Procedure code overutilization detection from healthcare claims using unsupervised deep learning methods.

Area of Science:

• Machine Learning
• Healthcare Analytics
• Data Science

Background:

• Fraud, Waste, and Abuse (FWA) in medical claims negatively impacts healthcare quality and cost.
• Procedure code overutilization, where procedures are irrelevant to diagnosis or patient profile, is a major FWA component.
• Identifying unwarranted procedures is crucial for mitigating FWA in healthcare claims.

Purpose of the Study:

• To identify unwarranted procedure codes within millions of healthcare claims using unsupervised machine learning.
• To apply deep autoencoders and compare their effectiveness against a density-based clustering model for FWA detection.
• To address the challenge of identifying FWA in the absence of labeled data.

Main Methods:

• Utilized deep autoencoders to detect anomalous procedure codes indicative of FWA in healthcare claims.
• Employed diagnoses, procedures, and demographic data as features for the machine learning models.
• Compared autoencoder performance against a baseline density-based clustering model using datasets of 100,000 and 33 million claims.

Main Results:

• The autoencoder model, particularly with a novel feature-weighted loss function, outperformed the density-based clustering approach.
• Performance was evaluated using synthetic and manually annotated outlier datasets.
• On the 33 million claims dataset, the autoencoder achieved precision, recall, and F1-scores of 0.48, 0.90, and 0.63, respectively, on manually annotated data.

Conclusions:

• Unsupervised, deep-learning methods are feasible for identifying potential procedure overutilization in healthcare claims.
• The developed autoencoder model demonstrates effectiveness in detecting FWA indicators.
• This approach offers a promising solution for enhancing the integrity of medical billing and reducing healthcare costs.