A machine learning method for predicting the probability of MODS using only non-invasive parameters.

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Timely prediction of Multiple Organ Dysfunction Syndrome (MODS) is crucial for trauma patient survival.
• Current MODS prediction methods are often invasive, prone to artifacts, and impractical in pre-hospital settings.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting MODS using solely non-invasive parameters.
• To compare the performance of ML models against traditional scoring systems for MODS prediction.

Main Methods:

• Utilized data from 2319 patients in the MIMIC-III database, employing Sequential Organ Failure Assessment (SOFA) scores.
• Developed real-time MODS prediction models using seven ML methods with 57 total parameters and 17 non-invasive parameters.
• Compared ML model performance (AUC) against four traditional scoring systems.

Main Results:

• LightGBM (LGBM) and Adaboost models using all parameters achieved an Area Under the ROC Curve (AUC) of 0.959.
• A reduced model with 17 non-invasive parameters still outperformed traditional scoring systems, with a minimal AUC decrease of 0.015.
• ML models demonstrated superior predictive performance compared to conventional scoring systems.

Conclusions:

• A novel, real-time MODS prediction method based on non-invasive parameters was successfully developed.
• This non-invasive ML approach offers a significant improvement over traditional scoring systems for early MODS diagnosis.
• The developed method has the potential to enhance pre-hospital trauma care and improve patient survival rates.