Identifying High-Risk Patients without Labeled Training Data: Anomaly Detection Methodologies to Predict Adverse...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Patient risk stratification

Background:

• Adverse outcomes in many clinical conditions affect a small patient subset.
• Traditional risk stratification requires large datasets, which are costly and time-consuming to acquire.
• Existing methods may be inadequate for novel or rare clinical phenomena.

Purpose of the Study:

• To explore anomaly detection techniques for identifying high-risk patients.
• To treat high-risk patients as outliers in sparse feature space regions.
• To evaluate the efficacy of anomaly detection for surgical patient risk stratification.

Main Methods:

• Investigated three categories of anomaly detection: classification-based, nearest neighbor-based, and clustering-based.
• Applied these methods to identify patients at elevated risk of adverse outcomes.
• Utilized data from the National Surgical Quality Improvement Program (NSQIP).

Main Results:

• Anomaly detection methods successfully identified patients at higher risk of mortality.
• These techniques also identified patients with elevated risk for rare morbidities.
• The methods proved effective in identifying high-risk cases within the NSQIP dataset.

Conclusions:

• Anomaly detection offers a viable approach for identifying high-risk surgical patients.
• This method can efficiently detect patients with elevated risks of mortality and rare morbidities.
• Anomaly detection provides a valuable tool for risk stratification in surgical outcomes.