In intensive care data, when a lab test is taken carries clinical meaning. Frequent lactate draws suggest a worried clinician, and a skipped test often means the patient is stable. Models for irregularly sampled records exploit this signal, and they gain a lot from it. The trouble is that each hospital also has its own measurement habits: one draws labs every four hours by protocol, another only when a value is suspicious. A model trained at one hospital learns that hospital's habits as if they were physiology, then misreads the schedule at a new site. Standard domain adaptation handles this badly in both directions. It either leaves the habit signal in place, or, when a discriminator forces features to look alike across hospitals, it erases the informative part of the schedule together with the institutional part.
We propose to separate the two, using a fact that makes this problem easier than it looks: the measurement schedule is fully observed in unlabeled data. We fit a model of each hospital's measurement policy, meaning the rate at which each variable gets measured given the patient's current estimated state. This model can be fit on the new hospital's records with no outcome labels at all. The predictor never sees raw measurement times. It sees how much more or less often a variable was measured than this hospital would normally measure a patient in this state. That surprise signal should carry the same clinical meaning everywhere. A light adversarial term then removes whatever site identity is left in the value-based state trajectory.
Core Claim
Cross-site failure of irregular-time-series models is driven mainly by site-specific, state-dependent measurement policies. Because these policies can be learned without labels, the right unsupervised fix is to condition on the site's policy rather than to align it away. Concretely, let $h(t)$ be a continuous-time patient state inferred from observed values, let $N_c(t)$ count the measurements of variable $c$ up to time $t$, and let $\lambda^s_c(t \mid h(t))$ be the measurement rate of variable $c$ at site $s$. The predictor receives the measurement residual
$$r^s_c(t) = N_c(t) - \int_0^t \lambda^s_c(u \mid h(u))\,du,$$
which is the number of measurements beyond what site $s$ would expect for this trajectory. In point-process theory this quantity is the martingale residual: it has mean zero under the correct site policy. Its meaning is therefore "clinician attention beyond local protocol," rather than "local protocol." The rate is parameterized as $\lambda^s_c = \exp(a_c(h) + b^s_c(h))$. The term $a_c$ is shared across sites. The term $b^s_c$ is site-specific and small, and at the target hospital it is fit only on measurement timestamps.
The claim is falsifiable in a specific way. If residual conditioning helps only by normalizing marginal measurement frequency, then a per-site rescaling of measurement counts should match it. We expect the gains to come from the state-dependent part of the policy. A semi-synthetic experiment described below isolates this directly.
Relation to Prior Work
Two papers motivate the problem. Adapt the Data, Not the Model found that transfer from MIMIC-IV to eICU and HiRID is dominated by cadence differences between hospitals, and that forward-filling turns those differences into step artifacts. That paper fixes the problem by learning an input-space mapping through a frozen predictor, which requires target labels. It also found that successful adaptation increased distributional distance from the source. That finding directly constrains our design: it warns against pure distribution alignment. We respond by aligning only the value-derived state and conditioning on the schedule rather than matching it. Domain Adaptation under Missingness Shift proves the target predictor is identifiable when only recording rates change. Its result holds for static, uninformative ("completely at random") underreporting, and its dropping filter works only when the target is sparser than the source. We treat the informative, temporal case it explicitly leaves open.
The architectural ingredients come from single-domain work on irregular sampling. GRU-ODE-Bayes evolves a hidden state continuously between observations and updates it at each measurement. Neural ODEs for Informative Missingness learns continuous decay as a function of elapsed time and reports large gains on PhysioNet 2012 from exploiting informative gaps. Temporal Belief Memory decays confidence in old observations over time. All three treat the schedule as a single-population signal, and none examines what happens when the schedule's meaning changes between sites. We reuse a continuous-time encoder of this kind for $h(t)$.
On the adaptation side, AdvSKM aligns time-series features adversarially with a learned spectral kernel. CALDA combines multi-source adversarial alignment with a contrastive loss that pulls same-class examples together, and it names unevenly spaced series as an open limitation. Both assume regular grids. We borrow the gradient-reversal adversary, a layer that flips gradients so the encoder learns to fool a site classifier (as in Yu et al.), but we apply it only to the value-derived state. KnowRare is the closest existing combination of adversarial adaptation and sparse ICU data. It aligns latent representations of an LSTM on discretized records across disease conditions, and it does not ask whether what it aligns is physiology or measurement frequency. Its "data-volume paradox," where pooling sources hurts, is a second constraint, and it is why we model sites separately rather than pooling them.
Research Program
The system has three parts. The first is a continuous-time state encoder in the style of GRU-ODE-Bayes, which consumes observed values and their times and produces $h(t)$. The second is a multivariate point-process head, which models per-variable measurement rates $\lambda^s_c$ using the shared-plus-site decomposition above. The third is a predictor that reads $h(t)$ together with the residual trajectory $r^s(t)$. Training at the source maximizes outcome likelihood plus measurement-time likelihood. At the target site, we fit only $b^{\text{target}}_c$ (plus optional batch statistics) by maximizing the likelihood of the target's measurement timestamps, with the encoder frozen. A gradient-reversal site classifier on $h(t)$ is trained on source and unlabeled target data together. Its purpose is to strip value-level site fingerprints, such as unit conventions and device offsets, and it is kept weak, with its weight tuned on source-only validation.
Beyond building this, we will analyze three things. The first is how much cross-site degradation of existing irregular-time-series models is attributable to schedule shift versus value shift. We measure this by swapping schedules between sites while holding values fixed, which is possible in the semi-synthetic setting. The second is whether the learned site-specific rate terms are clinically legible, for example whether they recover known protocol differences between HiRID's high-resolution monitoring and eICU hospitals. The third is when the method should fail. It will fail if a site's policy depends on information the values never reveal, since the residual then mixes policy with patient state.
Baselines
The naive reference is a source-trained continuous-time model applied unchanged to the target hospital, using the same encoder as ours with raw measurement indicators. This measures the size of the gap. We also report the same pipeline on forward-filled hourly grids, following the preprocessing in Yet Another ICU Benchmark, because that is what practitioners actually deploy.
The strongest unsupervised competitor is the obvious combination: domain-adversarial training with gradient reversal applied to the full representation of an informative-missingness encoder. We instantiate this with Extended ODE-GRU-D from Neural ODEs for Informative Missingness, which reports AUC 0.9147 on PhysioNet 2012 mortality in the single-domain setting, and alternatively with GRU-ODE-Bayes. We also run CALDA's contrastive-adversarial objective on gridded inputs for the multi-source eICU setting. The labeled reference point is INPUTADAPTER, given the same target labels it requires. It reports +14.1 AUCPR points for acute kidney injury prediction from MIMIC-IV to eICU (Adapt the Data, Not the Model). We treat this as a ceiling to approach, not a method to beat, since it uses labels ours does not. On the benchmark, YAIB reports within-dataset mortality AUROC of 0.84 to 0.87 across its databases. We found no reported result on Lacuna for unsupervised cross-site transfer of continuous-time irregular-series models on these databases, so we will establish those numbers ourselves.
Fairness is enforced as follows. All methods share the encoder backbone, the patient splits, and the cohort definitions from YAIB. Each gets the same hyperparameter search budget, with model selection on source-only validation so that no method peeks at target labels. Target-side fitting for every unsupervised method uses the same unlabeled target records.
The ablations isolate the key ingredient. We remove the residual and feed raw indicators. We replace the state-conditional policy with per-site marginal rates, $\lambda^s_c$ constant in $h$. We drop the target refit of $b^s$. We drop the adversary. We count the method as beating the obvious combination if it improves target AUROC and AUPRC averaged over transfer pairs, with confidence intervals across seeds excluding zero. It should also retain source-site performance within a small margin, because gains bought by destroying informative missingness would show up as a drop at the source.
Why not just…?
Why not just throw away the measurement times and use only values? This removes schedule shift trivially, but it also removes the signal that single-domain work shows is valuable. Extended ODE-GRU-D reaches 0.9147 AUC against 0.824 for standard GRU-D on PhysioNet 2012, and Temporal Belief Memory's time-aware decay reaches AUC 0.7840 against 0.7078 for mean substitution on septic shock. A values-only encoder is included as a baseline. If it matches our method at the target, that would show the informative part of the schedule does not transfer, which would itself be a useful negative result.
Why not just run domain-adversarial training on an existing irregular-series encoder? This is the obvious combination and the most serious competitor. Our argument against it is mechanistic. At a well-trained optimum, the discriminator can only be fooled by removing site-predictive content, and measurement frequency is strongly site-predictive. The adversary is therefore pushed to erase the informative schedule signal. The INPUTADAPTER finding, that good adaptation moved data away from the source distribution, is consistent with this. The decisive check is the semi-synthetic experiment below. If full-representation adversarial training matches residual conditioning there, the core claim is wrong.
Why not just normalize measurement counts per hospital? This is the cheap, non-adversarial version of our idea, and it is a real threat if hospitals differ mainly in overall measurement rate. It fails by construction when policies differ in which states trigger measurement. Suppose one site tests lactate on a fixed schedule and another only after hypotension. Then equal counts mean different things at the two sites, and no marginal rescaling fixes that. The per-site-marginal-rate ablation is exactly this alternative. We will report honestly if it recovers most of the gain on real data, in which case the practical recommendation becomes the simpler method.
Why not just apply the missingness-shift filter from DAMS? Its guarantees hold only when recording is uninformative, it only handles targets sparser than the source, and it ignores time. HiRID is denser than most sources, so the filter does not apply to transfers into it. We include the filter, applied to gridded features, on transfer pairs where the target is sparser.
Proposed Experiments
The first experiment is semi-synthetic and is the decisive one. We take PhysioNet 2012 and keep the real physiological values, but resample which measurements are "observed" under two synthetic hospital policies. Policy A measures on a fixed cadence. Policy B measures with a rate that rises with abnormal recent values plus a hidden severity signal tied to the outcome. We then vary how much the policies differ in marginal rate versus in state dependence. The ground truth is known here, so we can check three predictions. Residual conditioning should hold target AUROC near the oracle trained on the target policy. Full-representation adversarial training should lose the hidden-severity signal, which we can see as a drop at both source and target. Per-site marginal normalization should close the gap only when the policies differ in rate alone.
The second experiment uses real multi-center data through YAIB: MIMIC-III or MIMIC-IV, eICU, HiRID, and AUMCdb, on in-hospital mortality, acute kidney injury, and sepsis onset. We evaluate every ordered transfer pair between databases, plus leave-one-hospital-out within eICU's roughly 200 hospitals. The eICU setting gives many small, genuinely different target sites, which is where learning a target policy from timestamps alone is most tested. AUMCdb's higher mortality prevalence inflates AUPRC, so we report AUPRC gains relative to each target's no-adaptation baseline. We add calibration error, because a residual computed from a misfit site policy should show up as miscalibration before it shows up in ranking. We also track how many unlabeled target stays are needed to fit $b^{\text{target}}$, since that determines deployability.
The final diagnostic tests whether the method does what it claims. It trains a probe to predict site from the predictor's inputs. The probe should succeed on raw indicators, fail on residuals, and fail on $h(t)$ after adversarial training. If site stays predictable from the residuals, the policy model is under-fitting and the gains, if any, come from somewhere else.
References
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors
- Domain Adaptation under Missingness Shift
- GRU-ODE-Bayes: Continuous Modeling of Sporadically-Observed Time Series
- Neural ODEs for Informative Missingness in Multivariate Time Series
- Temporal Belief Memory: Imputing Missing Data during RNN Training
- Adversarial Spectral Kernel Matching for Unsupervised Time Series Domain Adaptation
- CALDA: Improving Multi-Source Time Series Domain Adaptation With Contrastive Adversarial Learning
- Fan-Net: Fourier-Based Adaptive Normalization for Cross-Domain Stroke Lesion Segmentation
- Bridging data gaps of rare conditions in ICU: a multi-disease adaptation approach for clinical prediction
- Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML