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

The DOI matches its Crossref record; 43 of the 44 references it cites were found in a registry or an index; 23 of the 35 sources its claims cite were checked for support.

Machine learning-driven prediction model for successful weaning of patients from mechanical ventilation in ICU

Open access
Qiu, Changcui, Tang, Lulu, Zhuang, Yugang, Chi, Chunwei, Zheng, Kangwei, Zhu, Xiaoping · 2026 · Intensive Care Medicine Experimental

Abstract

BACKGROUND: Mechanical ventilation is a critical life support technology in the intensive care unit. However, the weaning process remains complex, making the optimal timing for liberation from ventilation challenging to ascertain and imposing a considerable clinical workload. Additionally, advanced weaning assistance tools that integrate multidimensional clinical factors to help clinical staff make precise decisions during the weaning process are lacking. The aim of this study to develop and validate an interpretable machine learning model that comprehensively evaluates the factors influencing weaning to provide clinical decision support for weaning. METHOD: We collected data from the ICU of Shanghai Tenth People's Hospital and its affiliated hospitals. Ten distinct machine learning algorithms for predicting extubation outcomes in patients receiving mechanical ventilation were developed and internally validated. Model performance was quantified using the area under the receiver operating characteristic curve AUC, overall accuracy, sensitivity, specificity, and F1 score. NRI, IDI, and DCA were used to comprehensively identify the optimal model. The relative contribution of each predictor was ranked and compared through SHAP analysis, and the best-performing model was externally validated. RESULTS: Through univariate and LASSO analyses, 24 predictive variables for machine learning model construction were identified. Comprehensive evaluation showed that among the candidate algorithms, the LGB model demonstrated the highest overall performance. SHAP analysis revealed that the top-ranked features for predicting successful liberation from mechanical ventilation were creatinine levels, lactate levels, the level of consciousness, SpO2, systolic blood pressure, cough reflex, chronic respiratory disease, diastolic blood pressure, and age. CONCLUSIONS: The optimized predictive model, which was developed through the integration of multidimensional predictive factors with diverse machine learning algorithms, exhibits superior predictive accuracy and demonstrates significant clinical potential for determining the optimal timing for weaning patients receiving invasive mechanical ventilation. Trial registration Current Controlled Trials ChiCTR2400093658; registration date: December 10, 2024.

Overview

CoChat Summary

TL;DR

This multicenter study developed and internally/externally validated machine learning models to predict successful weaning from invasive mechanical ventilation in ICU patients, selecting Light Gradient Boosting (LGB) as the best-performing model. The model showed strong discrimination and clinical utility, with external validation AUC 0.972 and key predictors including creatinine, lactate, consciousness level, and SpO2.

Key findings

  • From univariate screening and LASSO regression, the authors selected 24 predictor variables for model building.
  • Among 10 classifiers, the LGB model was chosen as optimal; its internal validation performance was AUC 0.952, accuracy 0.954, sensitivity 0.980, specificity 0.803, F1 score 0.973, and precision 0.966. The RF model had slightly higher AUC (0.968) but the authors still selected LGB.
  • In the external validation set, the final model reached AUC 0.972, and DCA suggested significant net benefits across threshold probabilities.
  • SHAP ranked the top features for successful liberation as creatinine, lactate, consciousness level, SpO2, systolic blood pressure, cough reflex, chronic respiratory disease, diastolic blood pressure, and age.

Methods

  • Multicenter retrospective cohort using ICU data from Shanghai Tenth People’s Hospital and affiliated hospitals for model development/internal validation (Jan 1, 2021–Jun 30, 2024) and a separate ICU cohort for external validation (Dec 1, 2024–Jul 30, 2025).
  • Included adults ≥18 years who were in ICU >48 h and received mechanical ventilation >24 h; excluded ECMO, advanced cancer, and palliative-care cases. The analysis included 1371 patients in the development/internal-validation cohort and 402 in external validation.
  • Built and compared 10 machine-learning classifiers (KNN, SVM, NN, XGB, LR, RF, DT, ET, SGD, LGB), using a 7:3 train/test split, LASSO with fivefold cross-validation, SMOTE for imbalance, and evaluation by AUC, accuracy, sensitivity, specificity, F1, NRI, IDI, DCA, and SHAP.

Limitations

  • The paper states that it did not include imaging data such as diaphragmatic displacement/thickness, despite their potential relevance to weaning.
  • It also states that invasive hemodynamic indicators (e.g., CVP, PCWP, PiCCO) were difficult to obtain, so more accessible proxies were used instead.
  • The text makes evident that the study is retrospective and relies on data from a specific hospital system; this may limit generalizability beyond the studied setting (evident from the design and data source, not explicitly framed as a limitation by the authors).

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

Integrity signals

Registry

Crossref · 10.1186/s40635-026-00859-8

Cites

43 of 43 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Research flowchart. This flowchart illustrates the patient screening and machine learning model development process. Ten classifiers were used to develop the machine learning models. The models were trained and internally validated at a 7:3 ratio, and the optimal model was selected for external validation following SMOTE analysis
Fig. 1 Research flowchart. This flowchart illustrates the patient screening and machine learning model development process. Ten classifiers were used to develop the machine learning models. The models were trained and internally validated at a 7:3 ratio, and the optimal model was selected for external validation following SMOTE analysisCC BY · PubMed Central
Fig. 2 Comparison of the ROC curves for the 10 machine learning models. AUC area under the curve, knn K-Nearest Neighbors, svm Support Vector Machine, nn Neural Network, xgb Extreme Gradient Boosting, lr Logistic Regression, rf Random Forest, dt Decision Tree, et Extra Trees, sgd Stochastic Gradient Descent, lgb Light Gradient Boosting Machine
Fig. 2 Comparison of the ROC curves for the 10 machine learning models. AUC area under the curve, knn K-Nearest Neighbors, svm Support Vector Machine, nn Neural Network, xgb Extreme Gradient Boosting, lr Logistic Regression, rf Random Forest, dt Decision Tree, et Extra Trees, sgd Stochastic Gradient Descent, lgb Light Gradient Boosting MachineCC BY · PubMed Central
Fig. 3 DCA analysis based on the LGB model, lgb Light Gradient Boosting Machine
Fig. 3 DCA analysis based on the LGB model, lgb Light Gradient Boosting MachineCC BY · PubMed Central
Fig. 4 SHAP analysis of the LGB model. a SHAP summary bar plot showing feature importance ranking (internal validation); b feature importance based on the SHAP results. The features are arranged along the vertical axis, and the SHAP values are plotted on the horizontal axis. The colors of the points for each feature value are different; pink indicates a positive correlation with successful withdrawal, whereas blue indicates a negative correlation with successful withdrawal (internal validation); Cr creatinine, La lactate, CS2 consciousness state (Drowsy & Stupor & Coma), SP systolic pressure, CR2 cough reflex (cough reflex in response to airway stimulation, sputum only in the artificial airway and the need for suction to help the patient clear it), CR3 cough reflex (no stimulation needed to produce a cough reflex; sputum can be expelled by the patient’s efforts alone), HCRD history of chronic respiratory disease, DP diastolic pressure, K potassium, Ca calcium, TnT2 cardiac troponin T (0.1–1.0 μg/L), HI hemodynamic instability, NTproBN4 N-terminal pro B-type natriuretic peptide (> 900 ng/L), PCT4 procalcitonin (> 2 ng/L)
Fig. 4 SHAP analysis of the LGB model. a SHAP summary bar plot showing feature importance ranking (internal validation); b feature importance based on the SHAP results. The features are arranged along the vertical axis, and the SHAP values are plotted on the horizontal axis. The colors of the points for each feature value are different; pink indicates a positive correlation with successful withdrawal, whereas blue indicates a negative correlation with successful withdrawal (internal validation); Cr creatinine, La lactate, CS2 consciousness state (Drowsy & Stupor & Coma), SP systolic pressure, CR2 cough reflex (cough reflex in response to airway stimulation, sputum only in the artificial airway and the need for suction to help the patient clear it), CR3 cough reflex (no stimulation needed to produce a cough reflex; sputum can be expelled by the patient’s efforts alone), HCRD history of chronic respiratory disease, DP diastolic pressure, K potassium, Ca calcium, TnT2 cardiac troponin T (0.1–1.0 μg/L), HI hemodynamic instability, NTproBN4 N-terminal pro B-type natriuretic peptide (> 900 ng/L), PCT4 procalcitonin (> 2 ng/L)CC BY · PubMed Central
Fig. 5 External validation ROC curve for the LGB model a showing the model’s performance in distinguishing positive cases; decision curve analysis (DCA) b based on the LGB model for external validation. The external validation DCA results based on the LGB model demonstrate that this model exhibits significant net benefits across various threshold probabilities at different time points
Fig. 5 External validation ROC curve for the LGB model a showing the model’s performance in distinguishing positive cases; decision curve analysis (DCA) b based on the LGB model for external validation. The external validation DCA results based on the LGB model demonstrate that this model exhibits significant net benefits across various threshold probabilities at different time pointsCC BY · PubMed Central

Verification

Claims the paper makes

15 Valid · 3 Not valid · 6 Not checked · 40 Own results and interpretations

19 more claims

Create a free account to see all 64 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 26
  • Probably not a substantial claim. 20

Create a free account to see these claims.

Citations

Cites

Show all 44 references

Research works printed but not linked: 1

  1. Windisch WDDGJ (2020) Prolonged weaning from mechanical ventilation. Dtsch Arztebl Int 117(12):197–204

Metadata

Authors
Qiu, C., Tang, L., Zhuang, Y., Chi, C., Zheng, K., & Zhu, X.
Affiliations
  • Tongji University (CN)
  • Shanghai Tenth People's Hospital (CN)
Corresponding author
Zhu, Xiaoping
Venue
Intensive Care Medicine Experimental
Published
2026-01-21
Volume
14
Issue
1
Pages
6-6
ISSN
2197-425X (ISSN-L)
OpenAlex
W7125115277
Cited by
1
References
44
Citations by year
2026: 1
FWCI
4.75
Type
journal article
Language
en
Open access
Licence
Locations
  • Intensive Care Medicine Experimental (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
Funders
  • the Shanghai Nursing Association Research Project (2024MS-B21)
  • National Natural Science Foundation of China (72074168)
Topics
  • Health Sciences › Medicine › Pulmonary and Respiratory Medicine › Respiratory Support and Mechanisms
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
weaning from mechanical ventilation, extubation prediction, intensive care unit, clinical decision support, model interpretability, SHAP analysis, LASSO feature selection, gradient boosting, lactate, creatinine, level of consciousness, cough reflex, chronic respiratory disease

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-driven prediction model for successful weaning of patients from mechanical ventilation in ICU. https://app.cochat.ai/papers/10.1186/s40635-026-00859-8
The paper
Qiu, C., Tang, L., Zhuang, Y., Chi, C., Zheng, K., & Zhu, X. (2026). Machine learning-driven prediction model for successful weaning of patients from mechanical ventilation in ICU. Intensive Care Medicine Experimental. https://doi.org/10.1186/s40635-026-00859-8

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.