Identifying fraud in medical insurance based on blockchain and deep learning

Highlights

• Presents a novel framework to identify fraud of medical insurance effectively.

• Propose an explainable method to evaluate the reasonability of disease diagnosis.

• Design a protocol of medical data storage and access based on consortium blockchain.

• The blockchain is applied to ensure the data is secure, immutable, and traceable.

• The evaluation of the proposed approach is conducted on two real datasets.

Abstract

With the rapid growth of medical costs, the control of medical expenses has been becoming an important task of Health Insurance Department. Traditional medical insurance settlement is paid on a per-service basis, which leads to lots of unreasonable expenses. To cope with this problem, the single-disease payment mechanism has been widely used in recent years. However, the single-disease payment also has a risk of fraud. In this work, we propose a framework to identify fraud of medical insurance based on consortium blockchain and deep learning, which can recognize suspicious medical records automatically to ensure valid implementation on single-disease payment and lighten the work of medical insurance auditors. An explainable model BERT-LE is designed to evaluate the reasonability of ICD disease code for Medicare reimbursement by predicting the probability of a disease according to the chief complaint of a patient. We also put forward a storage and management process of medical records based on consortium blockchain to ensure the security, immutability, traceability, and auditability of the data. The experiments on two real datasets from two 3A hospitals demonstrate that the proposed solution can identify fraud effectively and greatly improve the efficiency in medical insurance reviews.

Introduction

With the rapid development of medical informatization, the information systems of hospitals have accumulated a large amount of data, which makes the medical industry enter the era of big data. Medical big data has brought tremendous value to medical field, and attracted lots of attention from both academia and industry [1]. The control of medical expense is an important research branch of medical big data.

In the traditional medical insurance system, medical expenses are paid on the basis of medical service items, which leads to excessive medical treatment and rising of medical expenses. To cope with the problems in item-based charge mechanism, the single-disease payment based on Diagnosis-Related Groups (DRGs) has been extensively studied and applied [2]. In the single-disease payment mechanism, a fixed payment standard is determined for each disease [3]. The social medical insurance agency pays the patients’ hospitalization fees to hospitals according to the prescribed standard of each disease. To better understand the difference between single-disease payment mechanism and item-based payment mechanism, let us see an example. Suppose that a patient’s diagnosis is pneumonia, according to the traditional medical insurance settlement mechanism, if the expenses covered by the medical insurance are 100,000 yuan and the reimbursement ratio is 90%, then the reimbursement expenses will be 90,000 yuan. In the DRGs payment model, assume that the DRG weighting rate of pneumonia is 6.2 and the unit payment price is 14,000 yuan, then the medical insurance reimbursement will be $6.2\times \text{14,000}=\text{86,800}$ yuan. After applying single-disease payment mechanism, the income of medical institutions is only related to each medical case and diagnosis and has nothing to do with the actual cost of treating a patient. If the treatment cost of a disease exceeds the standard of its group, the hospital need pay for the extra cost, so that the medical insurance expenditure can be controlled more effectively. It can standardize the medical resource utilization, that is, the consumption of medical institutions is directly proportional to the number of inpatients, the complexity of diseases, and the intensity of services. In short, the single-disease payment model stipulates a fixed medical insurance payment standard for each disease to avoid excessive medical behavior, and thereby decreases medical expenses. This mechanism guarantees the quality of medical services and is easy to operate.

However, the single-disease payment model may cause medical insurance fraud. For instance, when assigning the diagnostic code for an inpatient, the health care provider may change the actual low-cost disease code to another high-cost disease code to gain more revenue from medical insurance agency. Due to the large number of inpatients, the cost of reviewing each inpatient’s medical records manually is extremely high. Take Jiangsu province of China for example, the number of inpatients each year is about 15 million, so it is impossible to audit each discharge diagnosis manually. Therefore, how to efficiently detect possible fraud has become an urgent problem in the single-disease payment mechanism. In the light of the above problem, we propose a framework to identify fraud of medical insurance based on blockchain and deep learning. In the framework, we transform the fraud identification problem into a text classification problem, and design an explainable model BERT-LE to estimate the reasonability of disease diagnosis by predicting the probability of diagnostic code according to inpatients’ chief complaint. Only the identified abnormal diagnostic codes need to be manually audited, which can significantly improve the audit efficiency and reduce the workload of insurance auditing. In addition, in order to preserve the evidence of medical insurance fraud, we put forward a protocol of medical data storage and management based on consortium blockchain. The medical records are digitally signed and stored in the blockchain, which ensures the security and tamper-proofing of data, and at the same time enables the traceability and non-repudiation of fraudulent behavior. We also set up a credit rating mechanism for doctors and hospitals based on blockchain to carry out effective medical insurance supervision.

To validate our solution, we carry out experiments on two real datasets which contains 620,000 inpatients’ medical records from two large hospitals. The results reveal that our model works well for medical insurance anti-fraud and has good explainability. Our proposed model can also help health departments evaluate the quality of medical records and aid doctors in giving accurate diagnostic codes.

The contributions of this paper are summarized as follows:

• We propose a framework to transform the medical insurance anti-fraud problem into a text classification problem. The framework can effectively identify medical insurance fraud and reduce the workload of auditors.

• We design a label embedding method to explain the logic of classification decisions, which can help users better understand and trust our model.

• We put forward a protocol of medical data storage and access based on consortium blockchain, which can prevent the illegal tampering of medical records and make medical records traceable and auditable.

• We conduct lots of experiments on two real datasets from large hospitals, and the results show that our model can play an effective role in anti-fraud for medical insurance and have good explainability.

The rest of this paper is organized as follows. In Section 2, the preliminaries and problem statement are presented. In Section 3, an explainable deep learning-based method to identify fraud of medical insurance is described. In Section 4, the blockchain-based medical records management process is elaborated. In Section 5, the proposed framework is evaluated and discussed. In Section 6, previous studies are reviewed. Finally, Section 7 concludes the paper and provides some future work discussions.

A preliminary version of this work has been reported in a conference short paper [4].