One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification

Kaiwei Wang; Mengzhou Gao; Pengfei Jiao

arxiv: 2605.10179 · v1 · submitted 2026-05-11 · 💻 cs.LG · cs.AI

One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification

Mengzhou Gao , Kaiwei Wang , Pengfei Jiao This is my paper

Pith reviewed 2026-05-12 04:37 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords neural flowsgraph neural networksirregular time seriesself-supervised learninginteraction modelingmultivariate classificationODE trajectories

0 comments

The pith

One-step Graph-Structured Neural Flows capture inter-variable interactions through auxiliary trajectory self-supervision for irregular time series classification.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to establish that one-step neural flows can be made to learn meaningful interactions between variables in irregular multivariate time series, even though the single mapping removes the iterative refinement that usually allows interactions to emerge. It does this by structuring the flows on a graph and adding two self-supervision strategies: one that re-initializes trajectories to force divergence whose lower bound is derived theoretically, and one that generates reverse-time trajectories to enforce forward-backward consistency via the invertibility of flows. A reader would care because many sensor, medical, and financial records arrive at irregular intervals with coupled variables, yet repeated numerical integration of ODEs remains computationally expensive. The experiments report that the resulting model reaches state-of-the-art accuracy on five real-world datasets while keeping training time and memory competitive.

Core claim

GSNF introduces two auxiliary-trajectory self-supervision strategies to strengthen interaction learning inside one-step graph-structured neural flows: interaction-aware trajectory generation via re-initialization, which induces trajectory divergence to expose graph-induced interactions together with a theoretically derived lower bound on that divergence, and reverse-time trajectory generation, which enforces forward-backward consistency to regularize graph learning by exploiting flow invertibility.

What carries the argument

The pair of auxiliary-trajectory self-supervision strategies (re-initialization divergence and reverse-time consistency) that regularize graph learning inside the one-step neural flow.

Load-bearing premise

The two auxiliary self-supervision strategies reliably induce and regularize meaningful graph-induced interactions without introducing artifacts or overfitting to the datasets.

What would settle it

A controlled ablation in which removing both the re-initialization and reverse-time strategies causes classification accuracy to fall back to the level of independent-variable neural flow baselines on the same five datasets would falsify the claim that these strategies strengthen interaction learning.

Figures

Figures reproduced from arXiv: 2605.10179 by Kaiwei Wang, Mengzhou Gao, Pengfei Jiao.

**Figure 1.** Figure 1: Comparison of continuous-time models. Neural ODEs evolve latent states via step-by-step numerical integration, whereas Neural Flows perform one-step mappings. Graph Neural ODEs incorporate interaction modeling by conditioning dynamics on an interaction graph, while Graph-Structured Neural Flows integrate graph-conditioned interactions into one-step mappings. Existing continuous-time models for irregular mu… view at source ↗

**Figure 2.** Figure 2: Overall framework. Top:Inference and generation with one-step GSNF. During inference, segment-level interaction graphs {A (s) } C s=1 are aggregated into a posterior qϕ(A | X), and latent states are propagated backward through GSNF, conditioned on their corresponding segment-level graphs to infer qϕ(z0 | X). During generation, samples of A and z0 are used by GSNF to compute latent states forward at arbitra… view at source ↗

**Figure 3.** Figure 3: further examines computational efficiency. Within the top AUPRC tier, GSNF achieves the lowest training time. This efficiency is largely attributed to the IVP-VAE backbone, which exhibits the lowest training time among all the baselines. Notably, despite being the fastest baseline, IVP-VAE lies in the third performance tier; GSNF incorporating interaction modeling resolves this limitation by elevating GSN… view at source ↗

**Figure 4.** Figure 4: Interaction graphs with and without ITG on the P19 dataset. Attention weights are normalized by softmax, and edges highlighted by red boxes denote attention values above 0.035. With ITG, the interaction graph contains substantially more salient connections (78 edges) than without ITG (34 edges), with the additional connections primarily concentrated around key clinical variables. (0: HR, 3: SBP, 4: MAP, 22… view at source ↗

**Figure 7.** Figure 7: Training dynamics of the theoretically derived separation margin δlb and total loss on PhysioNet12. The theoretically derived margin δlb maintains a comparable scale to the manually selected margin δ, on the order of 10−6 , across training. small margins provide insufficient separation of interactioninduced differences, reflecting a trade-off between interaction exposure and dynamical fidelity. Theoretic… view at source ↗

**Figure 6.** Figure 6: Parameter sensitivity analysis of GSNF on PhysioNet12. Performance peaks at an intermediate t ∗ 0 and depends on δ mainly at the order-of-magnitude scale, with an optimum around 10−6 . remain weakly connected or isolated, leading to a sparse graph that captures only a small subset of interactions. 6.4. Parameters Sensitivity We analyze the sensitivity of ITG to the re-initialization time and the tuned marg… view at source ↗

**Figure 8.** Figure 8: Main Result using the first 48 hours of data. Due to class imbalance in these datasets, we assess classification performance using the area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC). All models were tested in the same computing environment. The details are as follows: • Operating System: Ubuntu 22.04.1 LTS • CPU: Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz • GPU: NVIDIA… view at source ↗

read the original abstract

Neural Flows efficiently model irregular multivariate time series by directly learning ODE solution trajectories with neural networks, bypassing step-by-step numerical solvers. Despite their efficiency, many existing approaches treat variables independently, leaving inter-variable interactions underexplored. Moreover, their one-step mapping makes interaction modeling inherently challenging, as it removes the iterative refinement of interactions during learning. To address this challenge, we propose one-step Graph-Structured Neural Flows (GSNF), which introduce two auxiliary-trajectory self-supervision strategies to strengthen interaction learning: (i) interaction-aware trajectory generation via re-initialization, which induces trajectory divergence to expose graph-induced interactions, with a theoretically derived lower bound on divergence; and (ii) reverse-time trajectory generation, which enforces forward-backward consistency to regularize graph learning, enabled by flow invertibility. Experiments on five real-world datasets show that GSNF achieves state-of-the-art classification performance with highly competitive training time and memory usage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GSNF pairs one-step neural flows with graphs and two self-supervision tricks for irregular time series, but the divergence lower bound looks vulnerable to the irregular sampling the method targets.

read the letter

This paper takes neural flows, which learn irregular multivariate trajectories in one shot without ODE solvers, and adds graph structure to model variable interactions. It then introduces two auxiliary self-supervision losses: re-initialization that forces divergence to surface those interactions (with a claimed lower bound) and reverse-time generation that uses flow invertibility for forward-backward consistency. The concrete combination is new relative to the cited neural-flow and graph baselines, and the reverse-time piece looks clean because invertibility is already a flow property. The efficiency claims also make sense for the target setting of sensor or medical data where one-step evaluation is attractive. Experiments report SOTA classification on five datasets alongside competitive training time and memory, which is the practical payoff if it holds. The main soft spot is the re-initialization strategy. The lower bound on divergence is presented as theoretically derived, yet the stress-test concern is fair: irregular, non-uniform sampling likely violates the regularity conditions (Lipschitz, uniform grids, etc.) needed for the bound to control meaningful interactions rather than sampling artifacts or capacity effects. Without seeing the derivation or ablations that isolate the graph contribution from the extra losses, it is hard to know whether the gains are robust. The abstract also omits error bars and controls, so the SOTA numbers are difficult to interpret at face value. This is aimed at the neural-ODE and irregular time-series crowd who already use flows and want to add interactions without losing the one-step speed. A reader working on similar methods would get value from the auxiliary-trajectory ideas if the math checks out. It deserves peer review so referees can examine the bound derivation and run the missing controls.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes one-step Graph-Structured Neural Flows (GSNF) for irregular multivariate time series classification. It augments neural flows with two auxiliary-trajectory self-supervision strategies: (i) interaction-aware trajectory generation via re-initialization that induces divergence to expose graph-induced interactions, supported by a theoretically derived lower bound on divergence, and (ii) reverse-time trajectory generation that enforces forward-backward consistency via flow invertibility. The method is evaluated on five real-world datasets and claims state-of-the-art classification accuracy together with competitive training time and memory usage.

Significance. If the lower bound holds under irregular sampling and the auxiliary objectives reliably regularize meaningful inter-variable interactions without artifacts, the work would offer an efficient route to interaction-aware modeling of irregular MTS that avoids iterative ODE solvers. The explicit theoretical bound and use of invertibility are positive features that could support reproducibility and generalization claims.

major comments (1)

[Abstract and §3] Abstract and §3 (theoretical derivation): the lower bound on divergence for the re-initialization strategy is presented as theoretically derived and central to exposing graph-induced interactions. However, the derivation appears to rely on regularity conditions (Lipschitz continuity, uniform grids, or invertibility properties) that are not guaranteed to hold for one-step flows on sparsely and non-uniformly sampled observations; if violated, the induced divergence may capture sampling artifacts or model capacity rather than variable interactions, directly undermining the claim that this auxiliary objective strengthens interaction learning.

minor comments (2)

[Abstract] Abstract: quantitative results are stated as SOTA without reference to error bars, number of runs, or statistical significance tests; this makes it difficult to assess whether the reported gains are robust.
The manuscript would benefit from an explicit statement of how the lower bound is incorporated into the training loss (e.g., as a regularizer term or constraint) and from ablation controls that isolate the contribution of each auxiliary strategy.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough review and valuable feedback on our work. We address the major comment point-by-point below, providing clarifications on the theoretical derivation while committing to revisions that strengthen the manuscript without altering its core contributions.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (theoretical derivation): the lower bound on divergence for the re-initialization strategy is presented as theoretically derived and central to exposing graph-induced interactions. However, the derivation appears to rely on regularity conditions (Lipschitz continuity, uniform grids, or invertibility properties) that are not guaranteed to hold for one-step flows on sparsely and non-uniformly sampled observations; if violated, the induced divergence may capture sampling artifacts or model capacity rather than variable interactions, directly undermining the claim that this auxiliary objective strengthens interaction learning.

Authors: We appreciate the referee's careful scrutiny of the theoretical section. The lower bound in §3 is derived from the Lipschitz continuity of the neural flow mapping, which is guaranteed by our architecture (bounded activations and weight constraints, as specified in the model description and supplementary material). The derivation does not assume uniform grids or iterative stepping; it applies directly to the one-step flow evaluated at arbitrary observed time points, using the difference in initial conditions induced by graph perturbations. Invertibility is not invoked for the re-initialization strategy (it applies only to the reverse-time auxiliary objective). We acknowledge that the presentation could more explicitly address irregular sampling to rule out potential artifacts. In the revision, we will expand §3 with a dedicated remark on the bound's validity under non-uniform sampling, include a proof sketch clarifying the Lipschitz assumption's independence from grid regularity, and add an ablation study demonstrating that divergence correlates with interaction strength rather than sampling density or model capacity. This addresses the concern while preserving the original claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's central proposal introduces auxiliary self-supervision (re-initialization divergence with a claimed theoretically derived lower bound, plus reverse-time consistency via invertibility) to regularize graph-structured one-step flows. These are presented as additional training objectives rather than as outputs derived from the main model parameters by construction. No equations or steps are shown reducing a 'prediction' (e.g., interaction strength or classification) to a fitted quantity or self-citation chain. The method is evaluated on five external real-world datasets for SOTA performance, providing independent falsifiability. Self-supervision depends on the same parameters by design of any auxiliary loss, but this does not constitute circularity per the rules when the bound is external and results are benchmarked outside the fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; the paper appears to rest on standard neural-ODE existence assumptions and graph-neural-network message-passing assumptions, plus the claim that the two self-supervision objectives expose true interactions. No explicit free parameters or invented entities are named in the abstract.

pith-pipeline@v0.9.0 · 5459 in / 1279 out tokens · 26697 ms · 2026-05-12T04:37:44.709891+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 5.1 (A Data-Dependent Lower Bound for the ITG Separation Margin)... η=σ_min(A)σ_min(W) Δ_in ... cumulative trajectory divergence satisfies ∑∥z∗(ti)−z(ti)∥ ≥ max{0,(L−k∗0+1)(η−Δin)}
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

GSNF ... z(t)=F(z(t0),t0,t,A) ... g(z(t0),t0,t,A)=MLP(...)⊙GCN(A,z(t0)||t||t0) ... invertibility if φ∈[0,1) and g contractive (spectral normalization)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 1 internal anchor

[1]

A., and Mark, R

Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L. A., and Mark, R. Mimic-iv.PhysioNet. Avail- able online at: https://physionet. org/content/mimi- civ/1.0/(accessed August 23, 2021), pp. 49–55,

work page 2021
[2]

Kingma, D. P. and Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

A., Josef, C

9 Submission and Formatting Instructions for ICML 2026 Reyna, M. A., Josef, C. S., Jeter, R., Shashikumar, S. P., Westover, M. B., Nemati, S., Clifford, G. D., and Sharma, A. Early prediction of sepsis from clinical data: the phys- ionet/computing in cardiology challenge 2019.Critical care medicine, 48(2):210–217,

work page 2026
[4]

In 2012 computing in cardiology, pp. 245–248. IEEE,

work page 2012
[5]

Proof of Theorem 4.1 (Invertibility of Graph-Structured Neural Flows) Proof

10 Submission and Formatting Instructions for ICML 2026 A. Proof of Theorem 4.1 (Invertibility of Graph-Structured Neural Flows) Proof. Let F(z(t 0), t0, t, A) denote the GSNF defined in Eq. (3). We assume that the interaction function g(·, t0, t, A) is contractive with Lipschitz constant Lg <1 , which is ensured by applying spectral normalization to all ...

work page 2026
[6]

We now lower-bound the residual term

+φ(t−t ∗ 0)g(z ∗ 0, t∗ 0, t, A)−φ(t−t 0)g(z 0, t0, t, A).(20) Taking norms and applying the reverse triangle inequality yields ∥δ(t)∥ ≥ φ(t−t ∗ 0)g(z ∗ 0, t∗ 0, t, A)−φ(t−t 0)g(z 0, t0, t, A) − ∥z∗ 0 −z 0∥.(21) Denote∆ in :=∥z ∗ 0 −z 0∥. We now lower-bound the residual term. Recall that g(z(t0), t0, t, A) = MLP(z(t0), t0, t)⊙GCN(z(t 0), t0, t, A), and ass...

work page 2026
[7]

12 Submission and Formatting Instructions for ICML 2026 δ(×10−6)Physionet12 P12 P19 MIMIC-IV eICU Hyperparameter 1.00 10.00 10.00 0.10 10.00 Lower Bound 3.51 22.54 15.76 0.47 29.82 Table 3.Comparison of manually selected separation margins (δ) and the calculated theoretical lower bound (δ lb). D. Comprehensive Experiments D.1. Memory Usage and Training Ti...

work page 2026
[8]

All five datasets are used for classification experiments

For testing, the model bypasses loss calculation and directly outputs the predicted labels via a forward pass. All five datasets are used for classification experiments. Each dataset is randomly split into 80% for training, 10% for validation, and 10% for testing. Following previous works (Rubanova et al., 2019; Shukla & Marlin, 2021; Zhang et al., 2022),...

work page 2019
[9]

Hyperparameter Value Scope Optimizer Adam All Weight decay1×10 −4 All Batch size 50 All Learning rate (LR)1×10 −3 All LR scheduler step 20 All LR decay factor 0.5 All Number of GSNF layers 2 GSNF Latent dimension Number of sensors GSNF Hidden layers 3 GSNF Hidden dimension 128 GSNF Cross-entropy weightα1000 GSNF ITG weightβ0.1 GSNF RTG weightγ0.1 GSNF Tab...

work page 2026
[10]

Physionet12 P12 P19 MIMIC-IV eICU #Samples 3,989 11,988 38,803 26,070 12,312 #Variables 37 36 39 96 14 Missing ratio (%) 84.34 88.4 94.9 97.95 65.25 Positive rate (%) 13.89 7 4 13.39 17.61 Table 6.Key information of the five datasets. ThePhysioNet 2012dataset (Silva et al., 2012)was released for the PhysioNet/Computing in Cardiology Challenge 2012, aiming...

work page 2012
[11]

We used the processed data provided by Raindrop (Zhang et al., 2022)

comprises data from 11,988 patients, including 36 sensor variables and a binary label indicating survival during hospitalization. We used the processed data provided by Raindrop (Zhang et al., 2022). TheP19dataset (Reyna et al.,

work page 2022
[12]

It contains patient information from ICU stays, comprising static demographics and sparse time-dependent physiological measurements

was released for the PhysioNet/Computing in Cardiology Challenge 2019, aiming to predict the onset of sepsis. It contains patient information from ICU stays, comprising static demographics and sparse time-dependent physiological measurements. In our experiments, we utilize 38803 variable-length time series, focusing on 39 features (5 static and 34 time-de...

work page 2019
[13]

is a multivariate time series dataset composed of sparse and irregularly sampled physiological data collected at the Beth Israel Deaconess Medical Center between 2008 and

work page 2008
[14]

A total of 26,070 patient stays are retained for use in classification tasks

Following a preprocessing approach similar to that of Neural Flow (Biloˇs et al., 2021), we extract 96 features — including patient intake/output, lab results, and medication prescriptions — from the first 48 hours post-ICU admission. A total of 26,070 patient stays are retained for use in classification tasks. TheeICUCollaborative Research Database (Poll...

work page 2021
[15]

contains data from patients admitted to ICUs across 208 hospitals in the United States between 2014 and

work page 2014
[16]

Following the preprocessing steps outlined by IVP-V AE (Xiao et al., 2024), we extract 14 features within the initial 48 hours post-ICU admission from a total of 12,312 patient stays. E.3. Baselines We compare our model against several baselines for the classification of multivariate irregular time-series. •Continuous-time model: – GRU-D(Che et al.,

work page 2024
[17]

•Graph-based models: 15 Submission and Formatting Instructions for ICML 2026 – Raindrop(Zhang et al.,

employs an invertible neural flow to learn the geometry of the control path, leveraging invertibility constraints to construct a continuous and data-adaptive manifold for robust modeling of sparse and irregularly-sampled time series. •Graph-based models: 15 Submission and Formatting Instructions for ICML 2026 – Raindrop(Zhang et al.,

work page 2026

[1] [1]

A., and Mark, R

Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L. A., and Mark, R. Mimic-iv.PhysioNet. Avail- able online at: https://physionet. org/content/mimi- civ/1.0/(accessed August 23, 2021), pp. 49–55,

work page 2021

[2] [2]

Kingma, D. P. and Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv

[3] [3]

A., Josef, C

9 Submission and Formatting Instructions for ICML 2026 Reyna, M. A., Josef, C. S., Jeter, R., Shashikumar, S. P., Westover, M. B., Nemati, S., Clifford, G. D., and Sharma, A. Early prediction of sepsis from clinical data: the phys- ionet/computing in cardiology challenge 2019.Critical care medicine, 48(2):210–217,

work page 2026

[4] [4]

In 2012 computing in cardiology, pp. 245–248. IEEE,

work page 2012

[5] [5]

Proof of Theorem 4.1 (Invertibility of Graph-Structured Neural Flows) Proof

10 Submission and Formatting Instructions for ICML 2026 A. Proof of Theorem 4.1 (Invertibility of Graph-Structured Neural Flows) Proof. Let F(z(t 0), t0, t, A) denote the GSNF defined in Eq. (3). We assume that the interaction function g(·, t0, t, A) is contractive with Lipschitz constant Lg <1 , which is ensured by applying spectral normalization to all ...

work page 2026

[6] [6]

We now lower-bound the residual term

+φ(t−t ∗ 0)g(z ∗ 0, t∗ 0, t, A)−φ(t−t 0)g(z 0, t0, t, A).(20) Taking norms and applying the reverse triangle inequality yields ∥δ(t)∥ ≥ φ(t−t ∗ 0)g(z ∗ 0, t∗ 0, t, A)−φ(t−t 0)g(z 0, t0, t, A) − ∥z∗ 0 −z 0∥.(21) Denote∆ in :=∥z ∗ 0 −z 0∥. We now lower-bound the residual term. Recall that g(z(t0), t0, t, A) = MLP(z(t0), t0, t)⊙GCN(z(t 0), t0, t, A), and ass...

work page 2026

[7] [7]

12 Submission and Formatting Instructions for ICML 2026 δ(×10−6)Physionet12 P12 P19 MIMIC-IV eICU Hyperparameter 1.00 10.00 10.00 0.10 10.00 Lower Bound 3.51 22.54 15.76 0.47 29.82 Table 3.Comparison of manually selected separation margins (δ) and the calculated theoretical lower bound (δ lb). D. Comprehensive Experiments D.1. Memory Usage and Training Ti...

work page 2026

[8] [8]

All five datasets are used for classification experiments

For testing, the model bypasses loss calculation and directly outputs the predicted labels via a forward pass. All five datasets are used for classification experiments. Each dataset is randomly split into 80% for training, 10% for validation, and 10% for testing. Following previous works (Rubanova et al., 2019; Shukla & Marlin, 2021; Zhang et al., 2022),...

work page 2019

[9] [9]

Hyperparameter Value Scope Optimizer Adam All Weight decay1×10 −4 All Batch size 50 All Learning rate (LR)1×10 −3 All LR scheduler step 20 All LR decay factor 0.5 All Number of GSNF layers 2 GSNF Latent dimension Number of sensors GSNF Hidden layers 3 GSNF Hidden dimension 128 GSNF Cross-entropy weightα1000 GSNF ITG weightβ0.1 GSNF RTG weightγ0.1 GSNF Tab...

work page 2026

[10] [10]

Physionet12 P12 P19 MIMIC-IV eICU #Samples 3,989 11,988 38,803 26,070 12,312 #Variables 37 36 39 96 14 Missing ratio (%) 84.34 88.4 94.9 97.95 65.25 Positive rate (%) 13.89 7 4 13.39 17.61 Table 6.Key information of the five datasets. ThePhysioNet 2012dataset (Silva et al., 2012)was released for the PhysioNet/Computing in Cardiology Challenge 2012, aiming...

work page 2012

[11] [11]

We used the processed data provided by Raindrop (Zhang et al., 2022)

comprises data from 11,988 patients, including 36 sensor variables and a binary label indicating survival during hospitalization. We used the processed data provided by Raindrop (Zhang et al., 2022). TheP19dataset (Reyna et al.,

work page 2022

[12] [12]

It contains patient information from ICU stays, comprising static demographics and sparse time-dependent physiological measurements

was released for the PhysioNet/Computing in Cardiology Challenge 2019, aiming to predict the onset of sepsis. It contains patient information from ICU stays, comprising static demographics and sparse time-dependent physiological measurements. In our experiments, we utilize 38803 variable-length time series, focusing on 39 features (5 static and 34 time-de...

work page 2019

[13] [13]

is a multivariate time series dataset composed of sparse and irregularly sampled physiological data collected at the Beth Israel Deaconess Medical Center between 2008 and

work page 2008

[14] [14]

A total of 26,070 patient stays are retained for use in classification tasks

Following a preprocessing approach similar to that of Neural Flow (Biloˇs et al., 2021), we extract 96 features — including patient intake/output, lab results, and medication prescriptions — from the first 48 hours post-ICU admission. A total of 26,070 patient stays are retained for use in classification tasks. TheeICUCollaborative Research Database (Poll...

work page 2021

[15] [15]

contains data from patients admitted to ICUs across 208 hospitals in the United States between 2014 and

work page 2014

[16] [16]

Following the preprocessing steps outlined by IVP-V AE (Xiao et al., 2024), we extract 14 features within the initial 48 hours post-ICU admission from a total of 12,312 patient stays. E.3. Baselines We compare our model against several baselines for the classification of multivariate irregular time-series. •Continuous-time model: – GRU-D(Che et al.,

work page 2024

[17] [17]

•Graph-based models: 15 Submission and Formatting Instructions for ICML 2026 – Raindrop(Zhang et al.,

employs an invertible neural flow to learn the geometry of the control path, leveraging invertibility constraints to construct a continuous and data-adaptive manifold for robust modeling of sparse and irregularly-sampled time series. •Graph-based models: 15 Submission and Formatting Instructions for ICML 2026 – Raindrop(Zhang et al.,

work page 2026