Development and external validation of machine learning-based models to predict patients with cellulitis developing...

Area of Science:

• Medical informatics
• Clinical prediction modeling
• Machine learning in healthcare

Background:

• Cellulitis is a common cause of hospitalization, with high mortality rates associated with sepsis.
• Existing stratification models for predicting sepsis in cellulitis patients have shown unsatisfactory performance in external validation.

Purpose of the Study:

• To develop and compare various machine learning models for predicting sepsis development in hospitalized patients with cellulitis.
• To evaluate the performance and robustness of these models in external validation.

Main Methods:

• Retrospective cohort study involving two independent international cohorts.
• Development phase: 6695 patients with cellulitis (MIMIC-IV database) using machine learning algorithms.
• External validation phase: 2506 patients with cellulitis (YiduCloud database).

Main Results:

• In internal validation, XGBoost achieved the highest AUC (0.780).
• In external validation, the Artificial Neural Network (ANN) model demonstrated the highest AUC (0.830), outperforming logistic regression (LR).
• Boosting and ANN models exhibited greater robustness compared to LR when variables were removed.

Conclusions:

• Boosting and neural network models offer improved performance and robustness for predicting sepsis in cellulitis patients.
• These models can serve as valuable tools for early detection of sepsis in hospitalized cellulitis patients.