Weakly Supervised Classification of Vital Sign Alerts as Real or Artifact.

Area of Science:

• Medical Technology
• Artificial Intelligence in Medicine
• Clinical Engineering

Background:

• A significant percentage of clinical physiologic monitoring alarms are false, leading to alarm fatigue among healthcare professionals.
• Alarm fatigue compromises patient safety by potentially causing clinicians to ignore critical alerts.
• Existing Machine Learning (ML) approaches for adjudicating vital sign (VS) alerts often rely on supervised methods requiring extensive manual data labeling.

Purpose of the Study:

• To investigate the efficacy of weak supervision techniques for developing ML models to classify bedside vital sign alerts.
• To reduce the reliance on costly and time-consuming manual data annotation in healthcare ML applications.
• To provide a practical alternative to supervised learning for ML in healthcare.

Main Methods:

• Exploration of weak supervision using multiple, imperfect heuristics to automatically generate probabilistic labels for unlabeled training data.
• Development and evaluation of weakly supervised ML models for adjudicating vital sign alerts.
• Comparison of weakly supervised model performance against traditional supervised ML techniques.

Main Results:

• Weakly supervised models demonstrated competitive performance compared to traditional supervised methods.
• The proposed approach significantly reduces the need for extensive manual data labeling by domain experts.
• Weak supervision offers an efficient and practical alternative for ML implementation in healthcare.

Conclusions:

• Weak supervision is a viable and effective strategy for developing ML models in healthcare settings.
• This method addresses the data labeling bottleneck, facilitating wider ML adoption in clinical practice.
• The findings support the use of weakly supervised ML for improving patient safety through accurate alert adjudication.