Pith. sign in

REVIEW 3 major objections 5 minor 33 references

A timestamp-attributed, bidirectional event graph predicts post-48h ICU mortality, and an ablation shows graph connectivity—not heterogeneous edge typing—carries most of the performance.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 00:46 UTC pith:6OGWNE7J

load-bearing objection A careful ablation study whose central claim is undermined by an internally contradictory reverse-edge ablation; the interesting findings are the negative ones. the 3 major comments →

arxiv 2608.02663 v1 pith:6OGWNE7J submitted 2026-08-01 cs.LG cs.AI

CT-HEG: A Bidirectional, Timestamp-Attributed Event Graph for ICU In-Hospital Mortality Prediction - An Architectural Ablation Study

Mohammad Nasir Uddin , Rahnuma Tabassum Orpita , Asaduzzaman Anik , Eklachur Rahman Bhuiyan , Marjahan Risalat , SM Wali Ullah , Asif Ahamed This is my paper
classification cs.LG cs.AI
keywords ICU mortality predictionheterogeneous graph neural networkirregular time seriescontinuous-time event graphedge timestampsablation studyMIMIC-IVGATv2Conv
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that irregular ICU observations can be encoded as a typed, timestamped event graph and that this representation supports accurate, reasonably calibrated post-48h in-hospital mortality prediction. Its central contribution is an architectural ablation: bidirectional connectivity between observation nodes and the visit readout is structurally required, timestamp-and-value edge attributes add a small but consistent gain, and heterogeneous edge typing does not help predictive performance in this cohort. If true, this gives an imputation-free backbone for ICU risk models and a concrete warning that relation-specific message passing is not automatically worth its extra parameters.

Core claim

On MIMIC-IV v3.1 (31,142 ICU stays with length of stay at least 48 hours, 13.4% mortality), the authors build each stay as a CT-HEG graph with visit, vital, and lab-event nodes, and edges carrying a 2D attribute of normalized time and value. The CHIRP-Net instantiation, a four-layer heterogeneous GATv2Conv network with edge-dim 2, reaches a 5-seed mean test AUROC of 0.8449 ± 0.0071, with an ensemble AUROC of 0.8618 and post-temperature-scaling ECE of 0.0307. The largest ablation effect is connectivity: removing reverse edges, which in the described topology leaves observation nodes without incoming messages, drops AUROC by 0.1968 ± 0.0073 to a 0.6482 floor that the authors interpret as a det

What carries the argument

The central object is the CT-HEG schema: each ICU stay becomes a typed, directed graph with visit, vital, and lab_event nodes, and every edge carries a 2D attribute (t_hours/48, value_norm) so that observation timing and value enter attention directly without imputation. The load-bearing mechanism in the ablation is bidirectional connectivity: reverse edges (visit-to-observation) allow observation nodes to be updated with visit context across stacked message-passing layers, and removing them collapses the four-layer network. GATv2Conv with edge-dim 2 is the message-passing engine that makes attention depend on both the clinical value and the timestamp of each observation.

Load-bearing premise

The load-bearing premise is that the edges removed in the central ablation really are the visit-to-observation reverse edges described in the schema; if the code instead removed the observation-to-visit edges, the performance collapse is just a disconnected graph and says nothing about bidirectionality.

What would settle it

The decisive check is to publish the exact edge-direction implementation and rerun the reverse-edge ablation with each direction removed independently: if removing only visit-to-observation edges does not produce the 0.6482 AUROC floor, the bidirectionality claim fails; if the floor appears only when observation-to-visit edges are removed, the result is trivial disconnection. A second falsifying test is external validation on an independent ICU cohort.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Post-48h ICU mortality can be predicted from an irregular, imputation-free event graph with AUROC around 0.845 on this cohort, and an ensemble of five seeds reaches 0.862.
  • Any variant of this topology that drops bidirectional connectivity forfeits most of the predictive signal; the effect is structural, not a tuning choice.
  • Encoding timestamps as continuous edge attributes contributes a small but consistent AUC gain over constant edge attributes, supporting the continuous-time encoding design.
  • Heterogeneous edge typing is not shown to help in this cohort: a single-relation GAT with 7x fewer parameters outperformed the full typed model on all five seeds.
  • Validation-fitted temperature scaling yields well-calibrated probabilities (ECE 0.0307) on the untouched test set, supporting calibrated use of the risk score.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The strongest conclusion—that bidirectional connectivity is necessary—should be treated as conditional until the released code and checkpoints confirm which edge directions the ablation actually removed, because the paper describes reverse edges inconsistently.
  • The same schema could plausibly transfer to other ICU outcomes such as sepsis onset, acute kidney injury, or ventilator weaning, but that is an extension the paper only sketches.
  • If the homogeneous-GAT advantage reproduces on external cohorts, the practical lesson may be to invest in temporal edge attributes and connectivity design rather than in fine-grained edge typing.
  • A streaming graph construction variant would test whether the offline 48h-snapshot build limits bedside deployment, which the paper explicitly leaves to future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces CT-HEG, a typed, timestamp-attributed heterogeneous graph schema for ICU in-hospital mortality prediction, and instantiates it as CHIRP-Net, a four-layer heterogeneous GATv2Conv network. On a MIMIC-IV v3.1 cohort of 31,142 ICU stays (LOS≥48h), the model achieves 5-seed mean AUROC 0.8449±0.0071 and AUPRC 0.4958±0.0209, with an ensemble AUROC of 0.8618 (95% CI 0.8485–0.8745) and post-temperature-scaling ECE 0.0307. The paper also presents three ablations: removing reverse edges (reported AUROC drop 0.1968 to 0.6482), removing time-attentive edge features (drop 0.0247), and collapsing heterogeneous edge types (which improves AUROC to 0.8638 with 7× fewer parameters). The authors conclude that bidirectional connectivity is structurally necessary, that time-aware edge features help modestly, and that heterogeneous typing is not shown to help in this cohort. They explicitly defer external validation, a pre-specified temporal evaluation, and a fairness audit to future work.

Significance. If the central claims were supported, the paper would make a useful contribution: it proposes a clean, imputation-free graph encoding of irregular EHR data, reports uncertainty honestly (5 seeds, bootstrap CIs, calibration), releases code, and includes a surprising and potentially important negative result (homogeneous GAT outperforming the full heterogeneous model). The candid limitations section and the explicit distinction between seed variation and bootstrap variation are exemplary. However, the paper's headline contribution—the claim that bidirectional connectivity is 'structurally necessary'—is undermined by an internal contradiction between the formal edge-direction definitions and the ablation narrative. Until that contradiction is resolved, the main architectural conclusion is not empirically supported.

major comments (3)
  1. [§3.2 vs. §5.6, Table 5, Abstract] The formal schema defines measured-at edges as vital/lab→visit and reverse edges as visit→vital/lab. If reverse edges are removed under this definition, the vital/lab→visit edges remain intact, so the visit node would still receive messages from observation nodes. Yet §5.6 and the Abstract state that removing reverse edges 'disconnects observation nodes from the visit readout entirely' and collapses AUROC to 0.6482±0.0008. These descriptions are mutually exclusive. If the ablation actually removed the vital/lab→visit edges, then the experiment is a reachability tautology—it removes the only path from input features to the readout—and it does not test whether the reverse (visit→observation) direction contributes anything. The paper's headline claim that bidirectional connectivity is 'necessary' is therefore unsupported by the reported experiment. The released repository does not currently
  2. [§5.6, Table 5] The AUROC collapse to 0.6482±0.0008 with a very small SD is exactly what one expects from an architectural disconnection—the visit node becomes an isolated embedding lookup. This is a graph-connectivity sanity check, not an empirical finding about learned representations. Presenting it as a 'verified' result and using it to conclude that bidirectional connectivity is necessary overstates what the experiment can show. At minimum, the paper should label this ablation as a structural integrity check and should not use it to support the bidirectional-necessity claim without a separate experiment that keeps observation→visit edges while removing visit→observation edges.
  3. [Abstract, Key Messages, Conclusion] The claim 'Bidirectional connectivity was necessary for the model to use its inputs at all' is not supported even under a corrected reading of the ablation. The described experiment can at most show that some observation-to-visit information path is required; it does not compare models with both edge directions against models with only the observation→visit direction. The term 'bidirectional' specifically implicates the visit→observation direction, but no experiment isolates that direction. Please either perform a proper ablation that preserves observation→visit edges while removing visit→observation edges, or revise the central claim to 'observation-to-visit connectivity is necessary' and adjust the title, abstract, and key messages accordingly.
minor comments (5)
  1. [§5.1 / Reproducibility] The paper states that split indices and checkpoints will be added 'prior to publication.' For a paper whose main claim hinges on the exact edge-direction mapping, these artifacts are essential. Please include them with the revision, or at least provide a code snippet that instantiates the forward/reverse edge types unambiguously.
  2. [§5.2 / §6.2, Table 3] GRU-D is reported from a single run while other models use 5 seeds, and no paired significance test is provided. The paper candidly acknowledges this in Limitations, but the text in the Results still says CHIRP-Net 'outperforms GRU-D by 0.002 AUROC'; given the seed SD of 0.0071, this phrasing is misleading. Please rephrase to 'performed comparably' or 'no reliable difference was observed.'
  3. [Table 5 note] The note about a 'normalization-leakage bug' in the original ablation scripts is vague. Please specify what the bug was and how the corrected logs differ from the original ones, since this directly affects trust in the reported ablation numbers.
  4. [§3.3 / §4.3] The term 'continuous-time' is used in the title and abstract, but §4.3 correctly clarifies that time is encoded as a continuous-valued edge attribute in an offline-built graph, not an online or streaming model. This clarification should appear earlier to avoid overclaiming.
  5. [General] Minor typographical and formatting issues: the affiliation list uses inconsistent spacing (e.g., 'V A', 'Alexandria, V A'), and the reference list includes several arXiv identifiers that should be updated to published versions where available.

Circularity Check

1 steps flagged

The 'bidirectional necessity' claim is a reachability tautology produced by mislabeling the ablated edge direction.

specific steps
  1. self definitional [Section 5.6 (Ablation Studies) and Table 5; cf. Section 3.2 edge definitions; Abstract and Key Messages]
    "First, removing reverse edges collapses AUROC to 0.6482±0.0008 ... In the CT-HEG topology, observation nodes are leaves with no incoming edges; without reverse edges, multi-layer message passing cannot propagate information across observations, reducing the 4-layer network to a single embedding lookup. ... The schema implemented and evaluated in this paper is: measured-at(vital/lab→visit) and its reverse (visit→vital/lab), giving the 4 forward+reverse edge types referenced in Fig. 1."

    Under the paper's own Section 3.2 definition, 'reverse' edges are visit→vital/lab. Removing them leaves the vital/lab→visit forward edges intact, so observation nodes would still send messages to the visit readout. The reported collapse to 0.6482±0.0008 and 'single embedding lookup' can only occur if the ablation actually removed the data-carrying observation→visit edges. That makes the result a reachability tautology: with no path from observation features to the readout, the model cannot use its inputs by construction. The abstract's conclusion that 'Bidirectional connectivity was necessary for the model to use its inputs at all' is therefore not an empirical discovery but an artifact of conflating 'remove reverse edges' with 'delete all observation-to-visit edges.' The paper's own forma

full rationale

The paper's central headline claim—that bidirectional connectivity is structurally necessary—is not supported by its own schema and ablation description, and the reported effect reduces by construction. Section 3.2 defines the edges as measured-at (vital/lab→visit) and its reverse (visit→vital/lab). Section 5.6 says removing reverse edges makes 'visit nodes receive no messages from observation types' and disconnects observation nodes from the readout, which is only true if the forward, data-carrying edges were removed instead. The deterministic AUROC floor (SD≈0.0008) and 'single embedding lookup' description confirm that the ablation severed all paths from inputs to the output—a definitional collapse, not evidence for bidirectionality. This affects the abstract, Key Messages, Section 5.6, Table 5, and Conclusion. The other empirical contributions (time-attentive edge features contributing 0.0247±0.0093, homogeneous GAT outperforming the full model, calibration ECE=0.0307, baseline comparisons) are self-contained and do not reduce to their inputs; there is no load-bearing self-citation. Because the paper's foremost architectural conclusion is a tautology under its own definitions, but substantial independent results remain, the circularity score is 6.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claims rest on a designed representation (CT-HEG/CHIRP-Net) plus data-handling assumptions. The fitted numbers are the validation-fitted temperature, untuned hyperparameters, and post hoc reporting thresholds. The most fragile unverifiable assumption is leakage-free normalization, given the disclosed normalization-leakage bug in the ablation scripts.

free parameters (4)
  • Temperature scaling T = 1.558
    Fitted on the validation split, applied to the untouched test set (Table 4). Standard calibration practice and disclosed, but it is a fitted number the ECE=0.0307 claim depends on.
  • LOS auxiliary loss weight lambda = 0.05
    Hand-chosen regularizer in Section 4.6; not tuned or ablated.
  • CHIRP-Net hyperparameters (untuned) = hidden=192, heads=4, layers=4, dropout=0.25, LR=3e-4, weight_decay=3e-4
    Single fixed configuration explicitly stated as not tuned (Table 2 note); all results depend on it.
  • Reporting threshold for operating points = 0.15 (with 0.10, 0.20, 0.25)
    Sensitivity/specificity/PPV/NPV reported at hand-picked thresholds without a pre-specified selection rule; does not affect AUROC but shapes the clinical reading.
axioms (4)
  • domain assumption The MIMIC-IV v3.1 cohort (LOS>=48h, first ICU stay per patient) defines a valid estimand for post-48h in-hospital mortality
    Section 3.1 and 5.1. The paper itself notes the LOS>=48h exclusion changes the estimand and calls it a limitation; all point estimates depend on this cohort choice.
  • domain assumption GATv2Conv with edge_dim=2 makes attention genuinely time- and value-aware
    Sections 4.3-4.4. The only support is the 0.0247+/-0.0093 AUROC delta; no mechanism check shows attention weights depend on the time attribute as intended.
  • domain assumption Final z-score normalization statistics for value_norm are computed on the training split only
    Table 5 note discloses a normalization-leakage bug in ablation scripts that was fixed, but does not demonstrate the final pipeline is leakage-free; normalization statistics are not yet released, so the assumption is unverifiable.
  • domain assumption Patient-level split disjointness holds exactly as claimed
    Section 5.1 asserts and reports a verification with the released cohort script; the claim is reasonable but is still an input assumption shaping all reported test metrics.
invented entities (2)
  • CT-HEG typed event graph schema (visit/vital/lab_event nodes; measured-at and reverse edge types with 2D attributes) no independent evidence
    purpose: Represent each ICU stay's irregular observations as a graph without imputation; the paper's core representational contribution.
    A designed representation, not an entity with an external falsifiable handle; its usefulness is measured only through the prediction task reported here.
  • CHIRP-Net (4-layer heterogeneous GATv2Conv instantiation of CT-HEG) no independent evidence
    purpose: Produce mortality logits from the event graph; the model whose results are reported.
    A model instance; its only evidence is this paper's experiments, so it carries no independent support.

pith-pipeline@v1.3.0-alltime-deepseek · 14630 in / 29335 out tokens · 297639 ms · 2026-08-05T00:46:38.746556+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CT-HEG: A Bidirectional, Timestamp-Attributed Event Graph for ICU In-Hospital Mortality Prediction - An Architectural Ablation Study." pith.science (2026). https://pith.science/paper/6OGWNE7J

@misc{pith2026260802663,
  author       = {Pith},
  title        = {Pith review of: CT-HEG: A Bidirectional, Timestamp-Attributed Event Graph for ICU In-Hospital Mortality Prediction - An Architectural Ablation Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OGWNE7J}},
  note         = {Machine review of arXiv:2608.02663}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate ICU mortality prediction requires modeling irregular clinical observations across heterogeneous entity types. Existing sequence models handle irregular sampling but ignore typed relational structure; existing graph models assume fixed-interval inputs. We introduce the Continuous-Time Heterogeneous EHR Graph (CT-HEG) schema and evaluate which architectural choices drive predictive performance. CT-HEG encodes each ICU stay as a typed, timestamped graph with three node types (visit, vital, lab_event) and 2D edge attributes (t_hours/48, value_norm) encoding timing and value without imputation. We instantiate CT-HEG as CHIRP-Net, a four-layer heterogeneous GATv2Conv network, evaluated on MIMIC-IV v3.1 (31,142 ICU stays, LOS>=48h, 13.4% mortality) with five seeds and bootstrapped confidence intervals, against logistic regression, mTAND, a Transformer, and GRU-D, plus an ablation study. CHIRP-Net achieved 5-seed mean AUROC 0.8449+/-0.0071 (AUPRC 0.4958+/-0.0209); the ensemble achieved AUROC 0.8618 (95% CI: 0.8485-0.8745). Removing reverse edges disconnected observation nodes from the visit readout, cutting AUROC by 0.1968+/-0.0073. Time-attentive edge features contributed 0.0247+/-0.0093 AUROC. Collapsing heterogeneous edge types into one relation (7x fewer parameters) outperformed the full model on all seeds. Post-calibration ECE was 0.0307. Temporal and demographic subgroup analyses were explored but not reported here, pending follow-up work. Bidirectional connectivity was necessary for the model to use its inputs at all, and CT-HEG was reasonably well calibrated after validation-fitted temperature scaling. These results support CT-HEG for irregular EHR data, while external validation, a pre-specified temporal evaluation, and a demographic fairness audit remain necessary before any claim of robustness. Code: https://github.com/nasiruddinstudents-ctrl/chirp-net-mimic-iv.

Figures

Figures reproduced from arXiv: 2608.02663 by Asaduzzaman Anik, Asif Ahamed, Eklachur Rahman Bhuiyan, Marjahan Risalat, Mohammad Nasir Uddin, Rahnuma Tabassum Orpita, SM Wali Ullah.

Figure 1
Figure 1. Figure 1: CHIRP-Net architecture overview. Left: CT-HEG with 3 node types (visit, vital, lab event); edges carry a 2D attribute (thours/48, value norm) encoding observation timing and value. Centre: 4 stacked HeteroConv layers (GATv2Conv, edge dim=2, hidden=192, 4 heads); the dashed box marks the 2 forward + 2 reverse edge types providing bidirectional message passing. Right: visit-node readout MLP producing mortali… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

  1. [1]

    Modeling Irregularly Sampled Clinical Time Series,

    S. N. Shukla and B. M. Marlin, “Modeling Irregularly Sampled Clinical Time Series,” arXiv:1812.00531, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    A new efficient ALignment-driven Neural Network for Mortality Predic- tion from Irregular Multivariate Time Series data,

    M. Bignoumba et al., “A new efficient ALignment-driven Neural Network for Mortality Predic- tion from Irregular Multivariate Time Series data,” Expert Systems with Applications, 2023

    work page 2023

  3. [3]

    Predictive Modeling with Temporal Graphical Representation on Electronic Health Records,

    Y. Chen et al., “Predictive Modeling with Temporal Graphical Representation on Electronic Health Records,” Proc. IJCAI, 2024

    work page 2024

  4. [4]

    Deep Learning with Heterogeneous Graph Embeddings for Mortality Pre- diction from Electronic Health Records,

    T. Wanyan et al., “Deep Learning with Heterogeneous Graph Embeddings for Mortality Pre- diction from Electronic Health Records,” Data Intelligence, 2020. 17

    work page 2020

  5. [5]

    CI-GNN: A Granger Causality-Inspired GNN for Interpretable Brain Network- Based Psychiatric Diagnosis,

    S. Zheng et al., “CI-GNN: A Granger Causality-Inspired GNN for Interpretable Brain Network- Based Psychiatric Diagnosis,” Neural Networks, 2023

    work page 2023

  6. [6]

    Graph Neural Network Causal Explanation via Neural Causal Models,

    R. Behnam and W. Wang, “Graph Neural Network Causal Explanation via Neural Causal Models,” Proc. ECCV, 2024

    work page 2024

  7. [7]

    Multi-Time Attention Networks for Irregularly Sampled Time Series,

    S. N. Shukla and B. M. Marlin, “Multi-Time Attention Networks for Irregularly Sampled Time Series,” in Proc. ICLR, 2021

    work page 2021

  8. [8]

    Time-aware Heterogeneous Graph Transformer with Adaptive Attention Merging for Health Event Prediction,

    Z. Li et al., “Time-aware Heterogeneous Graph Transformer with Adaptive Attention Merging for Health Event Prediction,” arXiv, 2024

    work page 2024

  9. [9]

    DynaGraph: interpretable dynamic graph learning for temporal electronic health records,

    M. Mesinovic, S. Molaei, P. Watkinson, and T. Zhu, “DynaGraph: interpretable dynamic graph learning for temporal electronic health records,” npj Digital Medicine, vol. 9, no. 216, 2026

    work page 2026

  10. [10]

    DyGraphTrans: A temporal graph representation learning framework for disease progression from EHRs,

    A. Rahman et al., “DyGraphTrans: A temporal graph representation learning framework for disease progression from EHRs,” bioRxiv, 2026

    work page 2026

  11. [11]

    CF-GNNExplainer: Counterfactual Explanations for Graph Neural Net- works,

    A. Lucic et al., “CF-GNNExplainer: Counterfactual Explanations for Graph Neural Net- works,” Proc. AISTATS, 2022

    work page 2022

  12. [12]

    PMTG: Personalized Multivariate Temporal Graph Learning for Clinical Pre- diction on EHRs,

    Z. Zhu et al., “PMTG: Personalized Multivariate Temporal Graph Learning for Clinical Pre- diction on EHRs,” Proc. IEEE BIBM, 2025

    work page 2025

  13. [13]

    MedGAITS: a graph autoencoder network for modeling irregular time series data in EMRs,

    K. Wang et al., “MedGAITS: a graph autoencoder network for modeling irregular time series data in EMRs,” Health Information Science and Systems, 2026

    work page 2026

  14. [14]

    CausGNN: A Causal-Based Explanation Framework for Graph Neural Networks,

    H. Debbi, “CausGNN: A Causal-Based Explanation Framework for Graph Neural Networks,” Expert Systems, 2026

    work page 2026

  15. [15]

    OrphicX: A Causality-Inspired Latent Variable Model for Interpreting GNNs,

    J. Lin et al., “OrphicX: A Causality-Inspired Latent Variable Model for Interpreting GNNs,” Proc. CVPR, 2022

    work page 2022

  16. [16]

    Explainable Graph-Neural Architectures for ICU Mortality Prediction Using Logic Explained Networks,

    R. Damian, “Explainable Graph-Neural Architectures for ICU Mortality Prediction Using Logic Explained Networks,” Proc. ICAART, 2026

    work page 2026

  17. [17]

    Interpretable Adaptive Graph Fusion Network for Mortality and Complication Prediction in ICUs,

    A. Cifci et al., “Interpretable Adaptive Graph Fusion Network for Mortality and Complication Prediction in ICUs,” Diagnostics, 2025

    work page 2025

  18. [18]

    ICU-TGNN: A Hybrid Multitask Transformer and GNN Model for ICU Clinical Outcomes,

    X. Shi et al., “ICU-TGNN: A Hybrid Multitask Transformer and GNN Model for ICU Clinical Outcomes,” Proc. IEEE SMC, 2024

    work page 2024

  19. [19]

    Continuous Graph Neural Networks,

    L.-P. Xhonneux et al., “Continuous Graph Neural Networks,” Proc. ICML, 2020

    work page 2020

  20. [20]

    Dynamic Graph Attention Modeling Based on Continuous Time Differentia- tion,

    J. Han et al., “Dynamic Graph Attention Modeling Based on Continuous Time Differentia- tion,” Proc. ICMLCA, 2025

    work page 2025

  21. [21]

    Graphical modelling in continuous-time: consistency guarantees via Neural ODEs,

    A. Bellot et al., “Graphical modelling in continuous-time: consistency guarantees via Neural ODEs,” arXiv, 2021

    work page 2021

  22. [22]

    Interpretability and fairness evaluation of deep learning models on MIMIC- IV,

    X. Meng et al., “Interpretability and fairness evaluation of deep learning models on MIMIC- IV,” Scientific Reports, 2022. 18

    work page 2022

  23. [23]

    MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV,

    X. Meng et al., “MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV,” arXiv, 2021

    work page 2021

  24. [24]

    Predicting Patient Outcomes with Graph Representation Learning,

    E. Rocheteau et al., “Predicting Patient Outcomes with Graph Representation Learning,” arXiv, 2021

    work page 2021

  25. [25]

    CiRLExplainer: Causality-Inspired Explainer for Graph Neural Networks via Reinforcement Learning,

    W. Hu, J. Wu, and Q. Qian, “CiRLExplainer: Causality-Inspired Explainer for Graph Neural Networks via Reinforcement Learning,” IEEE Trans. Neural Netw. Learn. Syst., 2025

    work page 2025

  26. [26]

    Generating In-Distribution Counterfactual Explanation for GNNs,

    Y. Chen et al., “Generating In-Distribution Counterfactual Explanation for GNNs,” Proc. AAAI, 2026

    work page 2026

  27. [27]

    CIDER: Counterfactual-Invariant Diffusion-based GNN Explainer,

    Y. Zhang et al., “CIDER: Counterfactual-Invariant Diffusion-based GNN Explainer,” arXiv, 2024

    work page 2024

  28. [28]

    VitalTCN: Enhancing ICU Mortality Prediction Through Temporal Convolutional Networks,

    T. Sk and M. Sundari, “VitalTCN: Enhancing ICU Mortality Prediction Through Temporal Convolutional Networks,” Proc. ICIMA, 2025

    work page 2025

  29. [29]

    Learning and Evaluating GNN Explanations based on Counterfactual and Factual Reasoning,

    S. Tan et al., “Learning and Evaluating GNN Explanations based on Counterfactual and Factual Reasoning,” Proc. WebConf, 2022

    work page 2022

  30. [30]

    Disease Risk Prediction via Heterogeneous Graph Attention Networks,

    L. Qu et al., “Disease Risk Prediction via Heterogeneous Graph Attention Networks,” Proc. IEEE BIBM, 2022

    work page 2022

  31. [31]

    DHGL: Dynamic hypergraph-based deep learning model for disease prediction,

    L. Qu et al., “DHGL: Dynamic hypergraph-based deep learning model for disease prediction,” Electronics Letters, 2024

    work page 2024

  32. [32]

    Recurrent Neural Networks for Multivariate Time Series with Missing Values,

    Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y. Liu, “Recurrent Neural Networks for Multivariate Time Series with Missing Values,” Scientific Reports, vol. 8, no. 1, p. 6085, 2018

    work page 2018

  33. [33]

    Multitask learning and benchmarking with clinical time series data,

    H. Harutyunyan, H. Khachatrian, D. C. Kale, G. Ver Steeg, and A. Galstyan, “Multitask learning and benchmarking with clinical time series data,” Scientific Data, vol. 6, no. 1, p. 96, 2019. 19

    work page 2019