Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 42 of the 42 references it cites were found in a registry or an index; 16 of the 27 sources its claims cite were checked for support.

Machine learning predicts sepsis deterioration trajectories.

Open access
Zhang, Rui, Long, Fang, Zhao, Zhanqi, Wu, Jingyi, Tan, Ruoming, Xu, Wen, Li, Lei, Long, Yun, Qu, Hongping · 2026 · NPJ Digital Medicine

Abstract

Sepsis has heterogeneous clinical trajectories, but conventional severity scores offer only static risk estimates. Timely, dynamic prediction could enable personalized intervention. In this multicenter retrospective study of 47,936 ICU patients meeting Sepsis-3 criteria from one institutional and two public datasets (MIMIC-III, eICU; sensitivity in MIMIC-IV), group-based trajectory modeling identified latent recovery patterns. An ensemble machine-learning model incorporating dynamic physiological variability was trained, temporally validated, and externally tested; clinical impact was assessed following implementation. Three trajectories emerged: rapid recovery (41.5%), slow recovery (36.4%), and clinical deterioration (22.1%). In the final binary classification task, AUROC was 0.92 (development), 0.89 (internal), 0.84 (MIMIC-III) and 0.77 (eICU); median warning time before deterioration was 17.6 h (Overall pooled across all cohorts). Reduced heart rate variability (SD < 10 bpm) predicted mortality (adjusted HR 2.17). Implementation reduced ICU stay by 1.8 days, machanical ventilation by 2.3 days, and 28-day mortality by 5.7%. This externally validated trajectory-based model offers accurate, early risk stratification for sepsis, supporting proactive, individualized critical care.

Overview

CoChat Summary

## TL;DR

This multicenter retrospective study developed and externally validated a trajectory-based machine-learning model to predict sepsis deterioration among 47,936 ICU patients meeting Sepsis-3 criteria. The model achieved AUROC 0.92 in development and 0.77 in eICU, identified three trajectories, and showed a median 17.6 h warning time before deterioration.

## Key findings

  • Group-based trajectory modeling identified three sepsis trajectories: rapid recovery (41.5%, n = 19,883), slow recovery (36.4%, n = 17,449), and clinical deterioration (22.1%, n = 10,604).
  • In the final binary classification task (clinical deterioration vs. non-deterioration), AUROC was 0.92 in development, 0.89 in internal validation, 0.84 in MIMIC-III, and 0.77 in eICU; corresponding AUPRC values were 0.87, 0.84, 0.80, and 0.73.
  • The model provided a median warning time of 17.6 h before clinical deterioration; within the 12-hour window, sensitivity was 0.83 and specificity was 0.88.
  • Reduced heart rate variability (HR SD < 10 bpm) was associated with higher 28-day mortality risk (adjusted HR 2.17, 95% CI 1.68–2.79); implementation was associated with reductions in ICU stay by 1.8 days, mechanical ventilation duration by 2.3 days, and 28-day mortality by 5.7%.

## Methods

  • Multicenter retrospective cohort study using 47,936 adult ICU patients with sepsis-3 from one institutional dataset plus MIMIC-III and eICU; MIMIC-IV was used for sensitivity analysis.
  • Patients had at least three observed time points (median 7); 45 candidate variables were narrowed using clinical relevance, missingness thresholds, LASSO, and random forest importance, with temporal features derived from vital signs, labs, SOFA, and treatment data.
  • Group-based trajectory modeling selected a three-class solution; machine-learning models (including random forest, gradient boosting, and deep learning approaches) were trained with temporal validation, external validation, SHAP interpretation, Cox models, and propensity score-matched pre/post implementation analyses.

## Limitations

  • The paper states that validation was limited to datasets from the United States and China, so broader geographic generalizability remains untested.
  • The implementation analysis was non-randomized, so causality cannot be definitively established.
  • The model requires high-resolution temporal data, which may limit applicability in resource-constrained settings.
  • Evident from the text: the study is retrospective, so temporal bias and unmeasured confounding are concerns; the paper explicitly notes these in the Discussion.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-026-02565-x

Cites

42 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Study design and methodology. This figure illustrates the three-phase study approach (trajectory identification, model development, and implementation) and patient cohort development. It shows how the initial multi-center sepsis database (n = 54,758) was refined to a final analysis cohort (n = 47,936) after exclusions, and further divided into development, validation, and implementation cohorts. ICU Intensive Care Unit, EHR Electronic Health Record, MIMIC-III Medical Information Mart for Intensive Care III, eICU Electronic Intensive Care Unit.
Fig. 1 Study design and methodology. This figure illustrates the three-phase study approach (trajectory identification, model development, and implementation) and patient cohort development. It shows how the initial multi-center sepsis database (n = 54,758) was refined to a final analysis cohort (n = 47,936) after exclusions, and further divided into development, validation, and implementation cohorts. ICU Intensive Care Unit, EHR Electronic Health Record, MIMIC-III Medical Information Mart for Intensive Care III, eICU Electronic Intensive Care Unit.CC BY-NC-ND · PubMed Central
Fig. 2 Sepsis trajectory patterns and clinical parameters. a Principal component analysis (PCA) shows the separation of patients according to their sepsis trajectories: Rapid Recovery, Slow Recovery, and Clinical Deterioration. b Representative longitudinal SOFA score patterns for each trajectory group over the first 60 h of ICU admission. c Key clinical parameters over time for each trajectory group: SOFA score, lactate, mean arterial pressure (MAP), heart rate, respiratory rate, and oxygen saturation; shaded areas represent ±95% CI around each trajectory curve. Several parameters with prominent inter-trajectory differences (lactate clearance, HR variability, SOFA trend, MAP variability, RR variability) were further evaluated in the final prediction model. d Distribution of the study population across the three identified trajectory groups. SOFA Sequential Organ Failure Assessment, MAP Mean Arterial Pressure, ICU Intensive Care Unit, bpm Beats Per Minute (heart rate) or Breaths Per Minute (respiratory rate), mmHg Millimeters of Mercury- mmol/L Millimoles per Liter.
Fig. 2 Sepsis trajectory patterns and clinical parameters. a Principal component analysis (PCA) shows the separation of patients according to their sepsis trajectories: Rapid Recovery, Slow Recovery, and Clinical Deterioration. b Representative longitudinal SOFA score patterns for each trajectory group over the first 60 h of ICU admission. c Key clinical parameters over time for each trajectory group: SOFA score, lactate, mean arterial pressure (MAP), heart rate, respiratory rate, and oxygen saturation; shaded areas represent ±95% CI around each trajectory curve. Several parameters with prominent inter-trajectory differences (lactate clearance, HR variability, SOFA trend, MAP variability, RR variability) were further evaluated in the final prediction model. d Distribution of the study population across the three identified trajectory groups. SOFA Sequential Organ Failure Assessment, MAP Mean Arterial Pressure, ICU Intensive Care Unit, bpm Beats Per Minute (heart rate) or Breaths Per Minute (respiratory rate), mmHg Millimeters of Mercury- mmol/L Millimoles per Liter.CC BY-NC-ND · PubMed Central
Fig. 3 ROC and precision-recall curves for the final binary classification task (clinical deterioration vs. non-deterioration). a Receiver Operating Characteristic (ROC) curves for the development, internal validation, and MIMIC III external validation datasets. The ROC curve plots the true positive rate against the false positive rate, and the area under the curve (AUC) quantifies discriminatory performance. b Precision–Recall curve for the eICU external validation dataset, showing the relationship between precision and recall for different classification thresholds. The dashed line represents the baseline precision equal to the prevalence of positive cases in the dataset. c SHAP summary plot showing the top 10 features contributing to the deterioration prediction model. Notably, five of these high-ranking features—lactate clearance rate, heart rate variability, SOFA score trend, mean arterial pressure variability, and respiratory rate variability—were among the dynamic physiological parameters that displayed marked trajectory differences in Fig. 2. This overlap highlights the consistency between trajectory-based observations and model-based predictive importance. Note: Metrics from the multi-time-point trajectory classification task are not plotted here to avoid mixing task types. All curves use conventional axis scaling to ensure visual alignment with the numerical values.
Fig. 3 ROC and precision-recall curves for the final binary classification task (clinical deterioration vs. non-deterioration). a Receiver Operating Characteristic (ROC) curves for the development, internal validation, and MIMIC III external validation datasets. The ROC curve plots the true positive rate against the false positive rate, and the area under the curve (AUC) quantifies discriminatory performance. b Precision–Recall curve for the eICU external validation dataset, showing the relationship between precision and recall for different classification thresholds. The dashed line represents the baseline precision equal to the prevalence of positive cases in the dataset. c SHAP summary plot showing the top 10 features contributing to the deterioration prediction model. Notably, five of these high-ranking features—lactate clearance rate, heart rate variability, SOFA score trend, mean arterial pressure variability, and respiratory rate variability—were among the dynamic physiological parameters that displayed marked trajectory differences in Fig. 2. This overlap highlights the consistency between trajectory-based observations and model-based predictive importance. Note: Metrics from the multi-time-point trajectory classification task are not plotted here to avoid mixing task types. All curves use conventional axis scaling to ensure visual alignment with the numerical values.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

9 Valid · 6 Not valid · 5 Not checked · 112 Own results and interpretations

65 more claims · 4 with their cited sources checked

Create a free account to see all 132 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 43
  • Probably not a substantial claim. 24
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 40 references

Research works printed but not linked: 3

  1. Futoma, J. et al. An improved multi-output Gaussian process RNN with real-time validation for early sepsis detection. Proc. Mach. Learn. Healthcare 68, 243–254 (2017).
  2. Futoma, J., Hariharan, S. & Heller, K. Learning to detect sepsis with a multitask Gaussian process RNN classifier. Proc. 34th Int. Conf. Mach. Learn. 70, 1174–1182 (2017).
  3. Futoma, J. et al. An improved multi-output Gaussian process RNN with real-time validation for early sepsis detection. Proc. Mach. Learn. Healthcare. 68, 243–254 (2017).

Cited by

See all 6

Metadata

Authors
Zhang, R., Long, F., Zhao, Z., Wu, J., Tan, R., Xu, W., Li, L., Long, Y., & Qu, H.
Affiliations
  • Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
  • Shanghai Jiao Tong University (CN)
Corresponding author
  • Wu, Jingyi
  • Tan, Ruoming
  • Xu, Wen
  • Long, Yun
  • Qu, Hongping
Venue
NPJ Digital Medicine
Published
2026-03-26
Volume
9
Issue
1
ISSN
2398-6352 (ISSN-L)
OpenAlex
W7140750475
Cited by
6
References
40
Citations by year
2026: 6
FWCI
27.67
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Natural Science Foundation of China (82470088)
  • Shanghai Science and Technology Commission Project (23Y11900100)
  • science and technology commission of shanghai municipality (24ZR1447000)
  • National High-Level Hospital Clinical Research Funding (2022- PUMCH-B-115)
Topics
  • Health Sciences › Medicine › Critical Care and Intensive Care Medicine › Sepsis Diagnosis and Treatment
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Endocrinology, Diabetes and Metabolism › Hyperglycemia and glycemic control in critically ill and hospitalized patients
Keywords
sepsis, clinical trajectories, group-based trajectory modeling, Sepsis-3, intensive care unit patients, heart rate variability, mortality prediction, risk stratification, ensemble learning, MIMIC-III, eICU

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Machine learning predicts sepsis deterioration trajectories. https://app.cochat.ai/papers/10.1038/s41746-026-02565-x
The paper
Zhang, R., Long, F., Zhao, Z., Wu, J., Tan, R., Xu, W., Li, L., Long, Y., & Qu, H. (2026). Machine learning predicts sepsis deterioration trajectories. NPJ Digital Medicine. https://doi.org/10.1038/s41746-026-02565-x

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.