Data-Centric AI for Healthcare Fraud Detection.

Area of Science:

• Health Informatics
• Machine Learning
• Data Science

Background:

• Automated detection of healthcare fraud is crucial for cost savings and patient care quality.
• Existing methods often rely on limited datasets and suboptimal evaluation techniques.
• Medicare claims data offers a rich resource for developing robust fraud detection models.

Purpose of the Study:

• To present a data-centric methodology for improving healthcare fraud classification.
• To create and enrich large-scale labeled datasets from Medicare claims for supervised learning.
• To propose an adjusted cross-validation technique for reliable model evaluation.

Main Methods:

• Curated nine large-scale labeled datasets from Medicare Part B, Part D, and DMEPOS claims (2013-2019).
• Enriched original datasets with up to 58 new provider summary features.
• Implemented an improved data labeling process and an adjusted cross-validation technique to mitigate target leakage.

Main Results:

• The newly enriched datasets consistently outperformed original Medicare datasets in fraud classification tasks.
• Extreme gradient boosting and random forest models demonstrated significant improvements with the enhanced data.
• The adjusted cross-validation technique provided more reliable evaluation results.

Conclusions:

• A data-centric machine learning workflow is highly effective for healthcare fraud detection.
• Data enrichment and improved evaluation are key to enhancing model performance and reliability.
• This study provides a strong foundation for future machine learning applications in combating healthcare fraud.