Advancing shock prediction: leveraging prior knowledge and self-controlled data for enhanced model accuracy and...

Area of Science:

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

Background:

• Timely intervention in shock is critical, as delays exceeding one hour significantly increase mortality.
• Existing prediction methods often rely on invasive tests or lack sufficient lead time.

Purpose of the Study:

• To develop an enhanced machine learning model for predicting shock one hour in advance.
• To improve predictive performance using self-controlled physiological waveform data and medical knowledge-based feature engineering.
• To enable early shock prediction without the need for blood tests.

Main Methods:

• Utilized patient data and physiological waveforms from the MIMIC-3 database.
• Defined shock based on hypotension and elevated lactate levels.
• Extracted 30-minute waveform segments preceding the shock event and self-controlled segments for comparison.
• Engineered 299 features from arterial blood pressure, electrocardiogram, respiratory, and SpO2 waveforms.

Main Results:

• The study included 389 ICU patients meeting shock criteria.
• A weighted ensemble model achieved an Area Under the Curve (AUC) of 0.93, 84.15% accuracy, and 79.64% sensitivity.
• Key predictive features included heart rate variability, respiratory cycle characteristics, and blood pressure waveform dynamics.

Conclusions:

• Demonstrated the feasibility of predicting shock one hour prior to onset using physiological waveforms.
• The developed model shows strong predictive performance (high AUC and sensitivity).
• Highlights the potential of waveform analysis and feature engineering for early clinical deterioration detection.