Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

A Lightweight LSTM-Transformer Fusion Architecture for Real-Time Sepsis Mortality Prediction.

Zekai Yu1, Feiwei Qin1, Zhu Zhu2,3

  • 1School of Computer Science and Technology, Hangzhou Dianzi University, Hangzhou, China.

Journal of Intensive Care Medicine
|May 6, 2026
PubMed
Summary

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Multimodal Feature Prototype Learning for Interpretable and Discriminative Cancer Survival Prediction.

IEEE journal of biomedical and health informatics·2026
Same author

Systemic predictors in middle ear cholesterol granuloma: methodological and translational considerations.

European archives of oto-rhino-laryngology : official journal of the European Federation of Oto-Rhino-Laryngological Societies (EUFOS) : affiliated with the German Society for Oto-Rhino-Laryngology - Head and Neck Surgery·2026
Same author

Machine learning phenotypes and heterogeneous albumin effects in cirrhotic AKI: a causal inference study.

Internal and emergency medicine·2026
Same author

Crossover effect: causal machine learning reveals opposing mortality responses to mean arterial pressure targets among phenotypically distinct hypertensive patients with septic shock.

Internal medicine journal·2026
Same author

From one to twelve: feasibility and clinical utility of deep learning-derived 12-lead ECGs for remote cardiac monitoring.

Frontiers in cardiovascular medicine·2026
Same author

ECG-GPT: considerations for advancing toward clinical implementation.

European heart journal. Digital health·2026

This study introduces a new deep learning model that accurately predicts sepsis patient mortality by considering treatment responses. The model shows improved performance and generalizability, offering a valuable tool for early warning systems.

Area of Science:

  • Critical Care Medicine
  • Artificial Intelligence in Healthcare
  • Biomedical Informatics

Background:

  • Accurate short-term mortality prediction in sepsis is vital for clinical decisions.
  • Existing models overlook dynamic responses to interventions, underestimating risk due to therapeutic masking.
  • Sepsis patient data from MIMIC-IV v3.1 database was utilized.

Purpose of the Study:

  • To develop a lightweight hybrid deep learning framework for predicting 24-hour sepsis mortality.
  • To integrate dynamic intervention responses, including vasopressor rates and urine output, into mortality prediction.
  • To address the limitations of static parameter-based models and the masking effect of treatments.

Main Methods:

  • A dual-branch LSTM-Transformer architecture was employed to capture temporal trends and long-range dependencies.
Keywords:
LSTM-TransformerMIMIC-IVdeep learningelectronic health recordsmortality predictionsepsis

Related Experiment Videos

  • A high-resolution feature set including vasopressor infusion rates and hourly urine output was constructed.
  • The model was trained on 13,788 adult sepsis patients from the MIMIC-IV database and validated on the eICU database.

    • Main Results:

    • The proposed model achieved an AUROC of 0.8139, outperforming seven mainstream baselines.
    • Feature importance analysis highlighted urine output and norepinephrine dosage as key predictive indicators.
    • External validation on the eICU database showed robust generalizability, with AUROC improving to 0.7347 after fine-tuning.

    Conclusions:

    • The LSTM-Transformer Fusion architecture effectively models "drug-physiology" interactions with low computational cost.
    • Explicitly modeling dynamic treatment responses enhances predictive accuracy and clinical interpretability.
    • The lightweight and generalizable model serves as a robust tool for early warning systems in intensive care settings.