Development and Validation of a Prediction Model for Acute Hypotensive Events in Intensive Care Unit Patients

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Persistent hypotension in the intensive care unit (ICU) is a significant predictor of increased mortality.
• Timely intervention for acute hypotensive events can improve patient outcomes.
• Developing accurate prediction models for these events is essential for proactive patient management.

Purpose of the Study:

• To develop and validate a predictive model for acute hypotensive events in ICU patients.
• To compare the performance of different machine learning algorithms for hypotension prediction.
• To identify the most effective features for predicting acute hypotensive events.

Main Methods:

• Adult patients from Nagoya City University (NCU) Hospital ICU (2018-2021) were used for internal validation.
• The MIMIC-III database was utilized for external validation of the prediction models.
• Machine learning algorithms including logistic regression, LightGBM, and LSTM were trained using vital signs and demographic data.
• A hypotensive event was defined as mean arterial pressure < 60 mmHg for at least 5 minutes within a 10-minute window.

Main Results:

• Acute hypotensive events occurred in 74.6% of NCU admissions and 51.1% of MIMIC-III admissions.
• The LightGBM model achieved the highest Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.835 in internal validation.
• The LSTM model, using only blood pressure-related features, demonstrated the highest AUROC of 0.843 and showed consistent performance across internal and external validation.

Conclusions:

• The Long Short-Term Memory (LSTM) model, when utilizing solely blood pressure-related features, demonstrated superior predictive performance for acute hypotensive events.
• The LSTM model's effectiveness was validated through comparable results in both internal and external datasets.
• These findings suggest that LSTM models focusing on hemodynamic data offer a promising approach for early detection of hypotension in critical care settings.