Explainable unsupervised anomaly detection for healthcare insurance data

Area of Science:

• Health Informatics
• Data Science
• Machine Learning

Background:

• Healthcare waste and fraud pose significant challenges for insurers.
• Big data analytics and machine learning offer potential solutions for detection.
• Acquiring labeled data for training fraud detection models is difficult and expensive.

Purpose of the Study:

• To develop a machine learning workflow for detecting healthcare waste and fraud.
• To assist investigators in identifying practitioners with unusual resource utilization.
• To improve the efficiency of combating waste and fraud in health insurance.

Main Methods:

• Combined categorical embeddings, unsupervised anomaly detection, and Shapley additive explanations (SHAP).
• Applied techniques to high-cardinality categorical variables and anomaly detection.
• Utilized SHAP for model interpretability in healthcare insurance anomaly detection.

Main Results:

• Categorical embeddings significantly improved performance over standard methods.
• Unsupervised anomaly detection techniques generally outperformed traditional methods.
• The workflow successfully identified a novel anomalous trend among general practitioners.

Conclusions:

• The proposed workflow effectively detects healthcare providers with atypical behavior.
• It aids expert investigators in making informed decisions regarding potential fraud and overconsumption.
• This approach enhances the fight against waste and fraud in health insurance.