Length of stay prediction for ICU patients using individualized single classification algorithm

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Prediction Models

Background:

• Accurate prediction of length of stay (LOS) in intensive care units (ICUs) is crucial for optimizing treatment and patient management.
• Existing universal prediction models often fail to meet the individualized needs of special ICU patients.
• Personalized LOS prediction is essential for tailored medical interventions.

Purpose of the Study:

• To develop a personalized model for predicting patient length of stay (LOS) in the ICU.
• To enhance patient care by moving beyond generic prediction models.
• To improve the accuracy of predicting hospitalization duration for individual patients.

Main Methods:

• Proposed a novel combination of just-in-time learning (JITL) and one-class extreme learning machine (one-class ELM), termed one-class JITL-ELM.
• Utilized JITL to identify personalized patient cases for comparison.
• Employed one-class ELM to predict patient discharge within 10 days.

Main Results:

• The one-class JITL-ELM model achieved an Area Under the Curve (AUC) of 0.8510 and a precision of 1.
• Model performance metrics included a G-mean of 0.7842, accuracy of 0.82, specificity of 1, and sensitivity of 0.6150.
• A novel mortality risk estimation system combining LOS and age demonstrated 66% accuracy with a 6.25% miss rate.

Conclusions:

• The one-class JITL-ELM model accurately predicts hospitalization days and patient mortality using early physiological parameters.
• A simple, interpretable mortality risk estimation system based on LOS and age offers significant clinical application value.
• Personalized prediction models offer a valuable advancement over universal models in critical care settings.