one ablation, not as the main contribution.
Baselines
The naive reference is last-observation-carried-forward per channel, which is hold-input compensation in the language of Zha et al. It is paired with a population-mean prior for channels never observed. Every method must beat it on forecast MSE for the comparison to mean anything.
The strongest classical competitor is the OKF of Greenberg et al. applied to data binned into regular intervals, with bin width tuned per dataset. It is the skeptic's preferred solution and shares our training objective. We give it the same latent dimension and tuning budget. Beating it means lower forecast MSE and better interval coverage at the best bin width, not just at a default one.
The strongest neural filtering competitor is NJ-ODE, which reports medical extrapolation MSE of 1.945 on PhysioNet'12, against 2.208–3.162 for Latent ODE variants. On USHCN it reports 0.40, against 0.43 for GRU-ODE-Bayes; both figures are from Neural Jump ODEs. We will reproduce NJ-ODE under its protocol and treat matching or beating 1.945 and 0.40 as the bar on those splits. We will also run it with observation masks and time-since-last-observation as extra inputs, so it has a fair chance to exploit timing.
For forecasting on MIMIC-IV we include GraFITi, and on PhysioNet'12 we include SDE-VI. GraFITi reports a 17.2% improvement over LinODEnet on MIMIC-IV. We found no reported absolute MSE on Lacuna for either GraFITi on MIMIC-IV or SDE-VI on PhysioNet, so both will be rerun under our shared splits.
For mortality classification, which we treat as secondary, the reference points are mTAN, reporting AUC 0.858 on PhysioNet'12 (mTAN); SLAN, reporting AUPRC 55.20 on PhysioNet'12 (SLAN); and IPN, reporting AUC 0.853 on MIMIC-III (IPN). Our classifier is a logistic head on the filtered state, so we do not expect to beat these by a margin. Staying within their range while also producing forecasts and calibrated intervals would be the claim.
The key ablation is CD-OKF with and without the timing model, identical in everything else. A second ablation replaces the in-filter timing update with the same timing features concatenated to the classifier head. This separates using timing as evidence from using timing as a feature.
All methods get the same splits, the same hyperparameter search budget in GPU-hours, and three seeds.
Why not just…?
Why not just bin the data and run OKF? This would be our own method's closest cousin, and it may well be competitive on dense channels such as heart rate. Its problems are structural. Missingness near 85% on PhysioNet'12, and up to 97.8% on MIMIC as reported by GraFITi, means most bins are empty. Binning also discards within-bin order and exact gap lengths, which is precisely the timing information we claim is useful. The bin-width sweep shows directly whether a fine-enough grid recovers our numbers. If it does, the timing component is the only remaining contribution, and the ablation judges it.
Why not just feed masks and observation intensity into a GRU or attention model, as IPN and mTAN do? This is the real threat. A flexible network given timing features can learn timing effects for classification, and IPN's intensity channel shows it does. Our argument is narrower. For forecasting values between observations, and for calibrated uncertainty, a model that turns timing into an update of the state's mean and covariance should do better with less data. The feature-concatenation ablation, the NJ-ODE-with-masks baseline, and a training-set-size sweep test this. If mask-fed neural models match us on forecast MSE and interval coverage at all data sizes, we will report that timing-as-evidence adds nothing beyond timing-as-feature.
Why not just use zero-input or hold-input compensation in a discrete filter? Zha et al. themselves find neither is universally better, and both are defined on a synchronous index with binary dropouts, not real-valued gaps. We include hold-input as the naive floor. Zero-input compensation is subsumed by the continuous-discrete filter, which simply skips the update when there is no measurement.
Why not just use a more expressive nonlinear latent model such as SDE-VI? Expressiveness and timing are separate axes. SDE-VI is a batch method, so it cannot give the online belief we need. We compare against it offline, and the nonlinear-readout ablation tells us whether capacity, rather than timing, drives any remaining gap.
Proposed Experiments
The decisive synthetic test simulates a latent linear SDE in which each channel's arrival rate grows with the state magnitude, with the strength controlled by a coefficient $\alpha$. When $\alpha = 0$, timing carries no information. As $\alpha$ grows, a timing-blind filter's estimates of the latent state should become increasingly biased, because it misreads silences, and our filter should track the exact posterior. The exact posterior can be approximated here with a large-sample particle filter. At $\alpha = 0$, the two should be indistinguishable and the learned $w_c$ should shrink to zero. If our method helps at $\alpha = 0$, something other than timing is at work, and we would say so.
On real data we run PhysioNet'12 and MIMIC-III/IV extrapolation using the NJ-ODE and GraFITi protocols, and USHCN as a control where station downtime should carry little state information. Metrics are forecast MSE, Gaussian negative log-likelihood, 90% interval coverage, mortality AUPRC and AUC from the filtered state, and per-observation update latency.
Two checks are decisive. The first is a timing-shuffle test. For each patient, arrival times are redistributed uniformly over the stay while values and counts are kept; the model is retrained on the shuffled data and evaluated on the shuffled test split. The timing model's advantage should disappear on shuffled clinical data, while the plain filter's performance should barely move. If it does not disappear, the gain is not coming from informative timing. The second is the cross-dataset contrast. The timing model should improve forecasting on PhysioNet and MIMIC and do essentially nothing on USHCN. A uniform gain everywhere would point to extra capacity rather than timing, and would count against the core claim.
References
- Neural Jump Ordinary Differential Equations: Consistent Continuous-Time Prediction and Filtering
- Optimization or Architecture: How to Hack Kalman Filtering
- GraFITi: Graphs for Forecasting Irregularly Sampled Time Series
- Markovian Gaussian Process Variational Autoencoders
- Generative Modeling of Irregular Time Series via SDE-Induced Continuous-Discrete Variational Inference
- Latent ODEs for Irregularly-Sampled Time Series
- Optimal Sensor Scheduling and Selection for Continuous-Discrete Kalman Filtering with Auxiliary Dynamics
- Interpolation-Prediction Networks for Irregularly Sampled Time Series
- No Imputation Needed: A Switch Approach to Irregularly Sampled Time Series
- Zero-Input Compensation Mechanism-Based Linear Estimation for Systems with Multiple Packet Dropouts and Multiplicative Noises
- Optimization or Architecture: How to Hack Kalman Filtering
- Multi-Time Attention Networks for Irregularly Sampled Time Series