Machine learning applied to multi-sensor information to reduce false alarm rate in the ICU

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Critical Care Medicine

Background:

• Intensive care unit (ICU) vital signs monitors exhibit high false alarm rates (FAR), ranging from 0.72 to 0.99.
• Existing expert-based rules for alarm detection can be unreliable when sensor data is missing.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) in reducing FAR for ICU vital signs monitoring.
• To evaluate ML model performance in diagnosing patient conditions using multi-sensor data, specifically when parameters are absent.

Main Methods:

• A random forest (RF) ML model was trained on tagged ICU multi-sensor data, simulating missing parameters.
• The RF model's performance was compared against full expert-based rules (FER) and partial expert-based rules (PER) using metrics like Youden index and positive predictive value (PPV).
• Seven clinical alarm scenarios were identified and tagged using an expert-based rules algorithm.

Main Results:

• The RF model achieved performance comparable to FER when all parameters were present.
• In the absence of 1-3 parameters, RF maintained high Youden index (0.94-0.97) and PPV (0.98-0.99), with a low FAR (0.01-0.02).
• PER performance significantly degraded with missing parameters (Youden index 0.54-0.8, PPV 0.76-0.88, FAR 0.17-0.39).

Conclusions:

• Machine learning models, particularly RF, demonstrate superior accuracy and lower FAR in ICU vital signs monitoring compared to partial expert rules, especially when sensor data is missing.
• The RF model's ability to fuse information from available sensors effectively compensates for missing data, maintaining diagnostic performance.
• This data-driven approach offers a promising solution to reduce alarm fatigue and improve patient care in ICUs.