Aiming for Relevance.

Area of Science:

• Critical care medicine
• Biomedical informatics
• Machine learning

Background:

• Vital signs are essential for patient monitoring in intensive care units (ICUs).
• Traditional machine learning metrics like Root Mean Squared Error (RMSE) do not adequately reflect clinical significance in vital sign prediction.
• Accurate prediction of vital sign trajectories is crucial for early detection of adverse events.

Purpose of the Study:

• To introduce novel performance metrics for vital sign prediction that align with clinical relevance in ICUs.
• To address the limitations of conventional metrics in evaluating the clinical utility of predictive models.
• To develop and validate metrics focusing on deviations from clinical norms, overall trends, and trend deviations.

Main Methods:

• Derived novel metrics from empirical utility curves based on ICU clinician interviews.
• Validated the proposed metrics using simulated and real-world clinical datasets (MIMIC and eICU).
• Utilized these metrics as loss functions for training neural networks.

Main Results:

• Models trained with the novel metrics demonstrated superior performance in predicting clinically significant events.
• The new metrics effectively capture clinically relevant aspects of vital sign prediction beyond standard error measures.
• Validation confirmed the usefulness of the metrics across different datasets.

Conclusions:

• The developed vital sign prediction metrics offer a clinically meaningful approach to model evaluation and optimization.
• These metrics can lead to improved machine learning models for ICU patient care.
• This work facilitates the development of more effective AI-driven tools for critical care settings.