Pith. sign in

REVIEW 4 major objections 4 minor 61 references

Causally-informed Deep Learning towards Explainable and Generalizable Outcomes Prediction in Critical Care

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read cDEEP predicts six critical-care outcomes from only their discovered direct causes, gaining explainability and out-of-distribution stability in one move.

desk verdict cDEEP is a serious, well-built ICU prediction system with a useful interpretation tool, but its central generalizability claim rests on a theorem whose intervention assumption doesn't match the age-split experiment. read the letter →

arxiv 2502.02109 v1 pith:ZDM76GNW submitted 2025-02-04 cs.LG cs.AI

Yuxiao Cheng , Xinxin Song , Ziqian Wang , Qin Zhong , Kunlun He , Jinli Suo This is my paper · ORCID
classification cs.LGcs.AI
keywords causaldiscoveryearlywarningscorecriticalcaregeneralizabilityinterpretabilityelectronichealthrecordstime-seriespredictiondeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

cDEEP is an ICU early-warning model that first learns a causal graph from electronic health records and then predicts each of six deteriorations—acute kidney injury, ARDS, circulatory failure, death, delirium, and sepsis—using only the variables the graph marks as direct causes of that outcome. The paper's central claim is that this restricted input set is what makes the model both interpretable and generalizable: causal parents stay stable under interventions on other variables, so predictions survive shifts in patient population and measurement practice. On MIMIC-IV and eICU, the model matches or exceeds generalizable-AI baselines on out-of-distribution splits by age and admission time while using far fewer inputs. If this holds, early warning systems could be deployed across hospitals with transparent reasoning for each risk score and without retraining on every new population.

What carries the argument

The load-bearing object is a learned causal probability matrix that is relaxed to continuous values and optimized alternately with an encoder-decoder prediction network. The graph is split into a variable-to-outcome (V2O) part, which selects each outcome's direct causes, and a variable-to-variable (V2V) part, which traces mediating pathways; a cumulative window graph downweights long time lags, and Gumbel-Softmax makes the binary parent sampling differentiable. The V2O parent set is what gets fed to the decoders, what supports the worst-case optimality theorem, and what cuts the average number of input variables from 39 to 18.5 and total input features by 80.1%. Controlled direct effects computed on this graph convert the learned structure into per-variable, per-patient quantitative explanations.

What would settle it

Learn the V2O parent sets separately for patients aged 75 and under and 76 and over; if the parent sets diverge, or if on a split defined by a documented treatment-protocol change—a genuine intervention on a non-parent variable—cDEEP fails to beat all-variable baselines, the theorem's premise is unmet and the claimed guarantee collapses.

Watch

Extended reading notes

Core claim

The central discovery is that a prediction model built on the direct causal parents of each outcome—rather than on all measured variables—retains accuracy while gaining a formal worst-case generalization guarantee. The paper proves, under the assumption that the discovered graph is the true causal graph, that the function predicting an outcome from its causal parents is optimal in the worst case over all distributions reachable by interventions, a result it notes is equivalent to the invariance theorem for causal transfer learning. Experimentally, age-based and time-based out-of-distribution splits of two large ICU databases show that cDEEP with causal variables only is competitive with or better than IRM, GroupDRO, VREx, and dropout baselines, and that cDEEP-full, which still uses all variables under the same training procedure, performs even better in most cases. The same causal graph also supplies the explanation: variable-to-variable edges plus variable-to-outcome edges give explicit pathways, and controlled direct effects quantify each variable's contribution.

Load-bearing premise

The entire generalization guarantee rests on the assumption that the graph learned from the training patients is the true causal graph and that the new patient group differs only through interventions on variables that are not direct causes; the experiments' age-based split is not shown to satisfy that condition.

Editorial extensions

If this is right

  • If only direct causes are used, predictions should remain stable when hospitals change how or whether they measure non-causal variables, easing deployment across institutions.
  • The same trained encoder with outcome-specific decoders can serve all six outcomes, each with its own parent set, so adding a new outcome reduces to learning and attaching its causal parents.
  • Clinicians can inspect the V2V pathways and CDE values to identify actionable variables upstream of a difficult-to-treat direct cause.
  • The 80.1% reduction in input features lowers the cost of data collection and speeds up full quantitative interpretation of a prediction.
  • The formal equivalence to invariant causal transfer gives a guarantee that no model using all measured variables can improve worst-case performance under interventions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A sharper test of the theorem than the age split would be an admission-time or site split that coincides with a documented change in measurement or treatment protocol, since that is a genuine intervention on a non-parent variable.
  • If the discovered parent sets prove stable across cohorts, the causal graph itself becomes a reusable asset: it could be transferred to a new hospital, or its V2V edges could seed new hypotheses for clinical studies.
  • The same recipe—learn parents, then predict from parents only—could generalize to other high-stakes medical predictions such as vasopressor dosing or unplanned readmission, where explaining the model and trusting it across sites are both bottlenecks.
  • Adding new variable types such as medications or imaging might change the learned parent sets, because the Markov boundary of an outcome depends on the measured variable set; checking this would show how complete the current graph is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes cDEEP, a deep learning early-warning system that jointly learns a causal graph (variable-to-outcome and variable-to-variable) and a set of outcome predictors. The model is trained on pooled MIMIC-IV and eICU data to predict six ICU outcomes, and only variables selected as direct causal parents are used as inputs at test time. The authors claim that this causal selection yields interpretable causal pathway explanations and improved out-of-distribution generalizability, and they evaluate this on age-based and admission-time-based split test sets against several OOD generalization baselines.

Significance. If the central claims were fully established, the paper would be a valuable contribution: it couples causal discovery with outcome prediction on a large EHR corpus, includes comparisons with IRM, VREx, and GroupDRO, introduces a causal-probability-matrix formulation with Gumbel-Softmax relaxation, and provides a publicly accessible visualization tool. The paper also correctly identifies the theoretical connection to Rojas-Carulla et al.'s invariant causal prediction results. However, the as-stated theoretical guarantee does not cover the evaluated OOD splits, the discovered graph is not validated as a true causal graph, and the empirical ablation with cDEEP-full points away from causal selection as the driver of OOD performance. These issues are load-bearing for the central claim, so the manuscript needs substantial revision.

major comments (4)
  1. [Methods, "Generalizability" (Theorem 1, Eq. 15-16); Results, "Generalizability"; Extended Data Fig. 1] The theoretical foundation in Theorem 1 does not apply to the age-based OOD evaluation. The theorem, following Rojas-Carulla et al., concerns distributions generated by interventions on variables outside the parent set Pa(Y). The OOD test set is defined by conditioning on Age ≥ 76, which is a selection on a static covariate that is itself an input and plausibly a parent of several outcomes. Selection is not a do-intervention, and the paper does not demonstrate that P(Y | Pa(Y)) or the parent set is invariant across age bands. Moreover, the theorem as stated defines P as "the set of all possible distributions of (X_T, Y_T)"; under that unrestricted set the minimax claim is false, because an adversary can alter P(Y | X) arbitrarily. The statement should be corrected to the intervention-restricted class from Rojas-Carulla et al., and an additional experiment should test invariance of the conditional outcome distribution given the estimated parents across age groups.
  2. [Methods, "Causal discovery" (Eq. 4-8) and "Granger causality"; Theorem 1 assumption] Theorem 1 assumes that G_v2o is the true causal graph, but the graph is estimated on the same training data used to fit the predictor, using a Granger-style prediction-sensitivity objective with a sparsity regularizer. Granger causality is not structural causality: a variable that improves prediction need not be a cause, and a true cause whose effect is mediated by another observed variable may not improve prediction once the mediator is included. The paper provides no independent validation that the selected parents are the true causal parents (e.g., no synthetic ground-truth experiment, no comparison with a known graph, no sensitivity analysis of the selected edge set). Without such validation, the theoretical guarantee cannot be invoked for the learned graph, and the interpretability claims are also only as strong as the graph's correctness.
  3. [Methods, "Interpretability" (Eq. 12; Fig. 2)] Eq. (12) labels the difference f_theta(x) - f_theta(x') as the "controlled direct effect" (CDE), but a CDE requires fixing mediators to a constant while intervening on the exposure. The quantity in Eq. (12) simply perturbs x_i and re-runs the network; it does not hold any intermediate variables fixed. Unless the network is a structural equation model with explicit mediator control, this quantity should not be called a CDE. The manuscript's interpretation section (Fig. 2 and the web tool) relies on this terminology, so the naming is misleading for clinical users. The authors should either implement a proper CDE computation using the V2V graph to identify and fix mediators, or rename the quantity and soften the causal interpretation.
  4. [Results, "Generalizability" (Fig. 3b-c)] The ablation results do not support the claim that causal variable selection is the source of OOD robustness. The text states that cDEEP-full, which uses all variables, is "even better than cDEEP" on OOD data (Fig. 3b-c). If selecting only direct causal parents were the mechanism behind the OOD advantage, cDEEP should outperform cDEEP-full under distribution shift. The observed pattern is consistent with the alternative explanation that the architecture, training procedure, or data-scale effects, rather than causal selection, drive the gains. The paper's remark that this "further validates the superiority of cDEEP" is internally inconsistent. The authors need to either provide an analysis showing how cDEEP-full also benefits from causal regularization (e.g., because the graph is used in its training loss), or re-frame the empirical generalizability claim as being about the full model rather than about causal input selection.
minor comments (4)
  1. [Extended Data Fig. 1; Table 1] The patient-count labels in Extended Data Fig. 1 are hard to reconcile with Table 1; please clarify which numbers refer to which database and subset.
  2. [Methods, "Theorem 1" notation] The notation in Theorem 1 is inconsistent: Eq. (15) defines f* as a conditional expectation, but the minimization in Eq. (16) is over functions f in C0 and uses X_T, Y_T without defining the subscript/superscript distinction; please define all symbols unambiguously.
  3. [Results, "High prediction accuracy" (calibration paragraph)] The calibration discussion reports that after isotonic regression the Brier score decreases by 0.0004; please state the baseline Brier score, the calibration method used, and how many samples were used for calibration.
  4. [Methods, "Time-series construction"] The paper reports an average of 74.42 prediction points per patient; it would be helpful to report the standard deviation or range, since the number of samples per patient varies widely and may affect the loss weighting.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: Theorem 1 is externally attributed, the OOD benchmarks are held-out and independent of graph fitting, and the main weaknesses are unverified assumptions rather than self-referential reductions.

full rationale

The paper's derivation chain is not circular in the sense of predicting its own fitted inputs. The generalizability claim rests on Theorem 1, which the paper explicitly attributes to Rojas-Carulla et al. (ref. 51) and states with the honest premise "Assuming that the causal graph Gv2o is the true causal graph"; this is an external mathematical result, not a self-citation. The empirical support is evaluated on held-out in-distribution patients and on out-of-distribution splits by age and admission time, and the OOD labels are not used in selecting the graph or fitting the predictor, so the reported AUROC/AUPRC gains are not statistically forced by construction. The self-citations to CUTS and CUTS+ (refs. 26-27) appear as background for neural causal discovery and are not the sole load-bearing justification; the paper supplies its own equations for the causal probability matrix, Gumbel-Softmax relaxation, cumulative window graphs, and alternating optimization. The real weaknesses are validity gaps rather than circularity: the discovered graph is not shown to be the true causal graph, the age-based split is a selection on a static covariate rather than an intervention on non-parent variables, and Eq. (12) labels a perturbation difference as a controlled direct effect without adjusting for mediators. These concerns affect whether Theorem 1's conclusion applies to the reported experiments, but they do not make any prediction equivalent to its input by definition, so no circular step is identified.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The causal graph, causal probability matrix, and CDE values are modeling constructs built from the data; they are presented as discovered truths, but they lack independent causal validation. The free parameters that most affect the central claim are the learned causal graph parameters and the unstated pruning and temperature choices.

free parameters (4)
  • Causal probability matrix parameters Q_v2o and Q_v2v = not reported
    Learned from training data to select causal parents for each outcome and for variable-to-variable relations; these selections drive the generalizability and interpretation claims.
  • Gumbel-Softmax temperature tau = not reported
    Controls the hardness of the causal parent sampling in Eq. 8 and is chosen by hand; not reported.
  • Causal graph pruning threshold for visualization = not reported
    Extended Data Fig. 3 shows only the most contributive causal relationships, but the threshold is not specified, so the displayed causal pathways are not fully reproducible.
  • Time window and chunking design = 168 time points, 2-hour slices, 14 chunks
    The 14-day window, 2-hour binning, and cumulative chunking in Eq. 9 are design choices that determine temporal resolution and the size of the causal probability matrix.
assumptions (6)
  • domain assumption G_v2o is the true causal graph
    Theorem 1 explicitly assumes this, and the paper does not validate the discovered graph against ground-truth causal benchmarks or known interventions.
  • domain assumption Distribution shifts in deployment are interventions that leave outcome parents unchanged
    Used to connect Theorem 1 to the age-based OOD split; the split is a selection on a static covariate, and the assumption is not tested.
  • domain assumption Outcomes cannot cause input variables or future outcomes
    The causal graph is split into V2O and V2V with empty outcome-to-variable edges; temporally plausible but not verified against alternative structures.
  • ad hoc to paper Nearer time points have stronger causal effects
    The cumulative window graph in Eq. 9 incorporates this prior to penalize longer time lags; it is a modeling choice, not derived from data.
  • domain assumption Granger-style predictive sensitivity corresponds to structural causation
    The V2O and V2V graphs are optimized to minimize prediction loss, so selected parents are predictive features under the model, not proven structural causes.
  • domain assumption Missing-value indicators do not bias causal discovery
    The paper states this is supported by Supplements Fig. S1, but that supplement is not included in the reviewed version.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causally-informed Deep Learning towards Explainable and Generalizable Outcomes Prediction in Critical Care." pith.science (2026). https://pith.science/paper/ZDM76GNW

@misc{pith2026250202109,
  author       = {Pith},
  title        = {Pith review of: Causally-informed Deep Learning towards Explainable and Generalizable Outcomes Prediction in Critical Care},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDM76GNW}},
  note         = {Machine review of arXiv:2502.02109}
}
read the original abstract

Recent advances in deep learning (DL) have prompted the development of high-performing early warning score (EWS) systems, predicting clinical deteriorations such as acute kidney injury, acute myocardial infarction, or circulatory failure. DL models have proven to be powerful tools for various tasks but come with the cost of lacking interpretability and limited generalizability, hindering their clinical applications. To develop a practical EWS system applicable to various outcomes, we propose causally-informed explainable early prediction model, which leverages causal discovery to identify the underlying causal relationships of prediction and thus owns two unique advantages: demonstrating the explicit interpretation of the prediction while exhibiting decent performance when applied to unfamiliar environments. Benefiting from these features, our approach achieves superior accuracy for 6 different critical deteriorations and achieves better generalizability across different patient groups, compared to various baseline algorithms. Besides, we provide explicit causal pathways to serve as references for assistant clinical diagnosis and potential interventions. The proposed approach enhances the practical application of deep learning in various medical scenarios.

Figures

Figures reproduced from arXiv: 2502.02109 by the authors.

Figure 1
Figure 1. The overall architecture of the proposed approach. a, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Demonstration of cDEEP’s interpretable predictions. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. cDEEP’s advantageous performance on out-of-distribution testing datasets. a, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Acceleration for quantitative interpretation. a, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 42 canonical work pages

  1. [1]

    Prosperi, M. et al. Causal inference and counterfactual predic- tion in machine learning for actionable healthcare. Nat. Mach. Intell. 2, 369–375, DOI: 10.1038/s42256-020-0197-y (2020)

    doi:10.1038/s42256-020-0197-y 2020

  2. [2]

    Machine learning and artificial intelligence re- search for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness

    V ollmer, S.et al. Machine learning and artificial intelligence re- search for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness. BMJ 368, l6927, DOI: 10.1136/bmj.l6927 (2020)

    work page doi:10.1136/bmj.l6927 2020

  3. [3]

    & Clifton, D

    Shamout, F., Zhu, T. & Clifton, D. A. Machine Learning for Clinical Outcome Prediction. IEEE Rev. Biomed. Eng. 14, 116–126, DOI: 10.1109/RBME.2020.3007816 (2021)

    work page arXiv 2021

  4. [4]

    & Xia, J

    Yang, G., Ye, Q. & Xia, J. Unbox the black-box for the medical explainable AI via multi-modal and multi-centre data fusion: A mini-review, two showcases and beyond.Inf. Fusion 77, 29–52, DOI: 10.1016/j.inffus.2021.07.016 (2022)

    doi:10.1016/j.inffus.2021.07.016 2022

  5. [5]

    Lauritsen, S. M. et al. Explainable artificial intelligence model to predict acute critical illness from electronic health records. Nat. Commun. 11, 3852, DOI: 10.1038/s41467-020-17431-x (2020)

    work page doi:10.1038/s41467-020-17431-x 2020

  6. [6]

    & Gutierrez, J

    Caicedo-Torres, W. & Gutierrez, J. ISeeU: Visually inter- pretable deep learning for mortality prediction inside the ICU. J. Biomed. Informatics 98, 103269, DOI: 10.1016/j.jbi.2019. 103269 (2019)

    doi:10.1016/j.jbi.2019 2019

  7. [7]

    Steyaert, S. et al. Multimodal data fusion for cancer biomarker discovery with deep learning. Nat. Mach. Intell. 1–12, DOI: 10.1038/s42256-023-00633-5 (2023)

    doi:10.1038/s42256-023-00633-5 2023

  8. [8]

    & Jurafsky, D

    Ethayarajh, K. & Jurafsky, D. Attention Flows are Shapley Value Explanations. In Zong, C., Xia, F., Li, W. & Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Associa- tion for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), 49–54, DOI: 10.18653/v1/2021.acl-short...

    work page doi:10.18653/v1/2021.acl-short.8

Show all 61 references
  1. [9]

    Yan, L. et al. An interpretable mortality prediction model for COVID-19 patients. Nat. Mach. Intell. 2, 283–288, DOI: 10.1038/s42256-020-0180-7 (2020)

    2020 doi

  2. [10]

    & Salakhutdinov, R

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res. 15, 1929– 1958 (2014)

    2014

  3. [11]

    & Hertz, J

    Krogh, A. & Hertz, J. A Simple Weight Decay Can Improve Generalization. In Advances in Neural Information Processing Systems, vol. 4 (Morgan-Kaufmann, 1991)

    1991

  4. [12]

    & Lopez-Paz, D

    Arjovsky, M., Bottou, L., Gulrajani, I. & Lopez-Paz, D. In- variant Risk Minimization, DOI: 10.48550/arXiv.1907.02893 (2020). 1907.02893

    2020 doi arXiv

  5. [13]

    W., Hashimoto, T

    Sagawa*, S., Koh*, P. W., Hashimoto, T. B. & Liang, P. Distri- butionally Robust Neural Networks. In International Confer- ence on Learning Representations (2019)

    2019

  6. [14]

    Krueger, D. et al. Out-of-Distribution Generalization via Risk Extrapolation (REx). In Proceedings of the 38th International Conference on Machine Learning, 5815–5826 (PMLR, 2021)

    2021

  7. [15]

    Thorsen-Meyer, H.-C. et al. Dynamic and explainable machine learning prediction of mortality in patients in the intensive care unit: A retrospective study of high-frequency data in electronic patient records. The Lancet Digit. Heal. 2, e179–e191, DOI: 10.1016/S2589-7500(20)300...

    2020 doi

  8. [16]

    Hyland, S. L. et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat. Medicine 26, 364–373, DOI: 10.1038/s41591-020-0789-4 (2020)

    2020 doi

  9. [17]

    Bates, D. W. & Syrowatka, A. Harnessing AI in sepsis care.Nat. Medicine 28, 1351–1352, DOI: 10.1038/s41591-022-01878-0 (2022)

    2022 doi

  10. [18]

    Demirjian, S. et al. Predictive accuracy of a perioperative laboratory test–based prediction model for moderate to severe acute kidney injury after cardiac surgery. JAMA 327, 956–964, DOI: 10.1001/jama.2022.1751 (2022)

    2022

  11. [19]

    Than, M. P. et al. Machine Learning to Predict the Likelihood of Acute Myocardial Infarction. Circulation 140, 899–909, DOI: 10.1161/CIRCULATIONAHA.119.041980 (2019)

    2019 doi

  12. [20]

    C., Walker, I

    Castro, D. C., Walker, I. & Glocker, B. Causality matters in medical imaging. Nat. Commun. 11, 3673, DOI: 10.1038/ s41467-020-17478-w (2020)

    2020

  13. [21]

    & Camps- Valls, G

    Runge, J., Gerhardus, A., Varando, G., Eyring, V . & Camps- Valls, G. Causal inference for time series. Nat. Rev. Earth & Environ. 4, 487–505, DOI: 10.1038/s43017-023-00431-y (2023)

    2023 doi

  14. [22]

    Causality: Models, Reasoning and Inference (Cam- bridge University Press, USA, 2009), 2nd edn

    Pearl, J. Causality: Models, Reasoning and Inference (Cam- bridge University Press, USA, 2009), 2nd edn

    2009

  15. [23]

    J., Camgoz, N

    V owels, M. J., Camgoz, N. C. & Bowden, R. D’ya like dags? A survey on structure learning and causal discovery, DOI: 10. 48550/arXiv.2103.02582 (2021). 2103.02582

    2021 doi arXiv

  16. [24]

    & Kunag, K

    Shen, Z., Cui, P., Zhang, T. & Kunag, K. Stable Learning via Sample Reweighting. Proc. AAAI Conf. on Artif. Intell. 34, 5692–5699, DOI: 10.1609/aaai.v34i04.6024 (2020)

    2020 doi

  17. [25]

    Moraffah, R. et al. Causal inference for time series analysis: Problems, methods and evaluation. Knowl. Inf. Syst. 63, 3041– 3085, DOI: 10.1007/s10115-021-01621-0 (2021)

    2021 doi

  18. [26]

    CUTS: Neural Causal Discovery from Irregular Time-Series Data

    Cheng, Y .et al. CUTS: Neural Causal Discovery from Irregular Time-Series Data. In The Eleventh International Conference on Learning Representations (2023)

    2023

  19. [27]

    CUTS+: High-Dimensional Causal Discovery from Irregular Time-Series

    Cheng, Y .et al. CUTS+: High-Dimensional Causal Discovery from Irregular Time-Series. Proc. AAAI Conf. on Artif. Intell. 38, 11525–11533, DOI: 10.1609/aaai.v38i10.29034 (2024)

    2024 doi

  20. [28]

    & Mukherjee, S

    Lagemann, K., Lagemann, C., Taschler, B. & Mukherjee, S. Deep learning of causal structures in high dimensions under data limitations. Nat. Mach. Intell. 5, 1306–1316, DOI: 10. 1038/s42256-023-00744-z (2023)

    2023

  21. [29]

    Johnson, A. E. W. et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci. Data 10, 1, DOI: 10.1038/s41597-022-01899-x (2023)

    2023 doi

  22. [30]

    Pollard, T. J. et al. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci. Data 5, 180178, DOI: 10.1038/sdata.2018.178 (2018)

    2018 doi

  23. [31]

    Kidney Int

    KDIGO. Kidney Int. Suppl. 2, 1, DOI: 10.1038/kisup.2012.1 (2012)

    2012 doi

  24. [32]

    & Ciapponi, A

    Santa Cruz, R., Villarejo, F., Irrazabal, C. & Ciapponi, A. High versus low positive end-expiratory pressure (PEEP) levels for mechanically ventilated adult patients with acute lung injury and acute respiratory distress syndrome. Cochrane Database Syst. Rev. (2021)

    2021

  25. [33]

    P., Anderson, M

    Basile, D. P., Anderson, M. D. & Sutton, T. A. Pathophysiology of Acute Kidney Injury. Compr. Physiol. 2, 1303–1353, DOI: 10.1002/cphy.c110041 (2012)

    2012 doi

  26. [34]

    Matuschak, G. M. & Rinaldo, J. E. Organ interactions in the adult respiratory distress syndrome during sepsis: Role of the liver in host defense. Chest 94, 400–406 (1988)

    1988

  27. [35]

    Herrero, R. et al. Liver–lung interactions in acute respiratory distress syndrome. Intensive Care Medicine Exp. 8, 48, DOI: 10.1186/s40635-020-00337-9 (2020)

    2020 doi

  28. [36]

    Trillo-Alvarez, C. et al. Acute lung injury prediction score: Derivation and validation in a population-based sample. Eur. Respir. J. 37, 604–609, DOI: 10.1183/09031936.00036810 (2011)

    2011

  29. [37]

    Writing Group for the Alveolar Recruitment for Acute Respi- ratory Distress Syndrome Trial (ART) Investigators. Effect of Lung Recruitment and Titrated Positive End-Expiratory Pressure (PEEP) vs Low PEEP on Mortality in Patients With Acute Respiratory Distress Syndrome: A Rand...

    2017

  30. [38]

    Riviello, E. D. et al. Hospital Incidence and Outcomes of the Acute Respiratory Distress Syndrome Using the Kigali Modi- fication of the Berlin Definition. Am. J. Respir. Critical Care Medicine 193, 52–59, DOI: 10.1164/rccm.201503-0584OC (2016)

    2016 doi

  31. [39]

    D., Matthay, M

    Wick, K. D., Matthay, M. A. & Ware, L. B. Pulse oximetry for the diagnosis and management of acute respiratory distress syndrome. The Lancet Respir. Medicine 10, 1086–1098, DOI: 10.1016/S2213-2600(22)00058-3 (2022)

    2022 doi

  32. [40]

    Morris, C. G. & Low, J. Metabolic acidosis in the critically ill: Part 2. Causes and treatment. Anaesthesia 63, 396–411, DOI: 10.1111/j.1365-2044.2007.05371.x (2008)

    2008

  33. [41]

    E., Monnet, X

    Marik, P. E., Monnet, X. & Teboul, J.-L. Hemodynamic param- eters to guide fluid therapy. Annals Intensive Care 1, 1, DOI: 10.1186/2110-5820-1-1 (2011)

    2011 doi

  34. [42]

    Li, D. et al. Immune dysfunction leads to mortality and organ injury in patients with COVID-19 in China: Insights from ERS- COVID-19 study. Signal Transduct. Target. Ther.5, 1–3, DOI: 10.1038/s41392-020-0163-5 (2020)

    2020 doi

  35. [43]

    Adrie, C. et al. Persistent lymphopenia is a risk factor for ICU- acquired infections and for death in ICU patients with sustained hypotension at admission. Annals Intensive Care 7, 30, DOI: 10.1186/s13613-017-0242-0 (2017)

    2017 doi

  36. [44]

    The effect of increased positive end expiratory pressure on brain tissue oxygenation and intracranial pressure in acute brain injury patients

    Gouvea Bogossian, E.et al. The effect of increased positive end expiratory pressure on brain tissue oxygenation and intracranial pressure in acute brain injury patients. Sci. Reports 13, 16657, DOI: 10.1038/s41598-023-43703-9 (2023)

    2023 doi

  37. [45]

    & Karanikolas, M

    Velissaris, D., Karamouzos, V ., Pierrakos, C., Aretha, D. & Karanikolas, M. Hypomagnesemia in Critically Ill Sepsis Pa- tients. J. Clin. Medicine Res. 7, 911–918, DOI: 10.14740/ jocmr2351w (2015)

    2015

  38. [46]

    Aliyu, M. et al. Interleukin-6 cytokine: An overview of the immune regulation, immune dysregulation, and therapeutic approach. Int. Immunopharmacol. 111, 109130, DOI: 10.1016/ j.intimp.2022.109130 (2022)

    2022

  39. [47]

    G., Amarasingham, R., Shah, A., Xie, B

    Cohen, I. G., Amarasingham, R., Shah, A., Xie, B. & Lo, B. The Legal And Ethical Concerns That Arise From Using Complex Predictive Analytics In Health Care. Heal. Aff. DOI: 10.1377/hlthaff.2014.0048 (2017)

    2017

  40. [48]

    Acute Respiratory Distress Syndrome: The Berlin Definition

    The ARDS Definition Task Force*. Acute Respiratory Distress Syndrome: The Berlin Definition. JAMA 307, 2526–2533, DOI: 10.1001/jama.2012.5669 (2012). 12/16

    2012

  41. [49]

    Gong, C., Yao, D., Zhang, C., Li, W. & Bi, J. Causal Discov- ery from Temporal Data: An Overview and New Perspectives (2023). 2303.10112

    2023 arXiv

  42. [50]

    K., Devijver, E

    Assaad, C. K., Devijver, E. & Gaussier, E. Survey and evalua- tion of causal discovery methods for time series. J. Artif. Intell. Res. 73, 767–819, DOI: 10.1613/jair.1.13428 (2022)

    2022 doi

  43. [51]

    & Peters, J

    Rojas-Carulla, M., Scholkopf, B., Turner, R. & Peters, J. In- variant Models for Causal Transfer Learning. J. Mach. Learn. Res. (2018)

    2018

  44. [52]

    Granger, C. W. J. Investigating causal relations by econometric models and cross-spectral methods. Econometrica 37, 424–438, DOI: 10.2307/1912791 (1969). 1912791

    1969 doi

  45. [53]

    & Stramaglia, S

    Marinazzo, D., Pellicoro, M. & Stramaglia, S. Kernel-granger causality and the analysis of dynamical networks. Phys. review E 77, 056215 (2008)

    2008

  46. [54]

    & Fox, E

    Tank, A., Covert, I., Foti, N., Shojaie, A. & Fox, E. B. Neural granger causality. IEEE Transactions on Pattern Analysis Mach. Intell. 44, 4267–4279, DOI: 10.1109/TPAMI.2021.3065601 (2022)

    2022

  47. [55]

    & Welling, M

    L¨owe, S., Madras, D., Zemel, R. & Welling, M. Amortized causal discovery: Learning to infer causal graphs from time- series data. In Proceedings of the First Conference on Causal Learning and Reasoning, 509–525 (PMLR, 2022)

    2022

  48. [56]

    & Karlen, W

    Schwab, P. & Karlen, W. CXPlain: Causal Explanations for Model Interpretation under Uncertainty. In Advances in Neural Information Processing Systems, vol. 32 (Curran Associates, Inc., 2019)

    2019

  49. [57]

    C., Abe, N., Liu, Y

    Lozano, A. C., Abe, N., Liu, Y . & Rosset, S. Grouped graphical Granger modeling methods for temporal causal modeling. In Proceedings of the 15th ACM SIGKDD International Confer- ence on Knowledge Discovery and Data Mining , KDD ’09, 577–586, DOI: 10.1145/1557019.1557085 (Asso...

  50. [58]

    & Michailidis, G

    Basu, S. & Michailidis, G. Regularized estimation in sparse high-dimensional time series models. The Annals Stat. 43, 1535–1567, DOI: 10.1214/15-AOS1315 (2015)

    2015 doi

  51. [59]

    & Poole, B

    Jang, E., Gu, S. & Poole, B. Categorical Reparameteriza- tion with Gumbel-Softmax, DOI: 10.48550/arXiv.1611.01144 (2017). 1611.01144

    2017 doi arXiv

  52. [60]

    & Ruttenberg, B

    Harradon, M., Druce, J. & Ruttenberg, B. Causal Learning and Explanation of Deep Neural Networks via Autoencoded Activations, DOI: 10.48550/arXiv.1802.00541 (2018). 1802. 00541

    2018 doi arXiv

  53. [61]

    E., Gnecco, N

    Christiansen, R., Pfister, N., Jakobsen, M. E., Gnecco, N. & Peters, J. A causal framework for distribution generalization, DOI: 10.48550/arXiv.2006.07433 (2021). 2006.07433. Acknowledgement This work was supported by the National Natural Science Foun- dation of China (Nos. 61...

    2021 doi arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.