Pith. sign in

REVIEW 5 major objections 5 minor 35 references

Self-supervised sepsis prediction from seven vital signs rivals a 40-marker benchmark.

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

2026-08-01 20:17 UTC pith:WNJFT4QR

load-bearing objection A promising empirical comparison undone by single-run validation numbers and inconsistent cohort counts; the horizon-robustness finding is worth checking, but the paper needs revision before its benchmarks are cited. the 5 major comments →

arxiv 2607.16681 v1 pith:WNJFT4QR submitted 2026-07-18 cs.LG

A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning

Umair bin Mansoor , Munaf Rashid , Roomi Naqvi This is my paper
classification cs.LG
keywords sepsis predictionself-supervised learningJEPAVICRegrepresentation learningtemporal convolutional networkfederated learningXGBoost
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.

This paper attempts to show that early sepsis prediction does not require dense laboratory panels. Using only seven continuously measured vital-sign channels, it reports that a JEPA encoder trained by masked latent prediction and scored by XGBoost reaches AUPRC 0.636 at the moment of sepsis onset, close to the 0.667 of a supervised pretraining benchmark that uses 40 inputs. It further claims that a model pretrained with VICReg and then fine-tuned retains its discrimination much better as the prediction horizon lengthens, dropping only 16.8% from onset to ten hours earlier, versus 47.5% for a supervised temporal CNN and 65.3% for JEPA. If these numbers hold, bedside vitals alone could support clinically useful sepsis alerts, and the choice of self-supervised objective could be guided by horizon stability rather than point accuracy alone.

Core claim

The central discovery is that a deliberately small input set—seven vital-sign biomarkers with hourly binning and forward-fill—can carry competitive sepsis prediction when paired with self-supervised representation learning and a tree-based scorer. The best configuration, JEPA with masked latent prediction, mean pooling, and XGBoost, achieves AUPRC 0.636 at onset, within 5% of the SupMix reference (0.667) while using 83% fewer biomarkers. A separate finding is that VICReg pretraining followed by semi-supervised fine-tuning gives the most temporally persistent representations: its AUPRC falls 16.8% from onset to ten hours before, compared with 47.5% for the supervised TCN and 65.3% for JEPA, s

What carries the argument

The load-bearing elements are (1) a sparsity-selected set of seven vital-sign biomarkers—heart rate, systolic and diastolic blood pressure, temperature, respiratory rate, oxygen saturation, and mean arterial pressure—encoded as 42 hourly descriptors per bin; (2) a shared causal temporal convolutional network that produces 128-dimensional latents; (3) two self-supervised pretraining objectives, JEPA's masked latent prediction with variance/covariance regularization and VICReg's two-view invariance with variance/covariance regularization; (4) a frozen-encoder XGBoost classifier that far outperforms logistic regression on those latents; and (5) a 'horizon gate' that uses JEPA's scores for short

Load-bearing premise

The claim that one training objective is best rests on a single chronological validation split whose reported size is inconsistent, so the AUPRC gaps among JEPA, fine-tuned VICReg, and TCN could be within run-to-run noise.

What would settle it

Recompute the main table with multiple random seeds and a single, clearly defined temporal validation split; if the H0 AUPRC gap between JEPA+XGB (0.636) and fine-tuned VICReg+XGB (0.510) collapses under confidence intervals, or if the H0-to-H10 degradation ordering (16.8% vs 47.5% vs 65.3%) reverses, the paper's central claims fail. An external evaluation on a different ICU cohort would settle whether seven vitals alone transfer.

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

If this is right

  • Sepsis screening could rely on vital signs alone, cutting the cost and wait time of laboratory tests.
  • Self-supervised pretraining followed by light fine-tuning may be the right recipe when a model must work across many prediction horizons at once.
  • A frozen encoder plus XGBoost yields large gains over linear heads, so representation learning and classifier choice should be optimized together.
  • Federated training with dynamic client correction can approach centralized performance, making privacy-preserving multi-hospital sepsis models realistic.
  • A horizon-gate deployment can output both risk score and estimated time-to-onset without knowing the true horizon at inference time.

Where Pith is reading between the lines

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

  • If the degradation ordering survives repeated runs, model selection for clinical early-warning systems should be based on horizon-robustness rather than H0 AUPRC alone, since real alerts fire before onset.
  • The 7-biomarker result suggests a deployment path for low-resource ICUs and step-down units where lab draws are infrequent; this testable claim would require a prospective or external-cohort validation.
  • The paper's own reporting inconsistencies—validation set stated as both 2,439 and 618 patients, and federated client counts that do not sum to the training set—imply the headline gaps need replication with a fixed, preregistered split before clinical adoption.
  • The horizon-gate idea implies that a patient's peak horizon shifting leftward over hours could serve as an early deterioration signal; this is a testable hypothesis the paper sketches but does not evaluate.

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

5 major / 5 minor

Summary. The paper compares four modeling paradigms—masked-latent JEPA, VICReg, fine-tuned VICReg, and supervised TCN—plus raw-feature baselines, for early sepsis prediction using 7 vital-sign biomarkers derived from MIMIC-III/PhysioNet-2019 data. On a single chronological validation split, it reports JEPA+XGBoost+mean pooling reaching AUPRC 0.636 at onset, fine-tuned VICReg+XGBoost reaching 0.510, and a horizon-robustness ordering in which fine-tuned VICReg degrades only 16.8% from H0 to H10 versus 47.5% for supervised TCN and 65.3% for JEPA. It also reports federated learning experiments in which FedDyn is within 5.8% of centralized validation loss, and it proposes a deployed hybrid horizon-gated alerting architecture.

Significance. If the reported numbers are reproducible, the paper offers a useful empirical contribution: a drastically reduced 7-channel vital-sign representation may rival dense-lab pipelines, and SSL pretraining followed by fine-tuning may yield more horizon-robust representations. The comparison is broad (four paradigms, two classifiers, two pooling strategies, eleven horizons, four FL strategies), the main table is internally consistent, and the authors disclose that the FL experiments use a reduced architecture. However, every headline claim—the 0.636 vs 0.510 vs 0.474 AUPRC ranking, the 16.8%/47.5%/65.3% degradation hierarchy, and the 83%-fewer-biomarkers comparison with SupMix—rests on a single run, an unexplained validation-set count, and an evaluation protocol that does not train horizon-specific models. These issues must be resolved before the core empirical claims can be considered reliable. The paper ships no code, data, or proof artifacts, so verification is limited to the text.

major comments (5)
  1. [§III.A.3, §V.A, §III.J] The validation-set size is inconsistent. Section III.A.3 says the chronological split assigns 37,897 stays to training and the remaining 2,439 to validation, but Section V.A reports results on "the validation set (618 patients)". Similarly, the five FL clients in Section III.J sum to 3,731 samples, a number that does not match the stated 37,897-stay training set. The filtering that maps 40,336 stays to 618 validation patients and 37,897 stays to 3,731 FL samples is never described. If 618 is a selected subset of the 2,439, the selection rule must be stated and justified; if it is a typo, the headline AUPRCs are unverifiable.
  2. [§IV, Table III] All results come from a single chronological split, a single training run, and no seeds, confidence intervals, or significance tests. With roughly 100 positives in a 618-patient validation set, the key gaps (0.636 vs 0.510 vs 0.474; FT 0.510 vs TCN 0.474; the 16.8%/47.5%/65.3% degradation drops) could shift or invert under re-runs. In addition, model choices—XGBoost vs LR, mean vs last pooling, and later the horizon gate and thresholds—are selected on the same validation set, with no independent test split. The paper should report multi-seed means with bootstrap CIs and either create a true test set or present the current table as model selection rather than final performance.
  3. [§IV, horizon protocol] The horizon evaluation is described as training each model once and then evaluating at each H by restricting the input window. At H=10, the classifier receives a 62-hour input, while training used 72-hour windows; mean pooling over fewer time steps and absent later latents is a covariate shift, not a controlled test of temporal persistence. The degradation hierarchy in Table II is therefore confounded with input-length mismatch. To support the claim that fine-tuned VICReg is intrinsically more horizon-robust, the authors must either train/validate a separate model per horizon, or include truncated windows in training and evaluate on same-length inputs.
  4. [§V.E, Table V, Abstract] The comparison with SupMix is not apples-to-apples. SupMix reports AP on the full PhysioNet 2019 test set with 40 biomarkers, while Table V reports our AUPRC on a MIMIC-III validation subset with 7 biomarkers. The abstract's statement that JEPA+XGBoost "approaches the SupMix benchmark (0.667) while using 83% fewer biomarkers" is unsupported because the datasets, feature sets, and evaluation protocols differ. Additionally, Section III.A.2 says the PhysioNet 2019 dataset includes MIMIC-III plus Emory data, while Table V labels our results as MIMIC-III; the paper must clarify exactly which cohort was used.
  5. [§VI.B, deployment design] The proposed deployment architecture—horizon gate using JEPA for H≤3 and FT for H≥4, watch/alert thresholds of 0.4/0.5, and a peak-shift escalation of ≥2 horizons—is selected from the same validation results that are presented as final. No independent evaluation of this hybrid rule is given, so its clinical utility is not established. This is post-hoc selection presented as a recommendation; it needs a held-out or external validation with the specified thresholds fixed a priori.
minor comments (5)
  1. [§III.A.4, Table I] The "7 biomarkers" count includes MAP, which the paper itself notes is a linear combination of SBP and DBP. The effective number of independent channels is 6; please clarify whether the count is meant as raw channels or independent physiological signals.
  2. [§III.H, XGBoost] The configuration says "scale pos weight for imbalance" but never reports the value used. Report the actual scale_pos_weight for reproducibility.
  3. [§V.A, Table III] Define "FT" and use consistent notation for pooling (mean vs last) across Tables II and III; currently the table mixes "m" and "l" with the prose using full words.
  4. [§III.A.2 vs §V.E] Clarify whether experiments use the full PhysioNet 2019 challenge dataset (MIMIC-III + Emory) or only the MIMIC-III subset; Table V labels the cohort as MIMIC-III while Section III.A.2 describes a combined dataset.
  5. [Fig. 2 vs Fig. 6] Figures 2 and 6 appear to show the same FL convergence curves. If they are different, distinguish them clearly; if not, remove the duplicate.

Circularity Check

0 steps flagged

No circularity found; central claims are empirical AUPRC measurements rather than derivations, and no load-bearing step reduces to its own inputs.

full rationale

The paper's load-bearing claims are empirical AUPRC/AUROC numbers measured on a validation split (Table III) and degradation percentages computed from those measurements (Table II). These are not derived from the model definitions or from fitted parameters renamed as predictions. The JEPA/XGBoost AUPRC 0.636, the fine-tuned VICReg AUPRC 0.510, and the H0-to-H10 degradation hierarchy are all reported measurements, not consequences of construction. The horizon gate (JEPA for H<=3, FT for H>=4) is a post-hoc design choice based on the same validation results, but the paper does not report a new 'predicted' performance for the hybrid gate; it merely proposes a deployment rule. That is model-selection/overfitting risk, not circularity. The paper contains no self-citation chain: references to prior work (SupMix, VICReg, FedDyn, etc.) are external, and none is invoked as an unverified uniqueness theorem to force the present conclusion. The unresolved inconsistencies—validation set described as 2,439 stays in Section III.A.3 but reported as 618 patients in Section V.A, and FL client sizes summing to 3,731 versus a training set of 37,897—are reproducibility and statistical-validity concerns, not circular reasoning. The comparison to SupMix is also a protocol-matching concern, not a circular step. Accordingly, no circularity is established and the score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper introduces no invented physical or clinical entities; its claims are empirical benchmarks on a public dataset. Free parameters are dominated by post-hoc design choices: the biomarker set, the horizon-gate split and alert thresholds selected from the same validation results they are then used to explain, and unreported XGBoost imbalance weights. Standard SSL hyperparameters follow cited literature values. The domain axioms are the PhysioNet/Sepsis-3 labels, forward-fill adequacy, the chronological split's immunity to drift, the sparsity filter, and the shared TCN backbone.

free parameters (6)
  • Biomarker inclusion rule (hourly observation-rate threshold) = 7 vitals only; effective threshold ~10% (soft, post hoc)
    Section III.A.4 states all lab markers above 10% observation are excluded, but Table I shows Glucose at 16.5% yet excluded; the real rule is 'vitals only,' a post-hoc selection that fixes the entire 42-dim feature space behind every AUPRC number.
  • Horizon gate split point = JEPA for H<=3, FT for H>=4
    Section VI.B/Fig. 5: the gate picks whichever model had better observed AUPRC in Table III on the same validation set; no held-out evaluation of the gated hybrid exists.
  • Alerting thresholds (watch/alert/peak shift) = 0.4 at H>=4; 0.5 at H<=3; >=2 horizons/hour
    Section VI.B: asserted as a three-tier strategy with no calibration, PRC analysis, or sensitivity study; chosen by the authors, not derived from data.
  • LDA concentration parameter alpha (FL partition) = 0.5
    Section III.J: controls client label skew (Clients 0-4 from 0.1% to 48.6% positive), which drives the FL comparison; no sensitivity analysis, and client sizes (3,731 total) are unexplained vs 37,897 training stays.
  • XGBoost scale_pos_weight = not reported
    Section III.H: 'scale pos weight for imbalance' with the value omitted; this knob directly affects AUPRC on a 5-15% prevalence task.
  • VICReg/JEPA regularizer weights = lambda_i=1.0, lambda_v=1.0, lambda_c=0.04, gamma=1.0, rho=0.4
    Standard values from VICReg [18] and MAE practice, not fitted to the target outcome; listed as design choices that shape the learned representations.
axioms (5)
  • domain assumption Sepsis-3 labels from PhysioNet 2019 are correct ground truth
    Section III.A.2: all training and evaluation inherits the challenge's label construction (suspected infection within +/-24h plus SOFA increase of 2+). Noisy labels would shift every AUPRC; the paper takes them as given.
  • domain assumption Forward-fill imputation preserves discriminative signal
    Section III.B: descriptors are forward-filled from the most recent actual measurement; the 7-biomarker strategy assumes zero/forward-filled descriptors carry more signal than imputed lab values, asserted in Section III.A.4 but never tested against learned imputation (MGP, GRU-D).
  • domain assumption Chronological split is leakage-free and stable
    Section III.A.3: first 37,897 stays train, last 2,439 validate; assumes no temporal drift within BIDMC 2001-2012, and (with the 618-patient table) that the filtered validation subset represents the held-out population.
  • domain assumption Hourly observation rate is the right biomarker filter
    Section III.A.4: sparsity analysis is the sole selection criterion; labs are excluded because imputation 'introduces more noise than signal,' contradicted by the Glucose 16.5% exclusion and never validated empirically.
  • domain assumption TCN is an adequate backbone for all four paradigms
    Section III.C: all models share the dilated TCN encoder; if the backbone is limiting, the comparisons across training objectives are confined to one architecture family.

pith-pipeline@v1.3.0-alltime-deepseek · 5130 in / 5523 out tokens · 227267 ms · 2026-08-01T20:17:25.342115+00:00 · methodology

0 comments
read the original abstract

Early sepsis prediction from electronic health records is challenged by irregular sampling, high missingness, and class imbalance. We systematically compare four modeling paradigms -- self-supervised Joint Embedding Predictive Architecture (JEPA) via masked latent prediction, self-supervised VICReg (variance-invariance-covariance regularization) with two-view augmentation, semi-supervised fine-tuning of a VICReg-pretrained encoder, and supervised Temporal Convolutional Network (TCN) -- alongside raw-feature baselines. All models share a common preprocessing pipeline of hourly binning with forward-fill imputation applied to 7 biomarkers selected via sparsity analysis from the MIMIC-III dataset. Our best model (JEPA + XGBoost + mean pooling) achieves AUPRC 0.636 at the time of onset (H0), approaching the SupMix benchmark (0.667) while using 83\% fewer biomarkers. The Tier 1 pipeline -- VICReg pretraining followed by semi-supervised fine-tuning and XGBoost -- achieves AUPRC 0.510 at H0, a 3.1$\times$ improvement over the raw-feature baseline (0.165) and a 7.6\% improvement over the end-to-end supervised TCN (0.474). Crucially, the fine-tuned VICReg encoder exhibits the most temporally persistent representations, degrading only 16.8\% from H0 to H10 compared to 47.5\% for supervised TCN and 65.3\% for JEPA, demonstrating that self-supervised pretraining with task-aware fine-tuning yields features that are both sharp near onset and robust across prediction horizons.

Figures

Figures reproduced from arXiv: 2607.16681 by Munaf Rashid, Roomi Naqvi, Umair Bin Mansoor.

Figure 1
Figure 1. Figure 1: Sliding-window prediction protocol. The 72-hour lookback window preceding the prediction time [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: AUPRC vs Prediction Horizon across all evaluated models. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: AUROC vs Prediction Horizon. H = 0 (right) = imminent prediction; H = 10 (left) = 10 hours before onset. strategy for producing clinically useful sepsis predictions across varying prediction windows. D. Federated Learning Results Table IV summarizes the FL results. FedDyn achieves the best validation loss (0.2480), degrading only 5.8% relative to the centralized baseline. FedAvg (0.2728, +16.3%) and Fed￾Pr… view at source ↗
Figure 5
Figure 5. Figure 5: Real-time sepsis prediction inference architecture deployed in an ICU [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Validation loss convergence curves for centralized training and four federated learning aggregation strategies (FedAvg, FedProx, FedAvgM, FedDyn) [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] 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

35 extracted references · 3 linked inside Pith

  1. [1]

    Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study,

    K. E. Ruddet al., “Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study,” The Lancet, vol. 395, no. 10219, pp. 200–211, 2020. 11 Fig. 6. Validation loss convergence curves for centralized training and four federated learning aggregation strategies (FedAvg, FedProx, FedAvgM, FedDyn) under...

    1990

  2. [2]

    Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human septic shock,

    A. Kumaret al., “Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human septic shock,”Critical Care Medicine, vol. 34, no. 6, pp. 1589–1596, 2006

    2006

  3. [3]

    Learning to detect sepsis with a multitask Gaussian process RNN classifier,

    J. Futoma, S. Hariharan, and K. Heller, “Learning to detect sepsis with a multitask Gaussian process RNN classifier,” inProc. ICML, 2017, pp. 1174–1182

    2017

  4. [4]

    Early recognition of sepsis with Gaussian process temporal convolutional networks and dynamic time warping,

    M. Moor, M. Horn, B. Rieck, D. Roqueiro, and K. Borgwardt, “Early recognition of sepsis with Gaussian process temporal convolutional networks and dynamic time warping,” inProc. MLHC, 2019

    2019

  5. [5]

    Optimizing embedding space with sub-categorical supervised pre-training,

    T. Wanyanet al., “Optimizing embedding space with sub-categorical supervised pre-training,” inProc. IEEE ICHI, 2023, pp. 362–373

    2023

  6. [6]

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

    T. Desautelset al., “Prediction of sepsis in the intensive care unit with minimal electronic health record data: A machine learning approach,” JMIR Med. Inform., vol. 4, no. 3, e28, 2016

    2016

  7. [7]

    The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unantic- ipated intensive care unit admission, and death,

    G. B. Smithet al., “The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unantic- ipated intensive care unit admission, and death,”Resuscitation, vol. 84, no. 4, pp. 465–470, 2013

    2013

  8. [8]

    The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3),

    M. Singeret al., “The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3),”JAMA, vol. 315, no. 8, pp. 801– 810, 2016

    2016

  9. [9]

    MIMIC-III, a freely accessible critical care database,

    A. E. Johnsonet al., “MIMIC-III, a freely accessible critical care database,”Scientific Data, vol. 3, 2016

    2016

  10. [10]

    Early prediction of sepsis in the intensive care unit using the GRU-D-MGP-TCN model,

    S. Leeet al., “Early prediction of sepsis in the intensive care unit using the GRU-D-MGP-TCN model,”IEEE Access, vol. 12, pp. 148291– 148304, 2024

    2024

  11. [11]

    Advancing early detection of sepsis with temporal convolutional networks using ECG signals,

    M. Apalak and K. Kiasaleh, “Advancing early detection of sepsis with temporal convolutional networks using ECG signals,”IEEE Access, vol. 12, pp. 3417–3427, 2024

    2024

  12. [12]

    Multi-branching temporal convolutional network for sepsis prediction,

    Z. Wang and B. Yao, “Multi-branching temporal convolutional network for sepsis prediction,”IEEE J. Biomed. Health Inform., vol. 26, no. 2, pp. 876–887, 2022

    2022

  13. [13]

    PoEMS: Policy network-based early warning monitoring system for sepsis in intensive care units,

    H. Daiet al., “PoEMS: Policy network-based early warning monitoring system for sepsis in intensive care units,”IEEE J. Biomed. Health Inform., vol. 27, no. 7, pp. 3534–3545, 2023

    2023

  14. [14]

    Improving early sepsis onset prediction through federated learning,

    C. D ¨using and P. Cimiano, “Improving early sepsis onset prediction through federated learning,” inProc. AIBio Workshop, 2025

    2025

  15. [15]

    Impact of a deep learning sepsis prediction model on quality of care and survival,

    A. Boussinaet al., “Impact of a deep learning sepsis prediction model on quality of care and survival,”npj Digit. Med., vol. 7, 14, 2024

    2024

  16. [16]

    Development and prospective implementation of a large language model based system for early sepsis prediction,

    S. P. Shashikumaret al., “Development and prospective implementation of a large language model based system for early sepsis prediction,”npj Digit. Med., vol. 8, 290, 2025

    2025

  17. [17]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  18. [18]

    VICReg: Variance-invariance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “VICReg: Variance-invariance- covariance regularization for self-supervised learning,” inProc. ICLR, 2022

    2022

  19. [19]

    Unsupervised represen- tation learning for time series with temporal neighborhood coding,

    S. Tonekaboni, D. Eytan, and A. Goldenberg, “Unsupervised represen- tation learning for time series with temporal neighborhood coding,” in Proc. ICLR, 2021

    2021

  20. [20]

    Online sepsis prediction using vital signs and multiscale temporal-aware contrastive learning: Model development and validation study,

    X. Yanget al., “Online sepsis prediction using vital signs and multiscale temporal-aware contrastive learning: Model development and validation study,”JMIR Med. Inform., vol. 14, e82762, 2026

    2026

  21. [21]

    Cross-modal contrastive learning for predicting sepsis onset in Medical Internet of Things (MIoT),

    Y .-S. Chen, W.-H. Wang, C.-T. Hu, and I. You, “Cross-modal contrastive learning for predicting sepsis onset in Medical Internet of Things (MIoT),”Internet of Things, vol. 29, 101456, 2025

    2025

  22. [22]

    VISReg: Variance-invariance- sketching regularization for JEPA training,

    H. Wu, R. Balestriero, and M. Levine, “VISReg: Variance-invariance- sketching regularization for JEPA training,” arXiv:2606.02572, 2026

    Pith/arXiv arXiv 2026

  23. [23]

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

    M. A. Reynaet al., “Early prediction of sepsis from clinical data: The PhysioNet/Computing in Cardiology Challenge 2019,”Critical Care Medicine, vol. 48, no. 2, pp. 210–217, 2020

    2019

  24. [24]

    The SOFA (Sepsis-related Organ Failure As- sessment) score to describe organ dysfunction/failure,

    J. L. Vincentet al., “The SOFA (Sepsis-related Organ Failure As- sessment) score to describe organ dysfunction/failure,”Intensive Care Medicine, vol. 22, no. 7, pp. 707–710, 1996

    1996

  25. [25]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  26. [26]

    XGBoost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” inProc. ACM SIGKDD, 2016, pp. 785–794

    2016

  27. [27]

    The precision-recall plot is more informa- tive than the ROC plot when evaluating binary classifiers on imbalanced datasets,

    T. Saito and M. Rehmsmeier, “The precision-recall plot is more informa- tive than the ROC plot when evaluating binary classifiers on imbalanced datasets,”PLoS ONE, vol. 10, no. 3, e0118432, 2015

    2015

  28. [28]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inProc. AISTATS, 2017, pp. 1273–1282

    2017

  29. [29]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” inProc. MLSys, 2020

    2020

  30. [30]

    Federated learning based on dynamic regularization,

    D. A. E. Acar, Y . Zhao, R. M. Navarro, M. Mattina, P. N. Whatmough, and V . Saligrama, “Federated learning based on dynamic regularization,” inProc. ICLR, 2021

    2021

  31. [31]

    Measuring the effects of non- identical data distribution for federated visual classification,

    T.-M. H. Hsu, H. Qi, and M. Brown, “Measuring the effects of non- identical data distribution for federated visual classification,” inProc. ICLR Workshop, 2019

    2019

  32. [32]

    A federated learning framework with knowledge graph and temporal transformer for early sepsis prediction in multi-center ICUs,

    Y . Chang, G. Lin, J. J. Chuang, S. Liu, X. Li, and Y . Li, “A federated learning framework with knowledge graph and temporal transformer for early sepsis prediction in multi-center ICUs,” arXiv:2603.15651, 2026

    arXiv 2026

  33. [33]

    Optimizing sepsis mortality prediction using hybrid federated learning and explainable AI framework,

    M. Z. Fuzail, I. ud Din, S. Ahmed, A. Alhumam, and A. H. Khan, “Optimizing sepsis mortality prediction using hybrid federated learning and explainable AI framework,”Sci. Rep., 2026

    2026

  34. [34]

    Cross-hospital sepsis early detection via semi-supervised optimal transport with self-paced ensemble,

    R. Dinget al., “Cross-hospital sepsis early detection via semi-supervised optimal transport with self-paced ensemble,”IEEE J. Biomed. Health Inform., vol. 27, no. 6, pp. 3049–3060, 2023

    2023

  35. [35]

    Machine learning predicts sepsis deterioration trajec- tories,

    R. Zhanget al., “Machine learning predicts sepsis deterioration trajec- tories,”npj Digit. Med., vol. 9, 385, 2026

    2026