Early Detection of Hypotension Using a Multivariate Machine Learning Approach.

Area of Science:

• Biomedical Engineering
• Machine Learning in Medicine
• Trauma Care Research

Background:

• Early hypotension detection in trauma is vital for improving patient outcomes.
• Limited research exists on combining multiple physiological signals for early hemorrhagic shock detection.
• This study investigates machine learning for early hypotension prediction using multivariate physiological data.

Purpose of the Study:

• To explore the viability of early hypotension detection using multiple physiological signals and machine learning.
• To develop and compare deep learning, ensemble learning, and classical learning models for predicting hypotension.
• To establish a proof of concept with a 5-minute prediction window for hypotension detection.

Main Methods:

• Utilized a publicly available dataset from experimental hemorrhage studies in sheep.
• Employed supervised machine learning: deep learning (LSTM), ensemble learning (Random Forest), and classical learning (SVM).
• Evaluated models using 3-fold cross-validation, comparing precision and recall for 5-minute hypotension prediction.

Main Results:

• Support Vector Machine (SVM) and Random Forest models outperformed LSTM networks.
• Random Forest achieved 84% recall and 56% precision; SVM achieved 62% recall and 82% precision.
• Electrocardiogram signals showed the highest importance, while arterial blood pressure had the least significance.

Conclusions:

• A multivariate approach using multiple physiological signals is potentially more effective for early hypotension detection than univariate methods.
• Machine learning models, particularly SVM and Random Forest, demonstrate potential for early hypotension prediction in trauma.
• Further research is warranted to validate these findings in clinical settings.