2026 Critical Care Congress Skip to main content
2026 Critical Care Congress Main banner
Final Program


Advance Program
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Session Icons
Thought Leader Session
Thought Leader Session
Guidelines Session
Guidelines Session
Aging Adult Track
Aging Adult Track
Additional Registration Fee
Additional Registration Fee

Research Snapshot Theater: Future of Critical Care, Adult I

Developed Machine Learning Early Warning System for Inpatient Deterioration: Comparative With NEWS

Sunday, March 22, 2026
12:15 PM - 1:15 PM Central Time
Location: Connections Central - RST 03

  • Askas Samansee, n/a (he/him/his)

    Narathiwat, THA

    Disclosure(s): No relevant financial relationship(s) to disclose.

  • SS

    Supakorn Suntudkan, n/a

    Medical Student, Doctor of Medicine Program (Research Development)
    Prince of Songkla University, Nakhon Si Thammarat, Thailand

    Disclosure information not submitted.

  • NA

    Nawaporn Assanangkornchai, n/a

    Critical care medicine, Internal medicine
    Prince of Songkla University, Songkhla, Thailand

    Disclosure information not submitted.

Introduction: Clinical deterioration during hospitalization significantly increases morbidity and mortality. Early recognition is critical but often delayed. National Early Warning Score (NEWS) has been widely used for detecting risk of intensive care unit (ICU) admission and death within 24 hours but limited by inability to capture complex relationships among multiple variables. Machine learning (ML) offers an approach to overcome these limitations by learn patterns from complex, high-dimensional data. This study aimed to develop and validate an ML model to predict unanticipated ICU admission (UIA) and in-hospital mortality (IHM) on medical wards, and to compare with the NEWS.


Methods: This retrospective study utilized electronic health information system data from Songklanagarind Hospital, Thailand (Jan 2014–May 2024). Patients over 18 years old admitted to the medical inpatient unit were included. Model development variables comprised demographic, vital signs, and basic laboratory. Data were randomly divided into training, testing, and validation sets. The outcomes were UIA and IHM in 6 hours. Various ML methodologies and deep learning models were employed. The NEWS served as the comparator. Evaluation was used accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUROC) for comparison.


Results: The study included 3,172 and 1,606 admissions for training and validation, 500 and 402 admissions for testing of UIA and IHM prediction respectively. The UIA rate was 4.55%, and the overall IHM rate was 3.38%. Patients with UIA had a mortality rate of 21.69%. Among the prediction models developed, Weighted Ensemble model showed the best performance. The model compared to NEWS on UIA, an AUROC of 0.84vs0.63;p < 0.0001, accuracy of 82%vs73%, sensitivity of 38%vs42%, and specificity of 96%vs83% respectively. In order of IHC an AUROC of 0.95vs0.63;p < 0.0001, accuracy of 93%vs81%, sensitivity of 79%vs78%, and specificity of 97%vs82% respectively.


Conclusions: This study demonstrates that our ML model is superior to the NEWS in predicting both ICU admission and IHM. This finding highlights the potential of our model as an effective screening tool for identifying high-risk patients, suggesting it could establish a new standard in clinical practice for medical inpatient care.