Predicting brain function status changes in critically ill patients via Machine learning.

Area of Science:

• Critical Care Medicine
• Neurology
• Artificial Intelligence

Background:

• Forecasting shifts in acute brain dysfunction (ABD) status in intensive care units (ICUs) is challenging.
• This unpredictability complicates effective hospital resource allocation.

Purpose of the Study:

• To develop a machine learning model for predicting next-day brain function status changes in ICU patients.
• To improve the forecasting of acute brain dysfunction (ABD) to aid clinical decision-making.

Main Methods:

• A light gradient boosting machine was trained and validated on multicenter prospective adult ICU cohorts.
• Shapley additive explanations were used to identify key predictive factors for a compact model.
• Performance was compared against existing state-of-the-art models.

Main Results:

• The boosting model achieved an AUROC of 0.824, significantly outperforming existing models (AUROC 0.697).
• A compact model using 13 factors retained 99.4% of the boosting model's predictive performance.
• Both models demonstrated strong generalizability in external validation (AUROC 0.812).

Conclusions:

• The developed machine learning models accurately predict next-day brain function status changes in ICU patients.
• The compact model utilizes simple, clinically relevant inputs, enabling direct prospective deployment.
• These models can significantly aid in critical hospital resource allocation by forecasting brain function status.