Supervised Machine-Learning Algorithms in Real-time Prediction of Hypotensive Events.

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Data Science

Background:

• Hypotension is a frequent complication in critically ill patients within Intensive Care Units (ICUs).
• Early detection of hypotensive events is crucial for timely intervention and preventing organ damage.

Purpose of the Study:

• To evaluate supervised machine learning algorithms for predicting acute hypotensive events in the ICU.
• To assess the effectiveness of a real-time labeling technique in conjunction with these algorithms.
• To develop a method for reducing false alarm rates from predictive models.

Main Methods:

• Investigated various supervised machine learning classification algorithms, including logistic regression and Support Vector Machines (SVM).
• Employed a real-time labeling technique for event prediction.
• Developed and implemented a high-level decision-making algorithm to filter false positives.

Main Results:

• Logistic regression achieved 85% sensitivity and 96% specificity with 81% positive predictive value (PPV).
• SVM demonstrated 83% sensitivity and 96% specificity with 82% PPV.
• The proposed decision-making algorithm filtered 24% of false alarms, saving significant medical staff time and reducing monitor disturbances.

Conclusions:

• Supervised machine learning, particularly logistic regression and SVM, shows strong performance in predicting ICU hypotension.
• The developed decision-making algorithm effectively reduces false alarms, enhancing clinical workflow and patient comfort.
• This approach offers a promising tool for proactive management of hypotension in critical care settings.