An Explainable Machine-Learning Model for Compensatory Reserve Measurement: Methods for Feature Selection and the...

Area of Science:

• Biomedical Engineering
• Physiological Monitoring
• Machine Learning in Medicine

Background:

• Accurate vital sign tracking is crucial for patient triage and timely intervention.
• Compensatory mechanisms can obscure injury severity, complicating patient assessment.
• Compensatory Reserve Measurement (CRM) from arterial waveforms aids early detection of hemorrhagic shock.

Purpose of the Study:

• To investigate classical machine learning models for estimating CRM using arterial waveform features.
• To identify specific arterial waveform elements predictive of CRM.
• To compare classical machine learning models with deep learning approaches for CRM estimation.

Main Methods:

• Extracted over 50 features from human arterial blood pressure datasets during simulated hypovolemic shock.
• Utilized progressive lower body negative pressure to induce controlled hypovolemia.
• Developed and optimized a bagged decision tree model using the ten most significant features for CRM estimation.

Main Results:

• The optimal bagged decision tree model achieved an average root mean squared error of 0.171 for CRM estimation.
• Performance was comparable to a deep-learning CRM algorithm (error of 0.159).
• Analysis revealed significant subject variability and differing key features across shock severity subgroups.

Conclusions:

• Classical machine learning models provide a transparent and effective method for CRM estimation.
• Identifying distinct features for different compensatory capacities can improve individual patient triage.
• This methodology holds potential for enhancing military and emergency medicine trauma patient assessment.