Researchers develop conditional anomaly detection method using soft harmonic functions for clinical alerting

Researchers developed a non-parametric approach for conditional anomaly detection based on soft harmonic solutions, designed to identify data instances with unusual responses such as omission of important lab tests in clinical settings.

The method estimates confidence of labels to detect anomalous mislabeling and includes regularization to avoid detecting isolated examples and examples on the boundary of the distribution support.

The approach was demonstrated on a real-world electronic health record dataset and compared against several baseline approaches.