An Explainable Detection Framework for Health Insurance Fraud via Temporal Capture and Confidence Assurance

Fraud detection is essential to managing health insurance funds, yet current research methods are too broad for specific medical contexts. This study aims to narrow this gap by dissecting the workflow of health service providers to create a focused detection framework. Through this approach, we not only achieve reliable prediction outcomes but also offer practical guidance for implementation in real-world medical practices.Health insurance systems are characterized by highly specialized workflows and data formats. Utilizing health insurance data, this study focuses on claims submitted by providers for comprehensive statistical analysis. Temporal features in varying claim sequences are significant indicators of fraud. Thus, an advanced long short-term memory (LSTM) model is used to analyze long claim sequences, capturing essential temporal and interval features. Furthermore, this study incorporates attention mechanism to adeptly aggregate sequences of varying lengths, facilitating focused analysis and prioritizing suspicious claims. To address data imbalance affecting classifier efficacy, conformal prediction is used to calibrate classifier biases. Besides, two design strategies offer probability guarantees for provider fraud detection under complex practical requirements. This method not only improves the classifier reliability but also helps to determining audit provider priorities, thereby improving audit fraud detection efficiency.The proposed framework was evaluated against nine prevalent models spanning machine learning, ensemble learning, and deep learning through comparative analysis and ablation experiments. It demonstrated superior performance, achieving an AUC value of 0.907 on a real-world dataset. Additionally, two conformal prediction schemes were implemented, showing effectiveness in coverage priority and size priority, with a coverage rate up to 0.992. Illustrative examples further highlight the practical value of attention weights and prediction sets in providing actionable fraud detection insights.This study constructs a framework for detecting health insurance fraud. Crucially, this design paradigm can provide actionable guidance for the practice of medical managers.