Evidence-based anomaly detection in clinical domains

Area of Science:

• Medical Informatics
• Machine Learning
• Clinical Decision Support

Background:

• Anomaly detection is crucial for identifying significant events in healthcare.
• Evaluating patient management decisions requires methods to identify unusual patterns.
• Existing methods may not adequately capture the nuances of patient-specific conditions.

Purpose of the Study:

• To develop and evaluate new probabilistic anomaly detection methods.
• To assess the ability to identify unusual management decisions for patients with similar conditions.
• To apply these methods to detect unusual hospitalization patterns in community-acquired pneumonia.

Main Methods:

• Utilizing probabilistic models, specifically Bayesian networks, learned from historical patient data.
• Developing statistical measures derived from these models to quantify unusualness.
• Evaluating detection performance using precision and recall metrics.

Main Results:

• The developed methods demonstrated encouraging performance in detecting unusual hospitalization patterns.
• Achieved a precision of 0.5 and a recall of 0.53 in the evaluation.
• Indicated the potential for these techniques in identifying concerning clinical events.

Conclusions:

• Probabilistic anomaly detection offers a promising approach for evaluating clinical decisions.
• These methods can form the basis for intelligent monitoring systems.
• Alerting clinicians to unusual events or decisions can improve patient care and outcomes.