Early Prediction of Sepsis Based on Machine Learning Algorithm.

Area of Science:

• Computational biology and bioinformatics
• Medical informatics and machine learning applications

Background:

• Sepsis is a life-threatening organ dysfunction caused by infection, characterized by high mortality rates.
• Accurate and early prediction of sepsis is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop and compare machine learning models for predicting sepsis onset 6 hours in advance.
• To evaluate the effectiveness of different data processing methods, including mean processing and feature generation.

Main Methods:

• Utilized XGBoost and LightGBM machine learning algorithms for sepsis prediction.
• Developed a feature generation method combining statistical, window, and medical features.
• Employed the Miceforest multiple imputation method to handle significant missing data.

Main Results:

• The feature generation method significantly outperformed the mean processing method in predictive accuracy.
• Both XGBoost and LightGBM demonstrated excellent prediction performance, with AUC values ranging from 0.910 to 0.979.
• LightGBM showed superior running speed and generalization ability, achieving an AUC of 0.979 with the feature generation method.

Conclusions:

• Machine learning, particularly the feature generation method with LightGBM, offers a powerful approach for early sepsis prediction.
• Key predictors for early sepsis include Prothrombin Time (PTT), White Blood Cell count (WBC), and platelet levels.
• The findings highlight the potential for improved sepsis management through advanced computational techniques.