Detecting Unusual Intravenous Infusion Alerting Patterns with Machine Learning Algorithms.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Patient Safety

Background:

• Intravenous administration of high-alert medications requires robust safety monitoring.
• Current methods for detecting unusual infusion alert patterns may not be sufficiently sensitive.
• Machine learning (ML) offers potential for advanced anomaly detection in clinical data.

Purpose of the Study:

• To evaluate the efficacy of unsupervised multivariate ML algorithms in detecting unusual infusion alerting patterns.
• To compare ML-based anomaly detection with traditional control chart methods.
• To advance safer inpatient intravenous medication administration.

Main Methods:

• Utilized one year of detailed propofol infusion data.
• Engineered interpretable clinical variables and aggregated data per day.
• Compared a univariate moving range (mr) control chart with three ML algorithms: Local Outlier Factor, Isolation Forest, and k-Nearest Neighbors.

Main Results:

• The mr-chart identified 15 alert pattern anomalies.
• ML algorithms detected 31 unique anomalies, with 10% agreement across all algorithms and 36% agreement across at least two.
• Each ML algorithm identified specific anomalies missed by the mr-chart, including a day with a high rate of overridden alerts.

Conclusions:

• Unsupervised ML algorithms demonstrate superior robustness compared to control charts for detecting unusual infusion alert patterns.
• A combination of ML algorithms provides a benchmarking function, improving the focus on high-agreeability anomalies.
• ML algorithms can assist clinicians in identifying critical alert patterns, contributing to safer infusion practices.