Feature Importance Analysis for Compensatory Reserve to Predict Hemorrhagic Shock

Area of Science:

• Biomedical Engineering
• Physiology
• Trauma Care

Background:

• Hemorrhage is a leading cause of preventable trauma death.
• Traditional vital signs lack sensitivity for early detection of blood loss due to compensatory mechanisms.
• Machine learning on arterial blood pressure (ABP) waveforms shows promise but lacks interpretability.

Purpose of the Study:

• To evaluate the importance of nine physiologically interpretable ABP-derived features for detecting central hypovolemia.
• To identify a reliable, interpretable feature for early detection of blood loss and hemorrhagic shock.
• To compare the performance of interpretable features against complex machine learning models.

Main Methods:

• Utilized a lower-body negative pressure model to induce progressive central hypovolemia in 40 human subjects.
• Analyzed arterial blood pressure waveforms to derive nine distinct physiological features.
• Employed linear regression and receiver operating curve analysis to assess feature performance.

Main Results:

• The half-rise to dicrotic notch (HRDN) feature emerged as significantly more important than others.
• HRDN measures the time delay between ejected and reflected ABP wave components, indicating compensatory mechanisms.
• Linear regression yielded an RMSE of 16.9% and R2 of 0.72; AUC for decompensation detection was 0.88.

Conclusions:

• A single, physiologically interpretable feature (HRDN) from ABP waveforms effectively monitors blood loss and impending hemorrhagic shock.
• HRDN performance is comparable to complex, less interpretable machine learning models.
• This finding offers a clinically relevant tool for early detection of hypovolemia in trauma patients.