Predicting Future Occurrence of Acute Hypotensive Episodes Using Noninvasive and Invasive Features

Area of Science:

• Critical care medicine
• Biomedical informatics
• Machine learning applications

Background:

• Early prediction of acute hypotensive episodes (AHE) in intensive care units (ICUs) is crucial for improving patient outcomes.
• Machine learning (ML) offers potential for developing predictive models using large ICU datasets.

Purpose of the Study:

• To apply and compare various ML algorithms for predicting AHE 30 minutes in advance.
• To evaluate the impact of invasive versus noninvasive features on AHE prediction accuracy.
• To predict continuous mean arterial pressure (MAP) for deeper clinical insights.

Main Methods:

• Utilized the MIMIC III Physionet dataset (>60,000 ICU records).
• Applied five classification algorithms: K-nearest neighbor, logistic regression, support vector machine (SVM), random forest, and long short-term memory (LSTM).
• Employed linear regression for continuous MAP prediction and analyzed feature inclusion effects.

Main Results:

• SVM achieved the highest recall (84%) for AHE prediction.
• Inclusion of invasive features significantly improved both recall and precision by over 20 percentage points.
• Achieved 91% recall and 68% precision for AHE prediction after converting continuous MAP forecasts.
• Predicted MAP with a root mean square error of 10 mmHg 60 minutes ahead.

Conclusions:

• ML models can effectively predict AHE with high precision and recall 30 minutes in advance.
• Invasive features enhance predictive model performance for AHE.
• Continuous MAP prediction offers interpretable clinical signals for timing and severity of AHE.