Pith. sign in

REVIEW 4 major objections 4 minor 68 references

End to End Autoencoder MLP Framework for Sepsis Prediction

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An end-to-end autoencoder–MLP network predicts sepsis better than traditional machine-learning baselines on three ICU cohorts.

desk verdict A standard bottleneck MLP is dressed as an autoencoder, with a likely test-set tuning leak; the clinical problem is real but this paper adds little beyond a routine benchmark. read the letter →

arxiv 2508.18688 v2 pith:22CIB7O6 submitted 2025-08-26 cs.LG

Hejiang Cai , Di Wu , Ji Xu , Xiang Liu , Yiziting Zhu , Xin Shu , Yujie Li , Bin Yi This is my paper
classification cs.LG
keywords sepsispredictionautoencodermultilayerperceptronend-to-endlearningelectronichealthrecordsmissingdatatimeseriesICUearlywarning
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

The paper claims that a single end-to-end neural network—an autoencoder that compresses hourly vital-sign and lab vectors into a 16-dimensional code, feeding a multilayer perceptron classifier—predicts sepsis better than tuned traditional machine-learning baselines on three ICU cohorts. Instead of hand-crafted features, the network learns its own representation directly from forward-filled, zero-filled vectors with explicit missingness indicators, and a custom 80% completeness rule decides which time segments to keep. The reported head-to-head results are 74.6%, 80.6%, and 93.5% accuracy on the three cohorts, with gains over gradient boosting concentrated in positive predictive value and specificity. If these numbers reproduce, the framework offers a practical real-time sepsis alerting system that does not require per-hospital feature engineering.

What carries the argument

The load-bearing mechanism is an end-to-end autoencoder–MLP trained with binary cross-entropy, plus a completeness-gated segmentation rule. The encoder (two fully connected layers of 32 and 16 units, ReLU and 50% dropout) creates a 16-dimensional code; the MLP head maps that code through 16→8 units to two logits. Around the network, two data-selection rules make the pipeline clinically usable: a custom down-sampling for training and a non-overlapping dynamic sliding window for inference, both triggered only when 80% of the feature set is present. This turns irregular, incomplete hourly EHR streams into fixed-size vectors while limiting alarm frequency.

What would settle it

Take the three cohorts, tune the learning rate and epoch count on the training split only, lock the model, and run it once on the held-out test split alongside gradient-boosted trees; the claim survives only if the reported accuracy and PPV margins reappear without test-set feedback.

Watch

Extended reading notes

Core claim

The central claim is that end-to-end co-training of an autoencoder and a small MLP classifier yields sepsis predictions that consistently beat traditional machine learning on the same ICU data. For each patient, hourly vital signs and labs are forward-filled and then segmented by an 80% feature-completeness rule: during training a customized down-sampling keeps the final complete vector of each qualifying segment, and during testing a non-overlapping dynamic sliding window emits exactly one prediction per qualifying segment. The chosen vector is Z-score normalized and zero-filled, with missingness marked, and passed through a 32→16 encoder to a 16-dimensional bottleneck, then a 16→8→2 MLP he

Load-bearing premise

The reported edge over the baselines assumes the learning rate and epoch count were chosen on training or validation data only; if the held-out test cohorts influenced those choices, the accuracy and PPV numbers are optimistic.

Editorial extensions

If this is right

  • A real-time sepsis alert can be produced directly from raw vitals and labs, with no manual feature engineering, once the 80% completeness threshold is met.
  • Because each qualifying segment yields one prediction, alarm frequency is naturally bounded; hospitals can tune the threshold to trade sensitivity against alert volume.
  • The reported pattern—larger gains in PPV and specificity than in sensitivity—means the framework's main practical benefit is fewer false alarms while keeping detection rates comparable to gradient boosting.
  • The statistical significance of the accuracy, PPV, NPV, and specificity differences supports using the model as a screening layer across ICUs with different measurement patterns.

Reading between the lines

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

  • The paper's own statistics show sensitivity differences are not significant, so the honest summary of the contribution is an alarm-reduction tool rather than a higher-detection tool; clinicians should expect similar true-positive counts with fewer false positives.
  • The 80% threshold and the one-prediction-per-segment rule create an implicit operating point on the sensitivity/specificity curve; sweeping that threshold across cohorts would reveal whether the reported superiority persists at other operating points.
  • The term 'unsupervised autoencoder' is used for the architecture, but training is purely supervised with classification loss; a reconstruction-based pretraining stage is an untested variant that the current design does not evaluate.
  • The three cohorts are drawn from two related public benchmark sources, so cross-institution generalization is claimed but not yet demonstrated on a wholly independent dataset; a blind external deployment would be the decisive test.
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 manuscript proposes an end-to-end deep learning framework for sepsis prediction that combines an autoencoder-based feature extractor with an MLP classifier. The preprocessing pipeline forward-fills missing values, extracts segments with at least 80% feature completeness using a customized down-sampling procedure during training and a non-overlapping dynamic sliding window during testing, and zero-fills remaining gaps. The framework is evaluated on three ICU cohorts (PhysioNet A, PhysioNet B, and FHC), reporting accuracies of 74.6%, 80.6%, and 93.5%, respectively, and claiming consistent improvement over Naive Bayes, SVM, Random Forest, and XGBoost baselines. The paper presents the autoencoder as unsupervised, but the only training objective given in Eq. (1) is a binary cross-entropy classification loss, with no reconstruction term or unsupervised pretraining stage.

Significance. If the reported results were obtained under an unbiased protocol, the proposed framework would be a useful, simple deep learning baseline for early sepsis prediction, and the segment-extraction and missingness-aware preprocessing ideas are worth reporting. The use of public datasets and the attempt at statistical testing (Friedman and Wilcoxon) are also positive features. However, the central methodological claim is not supported: the model is trained end-to-end with a classification loss only, so it is a bottleneck MLP rather than an autoencoder, and the novelty is substantially diminished. In addition, the hyperparameter selection data are not specified, and the small test sets with point-estimate-only reporting leave the main comparative claim insufficiently supported. The contribution is therefore more modest than presented, and the manuscript needs substantial revision to make its claims accurate and its evidence convincing.

major comments (4)
  1. [§2.2, Eq. (1)] The abstract, introduction, and Section 2.2 repeatedly describe the model as an 'unsupervised autoencoder' for automatic feature extraction. However, Eq. (1) defines only a binary cross-entropy classification loss over the entire encoder–MLP network; there is no reconstruction loss, no decoder, and no unsupervised pretraining. The bottleneck is trained purely by classification gradients. This invalidates the central methodological claim of unsupervised autoencoder-based feature learning. Please either add a genuine reconstruction objective (and rerun the experiments) or remove the autoencoder terminology and reframe the contribution as a bottleneck MLP; the title, abstract, and related claims must be updated accordingly.
  2. [§2.2] The grid search for learning rate and training epochs is described as selecting the combination that 'produce[s] the highest classification sensitivity and PPV,' but the data used for this selection are never specified. If the test sets described in Table 1 were used to choose hyperparameters, the performance comparisons in Table 2 become circular and the reported gains are optimistic. The authors must state clearly which split (training, validation, or test) was used for hyperparameter selection, and ideally use nested or repeated validation to avoid any potential leakage onto the test sets.
  3. [§3, Table 2] All results in Table 2 are point estimates without confidence intervals, error bars, or repeated-run variability. The test sets are small (457, 324, and 108 samples), so the reported 1–3 percentage-point accuracy improvements over baselines may be within sampling noise. Please report confidence intervals (e.g., bootstrap or exact binomial) for all metrics and, if feasible, repeated runs with different random seeds. Also, the text states that post hoc Wilcoxon analyses confirmed 'statistically significant pairwise improvements (all p < 0.05),' but no pairwise p-values are shown; provide those values and account for multiple comparisons.
  4. [§3, Statistical Analysis] The Friedman test for sensitivity is not significant (p = 0.2288), yet the surrounding text states that the findings 'substantiate' significant improvements across metrics. This is an overstatement. The manuscript should explicitly acknowledge the non-significant sensitivity result and temper the claim of consistent superiority, or provide additional evidence that the sensitivity differences are meaningful despite the global test result.
minor comments (4)
  1. [§2.1] The abstract and introduction promise 'explicit missingness indicators,' but the implementation is not described beyond zero-filling after forward-fill. Does the input vector concatenate a binary mask, or are missingness indicators implicitly represented by zeros? Please clarify the exact input representation.
  2. [§2.1] The FHC dataset is cited via reference [44] but is not described in the text. Please provide at least a brief description of the dataset and its recording protocol so that the three-cohort comparison is interpretable.
  3. [§2.1.2 and §2.2] The term 'non-overlapping dynamic sliding window' is potentially confusing because a sliding window usually implies overlap or continuous movement. Please clarify the mechanism, possibly renaming it to 'non-overlapping adaptive window' or similar.
  4. [References] Several references contain author-name errors or likely misattributions (e.g., reference [50] 'D. Celi, L. Ghassemi, F. Naumann' and reference [59] 'K. Baldi'). Please verify all references carefully, especially the large block of self-citations in the introduction that appear unrelated to sepsis prediction.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; one unresolved grid-search/validation-split risk prevents a clean 0

  1. other [Section 2.2 (hyperparameter grid search) and Table 2 (reported performance metrics)]
    "An extensive grid search was conducted to optimize the learning rate and training epochs of our end-to-end autoencoder–MLP framework. On the PhysioNet dataset, a learning rate of 7×10⁻⁴ with 550 epochs produced the highest classification sensitivity and positive predictive value (PPV)."

    The grid search criterion is literally the same pair of metrics (sensitivity and PPV) that Table 2 later reports as evidence of superiority. The paper never states that the grid search was confined to a held-out validation split or that the test sets were excluded from hyperparameter selection. As written, the possibility remains that the reported test-set sensitivities and PPVs are the very quantities that were maximized during model selection. If so, the reported 'predictions' reduce to fitted values: the model was chosen to make these numbers high, and the comparison against baselines is not an out-of-sample test. This is a missing-support circularity risk rather than a demonstrated equation-level reduction, because the text does not explicitly say the test set was used. But the omissio

full rationale

This is an empirical benchmark paper, not a mathematical derivation, so most circularity patterns (self-definitional equations, uniqueness theorems, ansatz smuggling) do not apply. The autoencoder–MLP architecture is a standard supervised MLP with a bottleneck; the term 'autoencoder' is a misnomer because no reconstruction loss is used, but that is a correctness/naming issue, not circularity. The down-sampling and sliding-window preprocessing are explicitly attributed to reference [44]; there is no evidence in the text that this is a self-citation, and even if it were, it is an adopted heuristic rather than a derived claim. The extensive self-citations in the reference list are background citations and do not carry the argument. The only significant circularity-adjacent concern is the hyperparameter grid search in Section 2.2: the objective metrics (sensitivity, PPV) are identical to the headline evaluation metrics in Table 2, and the paper never identifies the data split used for model selection. If the test sets were used for tuning, the reported accuracy advantages are fitted rather than predicted. However, the paper does not explicitly admit to test-set tuning, so this is an unresolved leakage risk rather than a proven circular step. A score of 2 reflects one potentially load-bearing omission, while acknowledging that the core empirical comparison would be independent if the grid search used a proper held-out split.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities or forces are introduced. The free parameters are all model hyperparameters and the preprocessing threshold, none of which are derived from theory. The axioms are standard machine learning background plus domain assumptions about the data and the ad hoc completeness threshold.

free parameters (5)
  • learning_rate = 7e-4 (PhysioNet), 1e-3 (FHC)
    Selected by grid search to maximize sensitivity and PPV; not fixed by theory.
  • training_epochs = 550 for both datasets
    Chosen by grid search along with learning rate; no early stopping described.
  • completeness_threshold = 80%
    Assumed to define high information density segments; taken from ref [44], no sensitivity analysis.
  • bottleneck_size = 16
    Architecture choice not justified by ablation or theory.
  • dropout_rate = 0.5
    Standard choice, not tuned or justified for this task.
assumptions (5)
  • standard math Cross-entropy loss and gradient descent are appropriate for optimizing the classifier.
    Invoked in Section 2.2 without proof; standard background.
  • domain assumption Sepsis labels in the PhysioNet and FHC cohorts are accurate and consistent.
    The paper does not describe label definitions or verify the FHC cohort; this is necessary for any performance claim.
  • domain assumption Forward-filling followed by zero-filling with missingness indicators adequately represents the EHR missingness process.
    Implied in Section 2.1; the paper does not evaluate alternative imputation strategies.
  • ad hoc to paper A segment is information-dense if at least 80 percent of features are present.
    This threshold is taken from prior work [44] and is assumed without analysis in this dataset context.
  • ad hoc to paper The bottleneck architecture yields useful representations for sepsis classification.
    No ablation comparing different depths or widths is provided; the architecture is declared without evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End to End Autoencoder MLP Framework for Sepsis Prediction." pith.science (2026). https://pith.science/paper/22CIB7O6

@misc{pith2026250818688,
  author       = {Pith},
  title        = {Pith review of: End to End Autoencoder MLP Framework for Sepsis Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22CIB7O6}},
  note         = {Machine review of arXiv:2508.18688}
}
read the original abstract

Sepsis is a life threatening condition that requires timely detection in intensive care settings. Traditional machine learning approaches, including Naive Bayes, Support Vector Machine (SVM), Random Forest, and XGBoost, often rely on manual feature engineering and struggle with irregular, incomplete time-series data commonly present in electronic health records. We introduce an end-to-end deep learning framework integrating an unsupervised autoencoder for automatic feature extraction with a multilayer perceptron classifier for binary sepsis risk prediction. To enhance clinical applicability, we implement a customized down sampling strategy that extracts high information density segments during training and a non-overlapping dynamic sliding window mechanism for real-time inference. Preprocessed time series data are represented as fixed dimension vectors with explicit missingness indicators, mitigating bias and noise. We validate our approach on three ICU cohorts. Our end-to-end model achieves accuracies of 74.6 percent, 80.6 percent, and 93.5 percent, respectively, consistently outperforming traditional machine learning baselines. These results demonstrate the framework's superior robustness, generalizability, and clinical utility for early sepsis detection across heterogeneous ICU environments.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 66 canonical work pages

  1. [44]

    Neural Netw

    Yuan, Y., Wang, Y., and Luo, X.: ‘A node -collaboration-informed graph convolutional network for highly-accurate representation to undirected weighted graph’, IEEE Trans. Neural Netw. Learn. Syst., 2024, 35, (3), pp. 3845–3858

    work page 2024

  2. [1]

    Automatica Sinica, 2021, 8, (4), pp

    Wu, D., Luo, X.: ‘Robust Latent Factor Analysis for Precise Representa- tion of High‑Dimensional and Sparse Data’, IEEE/CAA J. Automatica Sinica, 2021, 8, (4), pp. 796–805

    work page 2021

  3. [2]

    Remaining gaps are zero-filled to create fixed-size vectors marking miss- ingness

    Methodology To handle irregular, incomplete, and imbalanced ICU time-series, we propose an end -to-end framework: after for- ward-filling missing vitals and labs, we extract segments with ≥ 80 % completeness using customized down -sampling (training) and dynamic sliding windows (testing). Remaining gaps are zero-filled to create fixed-size vectors marking...

    work page 2019

  4. [3]

    Wu, D., Luo, X., Shang, M., et al.: ‘A Deep Latent Factor Model for High‑Dimensional and Sparse Matrices in Recommender Systems’, IEEE Trans. Syst. Man Cybern. Syst., 2021, 51, (7), pp. 4285–4296

    work page 2021

  5. [4]

    Wu, D., Luo, X., Shang, M., et al.: ‘A Data Characteristic Aware Latent Factor Model for Web Services QoS Prediction’, IEEE Trans. Knowl. Data Eng., 2022, 34, (6), pp. 2525–2538

    work page 2022

  6. [5]

    Neur al Netw

    Wu, D., Shang, M., Luo, X., et al.: ‘An L1 and L2‑norm Oriented Latent Factor Model for Recommender Systems’, IEEE Trans. Neur al Netw. Learn. Syst., 2022, 33, (10), pp. 5775–5788

    work page 2022

  7. [6]

    Wu, D., He, Y., Luo, X., et al.: ‘A Latent Factor Analysis‑Based Ap- proach to Online Sparse Streaming Feature Selection’, IEEE Trans. Syst. Man Cybern. Syst., 2022, 52, (11), pp. 6744–6758

    work page 2022

  8. [7]

    Big Data, 2024, 10, (1), pp

    Wu, D., Li, Z., Chen, F., et al.: ‘Online Sparse Streaming Feature Selec- tion With Gaussian Copula’, IEEE Trans. Big Data, 2024, 10, (1), pp. 92–107

    work page 2024

Show all 68 references
  1. [8]

    Neural Netw

    Wu, D., Luo, X., He, Y., et al.: ‘A Prediction‑Sampling‑Based Multi- layer‑Structured Latent Factor Model for Accurate Representation to High‑Dimensional and Sparse Data’, IEEE Trans. Neural Netw. Learn. Syst., 2024, 35, (3), pp. 3845–3858

    2024

  2. [9]

    Wu, D., Hu, Y., Liu, K., et al.: ‘An Outlier‑Resilient Autoencoder for Representing High‑Dimensional and Incomplete Data’, IEEE Trans. Emerg. Topics Comput., 2024

    2024

  3. [10]

    Chen, J., Wang, R., Wu, D., Luo, X.: ‘A Differential Evolution‑Enhanced Position‑Transitional Approach to Latent Factor Analysis’, IEEE Trans. Emerg. Topics Comput. Intell., 2023, 7, (2), pp. 389–401

    2023

  4. [11]

    Neural Netw

    Wu, D., Li, Z., Yu, Z., et al.: ‘Robust Low‑Rank Latent Feature Analysis for Spatio‑Temporal Signal Recovery’, IEEE Trans. Neural Netw. Learn. Syst., 2023, 34, (11)

    2023

  5. [12]

    Wu, D., He, Y., Luo, X.: ‘A Graph‑Incorporated Latent Factor Analysis Model for High‑Dimensional and Sparse Data’, IEEE Trans. Emerg. Topics Comput., 2023, 11, (4), pp. 907–917

    2023

  6. [13]

    Wu, D., Pang, P., He, Y., Luo, X.: ‘A Double‑Space and Double‑Norm Ensembled Latent Factor Model for Highly Accurate Web Service QoS Prediction’, IEEE Trans. Serv. Comput., 2023, 16, (2), pp. 802–814

    2023

  7. [14]

    Circuits Syst

    Li, Z., Li, S., Luo, X.: ‘A Novel Machine Learning System for Industrial Robot Arm Calibration’, IEEE Trans. Circuits Syst. II: Express Briefs, 2023

    2023

  8. [15]

    Big Data, 2024, 10, (1), pp

    Qin, W., Luo, X., Zhou, M.: ‘Adaptively‑Accelerated Parallel Stochastic Gradient Descent for High‑Dimensional and Incomplete Data Representation Learning’, IEEE Trans. Big Data, 2024, 10, (1), pp. 92–107

    2024

  9. [16]

    Chen, T., Li, S., Qiao, Y., Luo, X.: ‘A Robust and Efficient Ensemble of Diversified Evolutionary Computing Algorithms for Accurate Robot Cali- bration’, IEEE Trans. Instrum. Meas., 2024, 73, pp. 1–14

    2024

  10. [17]

    Automatica Sinica, 2024, 11, (2), pp

    Zeng, N., Li, X., Wu, P., Li, H., Luo, X.: ‘A Novel Tensor Decomposi- tion‑Based Efficient Detector for Low‑Altitude Aerial Objects With Knowledge Distillation Scheme’, IEEE/CAA J. Automatica Sinica, 2024, 11, (2), pp. 487–501

    2024

  11. [18]

    Neural Netw

    Yan, J., Jin, L., Luo, X., Li, S.: ‘Modified RNN for Solving Comprehen- sive Sylvester Equation With TDOA Application’, IEEE Trans. Neural Netw. Learn. Syst., 2023

    2023

  12. [19]

    Cybern., 2023, 53, (2), pp

    Chen, X., Luo, X., Jin, L., Li, S., Liu, M.: ‘Growing Echo State Network With an Inverse‑Free Weight Update Strategy’, IEEE Trans. Cybern., 2023, 53, (2), pp. 753–764

    2023

  13. [20]

    Pattern Anal

    Luo, X., Wu, H., Wang, Z., Wang, J., Meng, D.: ‘A Novel Approach to Large‑Scale Dynamically Weighted Directed Network Representation’, IEEE Trans. Pattern Anal. Mach. Intell., 2022, 44, (12), pp. 9756–9773

    2022

  14. [21]

    S., Luo, X.: ‘Graph Linear Convolution Pooling for Learning in Incomplete High‑Dimensional Data’, IEEE Trans

    Bi, F., He, T., Ong, Y. S., Luo, X.: ‘Graph Linear Convolution Pooling for Learning in Incomplete High‑Dimensional Data’, IEEE Trans. Knowl. Data Eng., 2024

    2024

  15. [22]

    Wu, H., Qiao, Y., Luo, X.: ‘A Fine‑Grained Regularization Scheme for Nonnegative Latent Factorization of High‑Dimensional and Incomplete Tensors’, IEEE Trans. Serv. Comput., 2024

    2024

  16. [23]

    Zhou, Y., Liu, Z., Luo, X., Hu, L., Zhou, M.: ‘Generalized Nesterov’s Acceleration‑Incorporated, Non‑negative and Adaptive Latent Factor Analy- sis’, IEEE Trans. Serv. Comput., 2022, 15, (5), pp. 2809–2823

    2022

  17. [24]

    Luo, X., Chen, M., Wu, H., Liu, Z., Yuan, H., Zhou, M.: ‘Adjusting Learning Depth in Non‑negative Latent Factorization of Tensors for Accu- rately Modeling Temporal Patterns in Dynamic QoS Data’, IEEE Trans. Autom. Sci. Eng., 2022, 18, (4), pp. 2142–2155

    2022

  18. [25]

    Neural Netw

    Luo, X., Liu, Z., Jin, L., Zhou, Y., Zhou, M.: ‘Symmetric Non‑negative Matrix Factorization‑Based Community Detection Models and Their Con- vergence Analysis’, IEEE Trans. Neural Netw. Learn. Syst., 2022, 33, (3), pp. 1203–1215

    2022

  19. [26]

    Li, J., Tan, F., He, C., Wang, Z., Song, H., Hu, P., Luo, X.: ‘Salien- cy‑Aware Dual Embedded Attention Network for Multivariate Time‑Series Forecasting in Information Technology Operations’, IEEE Trans. Ind. Informatics, 2024, 20, (3), pp. 4206–4217

    2024

  20. [27]

    Luo, X., Zhou, Y., Liu, Z., Zhou, M.: ‘Fast and Accurate Non‑negative Latent Factor Analysis on High‑Dimensional and Sparse Matrices in Rec- ommender Systems’, IEEE Trans. Knowl. Data Eng., 2023, 35, (4), pp. 3897– 3911

    2023

  21. [28]

    Neural Netw

    Luo, X., Zhong, Y., Wang, Z., Li, M.: ‘An Alternat- ing‑Direction‑Method of Multipliers‑Incorporated Approach to Symmetric Non‑negative Latent Factor Analysis’, IEEE Trans. Neural Netw. Learn. Syst., 2023, 34, (8), pp. 4826–4840

    2023

  22. [29]

    Luo, X., Wang, L., Hu, P., Hu, L.: ‘Predicting Protein –Protein Interac- tions Using Sequence and Network Information via Variational Graph Auto- encoder’, IEEE/ACM Trans. Comput. Biol. Bioinform., 2023, 20, (5), pp. 3182–3194

    2023

  23. [30]

    Luo, X., Chen, J., Yuan, Y., Wang, Z.: ‘Pseudo Gradient‑Adjusted Parti- cle Swarm Optimization for Accurate Adaptive Latent Factor Analysis’, IEEE Trans. Syst. Man Cybern. Syst., 2023

    2023

  24. [31]

    Automatica Sinica, 2024

    Chen, J., Liu, K., Luo, X., Yuan, Y., Sedraoui, K., Al‑Turki, Y., Zhou, M.: ‘A State‑Migration Particle Swarm Optimizer for Adaptive Latent Factor Analysis of High‑Dimensional and Incomplete Data’, IEEE/CAA J. Automatica Sinica, 2024

    2024

  25. [32]

    Automatica Sinica, 2024

    Yang, W., Li, S., Luo, X.: ‘Data‑Driven Vibration Control: A Review’, IEEE/CAA J. Automatica Sinica, 2024

    2024

  26. [33]

    Automatica Sinica, 2024

    Liao, X., Hoang, K., Luo, X.: ‘Local Search‑Based Anytime Algorithms for Continuous Distributed Constraint Optimization Problems’, IEEE/CAA J. Automatica Sinica, 2024

    2024

  27. [34]

    Zhong, Y., Liu, K., Gao, S., Luo, X.: ‘Alternating‑Direction‑Method of Multipliers‑Based Adaptive Nonnegative Latent Factor Analysis’, IEEE Trans. Emerg. Topics Comput., 2024

    2024

  28. [35]

    Chen, T., Yang, W., Zhang, Z., Luo, X.: ‘An Efficient Industrial Robot Calibrator With Multi‑Planer Constraints’, IEEE Trans. Ind. Informatics, 2024

    2024

  29. [36]

    Neural Netw

    Yuan, Y., Wang, Y., Luo, X.: ‘A Node -Collaboration-Informed Graph Convolutional Network for Highly -Accurate Representation to Undirected Weighted Graph’, IEEE Trans. Neural Netw. Learn. Syst., 2024

    2024

  30. [37]

    Lin, M., Liu, J., Chen, H., Xu, X., Luo, X., Xu, Z.: ‘A 3D Convolu- tion-Incorporated Dimension Preserved Decomposition Model for Traffic Data Prediction’, IEEE Trans. Intell. Transp. Syst., 2024

    2024

  31. [38]

    Automatica Sinica, 2024

    Luo, Z., Jin, X., Luo, Y., Zhou, Q., Luo, X.: ‘Analysis of Students’ Posi- tive Emotion and Smile Intensity Using Sequence -Relative Key-Frame La- beling and Deep-Asymmetric Convolutional Neural Network’, IEEE/CAA J. Automatica Sinica, 2024

    2024

  32. [39]

    Automati- ca Sinica, 2024

    Huang, P., Luo, X.: ‘FDTs: A Feature Disentangled Transformer for Interpretable Squamous Cell Carcinoma Grading’, IEEE/CAA J. Automati- ca Sinica, 2024

    2024

  33. [40]

    Automatica Sinica, 2024

    Chen, J., Yuan, Y., Luo, X.: 'SDGNN: Symmetry Preserving Dual Stream Graph Neural Networks', IEEE/CAA J. Automatica Sinica, 2024

    2024

  34. [41]

    Fuzzy Syst., 2024, 32, (2), pp

    Yuan, Y., Li, J., and Luo, X.: ‘A fuzzy PID -incorporated stochastic gradient descent algorithm for fast and accurate latent factor analysis’, IEEE Trans. Fuzzy Syst., 2024, 32, (2), pp. 445–459

    2024

  35. [42]

    Neural Netw

    He, Z., Lin, M., and Luo, X.: ‘Structure -preserved self -attention for fusion image information in multiple color spaces’, IEEE Trans. Neural Netw. Learn. Syst., 2024, 35, (11), pp. 8778–8790 5

    2024

  36. [43]

    Huang, P., and Luo, X.: ‘FDTs: A feature disentangled transformer for interpretable squamous cell carcinoma grading’, IEEE/CAA J. Autom. Sini- ca, 2024, 11, (2), pp. 487–501

    2024

  37. [45]

    Wu, Q., Ye, F., Gu, Q., et al.: ‘A Customised Down‑Sampling Machine Learning Approach for Sepsis Prediction’, Int. J. Med. Inform., 2024, 184, 105365

    2024

  38. [46]

    J., Alvarez, J

    Delahanty, R. J., Alvarez, J. A., Flynn, L. M., et al.: ‘Development and Evaluation of a Machine Learning Model for the Early Identification of Patients at Risk for Sepsis’, Ann. Emerg. Med., 2019, 73, (3), pp. 334–344

    2019

  39. [47]

    Bloch, E., Rotem, T., Cohen, J., et al.: ‘Machine Learning Models for Analysis of Vital Signs Dynamics: A Case for Sepsis Onset Prediction’, J. Healthc. Eng., 2019, 2019, Article 5930379

    2019

  40. [48]

    D., Futoma, J., Clement, M

    Bedoya, A. D., Futoma, J., Clement, M. E., et al.: ‘Machine Learning for Early Detection of Sepsis: An Internal and Temporal Validation Study’, JAMIA Open, 2020, 3, pp. 252–260

    2020

  41. [49]

    An improved multi -output Gaussian process RNN for the early detection of sepsis,

    J. Futoma, S. Hariharan, K. Heller, et al., “An improved multi -output Gaussian process RNN for the early detection of sepsis,” Machine Learning for Healthcare, 2017, pp. 243–254

    2017

  42. [50]

    Prediction of sepsis in the intensive care unit with minimal electronic health record data: a machine learning approach,

    T. Desautels, M. Calvert, P. Hoffman, et al., “Prediction of sepsis in the intensive care unit with minimal electronic health record data: a machine learning approach,” JMIR Medical Informatics, 2016, 4(3): e28

    2016

  43. [51]

    ICU data collaboration for sepsis prediction and early warning systems,

    D. Celi, L. Ghassemi, F. Naumann, et al., “ICU data collaboration for sepsis prediction and early warning systems,” Nature Medicine, 2019, 25(1), pp. 44–50

    2019

  44. [52]

    Early prediction of sepsis in the ICU using machine learning: a systematic review,

    S. Moor, C. R. Wulff, P. Svensson, et al., “Early prediction of sepsis in the ICU using machine learning: a systematic review,” Critical Care Medi- cine, 2021, 49(12), pp. e1221–e1232

    2021

  45. [53]

    Early prediction of sepsis from clinical data: the PhysioNet/Computing in Cardiology Challenge 2019,

    M. Reyna, C. Josef, S. Jeter, et al., “Early prediction of sepsis from clinical data: the PhysioNet/Computing in Cardiology Challenge 2019,” Critical Care Medicine, 2020, 48(2), pp. 210–217

    2019

  46. [54]

    An interpretable machine learn- ing model for accurate prediction of sepsis in the ICU,

    J. Nemati, M. Holder, A. Razmi, et al., “An interpretable machine learn- ing model for accurate prediction of sepsis in the ICU,” Critical Care Explo- rations, 2021, 3(1): e0331

    2021

  47. [55]

    Artificial intelligence approaches for sepsis: An updated review,

    A. Kamaleswaran, M. Akbilgic, “Artificial intelligence approaches for sepsis: An updated review,” Frontiers in Digital Health, 2022, 4: 867107

    2022

  48. [56]

    Early prediction of sepsis in the ICU using machine learning: A systematic survey,

    J. van Wyk, E. Khojandi, D. Davis, “Early prediction of sepsis in the ICU using machine learning: A systematic survey,” Computers in Biology and Medicine, 2019, 115: 103375

    2019

  49. [57]

    Sepsis prediction in intensive care using machine learning and temporal data,

    A. Schamoni, T. Frank, F. Fiedler, et al., “Sepsis prediction in intensive care using machine learning and temporal data,” Artificial Intelligence in Medicine, 2021, 118: 102104

    2021

  50. [58]

    Predicting sepsis with a recurrent neural network using electronic health records,

    M. Scherpf, C. Gräßer, T. Malberg, et al., “Predicting sepsis with a recurrent neural network using electronic health records,” Computers in Biology and Medicine, 2019, 113: 103395

    2019

  51. [59]

    Representation learning: A review and new perspectives,

    Y. Bengio, A. Courville, P. Vincent, “Representation learning: A review and new perspectives,” IEEE Trans. Pattern Anal. Mach. Intell., 2013, 35(8), pp. 1798–1828

    2013

  52. [60]

    Autoencoders, unsupervised learning, and deep architec- tures,

    K. Baldi, “Autoencoders, unsupervised learning, and deep architec- tures,” ICML Workshop on Unsupervised Feature Learning, 2012, pp. 37–49

    2012

  53. [61]

    Why does unsupervised pre - training help deep learning?,

    D. Erhan, Y. Bengio, A. Courville, et al., “Why does unsupervised pre - training help deep learning?,” Journal of Machine Learning Research, 2010, 11, pp. 625–660

    2010

  54. [62]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016

    2016

  55. [63]

    Tutorial on variational autoencoders,

    C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint, arXiv:1606.05908, 2016

    2016 arXiv

  56. [64]

    Estimating missing data in tem- poral data streams using multi -directional recurrent neural networks,

    A. Yoon, J. Zame, M. van der Schaar, “Estimating missing data in tem- poral data streams using multi -directional recurrent neural networks,” IEEE Trans. Biomed. Eng., 2019, 66(5), pp. 1477–1490

    2019

  57. [65]

    Directly modeling missing data in sequences with RNNs: Improved classification of clinical time series,

    M. Lipton, D. Kale, C. Wetzel, et al., “Directly modeling missing data in sequences with RNNs: Improved classification of clinical time series,” Machine Learning for Healthcare, 2016, pp. 253–270

    2016

  58. [66]

    GRAM: Graph - based attention model for healthcare representation learning,

    E. Choi, M. T. Bahadori, L. Song, W. Stewart, J. Sun, “GRAM: Graph - based attention model for healthcare representation learning,” KDD, 2017, pp. 787–795

    2017

  59. [67]

    Recurrent neural networks for multivariate time series with missing values,

    P. Che, Z. Cheng, J. Sun, “Recurrent neural networks for multivariate time series with missing values,” Scientific Reports, 2018, 8, 6085

    2018

  60. [68]

    GAIN: Missing data imputa- tion using generative adversarial nets,

    J. Yoon, W. R. Zame, M. van der Schaar, “GAIN: Missing data imputa- tion using generative adversarial nets,” ICML, 2018, pp. 5689–5698

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.