Posted by Admin: System Admin
Patients depend on health insurance provided by the government systems, private systems, or both to utilize the high-priced healthcare expenses. This dependency on health insurance draws some healthcare service providers to commit insurance frauds. Although the number of such service providers is small, it is reported that the insurance providers lose billions of dollars every year due to frauds. In this paper, we formulate the fraud detection problem over a minimal, definitive claim data consisting of medical diagnosis and procedure codes. We present a solution to the fraudulent claim detection problem using a novel representation learning approach, which translates diagnosis and procedure codes into Mixtures of Clinical Codes (MCC). We also investigate extensions of MCC using Long Short Term Memory networks and Robust Principal Component Analysis. Our experimental results demonstrate promising outcomes in identifying fraudulent records. Machine learning is an important component of the growing field of data science. Through the use of statistical methods, different type of algorithms is trained to make classifications or predictions, and to uncover key insights in this project. These insights subsequently drive decision making within applications and businesses, ideally impacting key growth metrics. Machine learning algorithms build a model based on this project data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of datasets, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
Yang and Hwang developed a fraud detection model using the clinical pathways concept and process-mining framework that can detect frauds in the healthcare domain [13]. The method uses a module that works by discovering structural patterns from input positive and negative clinical instances. The most frequent patterns are extracted from every clinical instance using the module. Next, a feature-selection module is used to create a filtered dataset with labeled features. Finally, an inductive model is built on the feature set for evaluating new claims. Their method uses clustering, association analysis, and principal component analysis. The technique was applied on a real-world data set collected from National Health Insurance (NHI) program in Taiwan. Although the authors constructed different features to generate patterns for both normal and abusive claims, the significance of those features is not discussed. Bayerstadler et al. [14] presented a predictive model to detect fraud and abuse using manually labeled claims as training data. The method is designed to predict the fraud and abuse score using a probability distribution for new claim invoices. Specifically, the authors proposed a Bayesian network to summarize medical claims’ representation patterns using latent variables. In the prediction step, a multinomial variable modeling predicts the probability scores for various fraud events. Additionally, they estimated the model parameters using Markov Chain Monte Carlo (MCMC) [15]. Zhang et al. [16] proposed a Medicare fraud detection framework using the concept of anomaly detection [17]. First part of the proposed method consists of a spatial density based algorithm which is claimed to be more suitable compared to local outlier factors in medical insurance data. The second part of the method uses regression analysis to identify the linear dependencies among different variables. Additionally, the authors mentioned that the method has limited application on new incoming data. Kose et al. [18] used interactive unsupervised machine learning where expert knowledge is used as an input to the system to identify fraud and abuse related legal cases in healthcare. The authors used a pairwise comparison method of analytic hierarchical process (AHP) to incorporate weights between actors (patients) and attributes. Expectation maximization (EM) is used to cluster similar actors. They had domain experts involved at different levels of the study and produced storyboard based abnormal behavior traits. The proposed framework is evaluated based on the behavior traits found using the storyboard and later used for prescriptions by including all related persons and commodities such as drugs. Bauder and Khoshgoftaar [19] proposed a general outlier detection model using Bayesian inference to screen healthcare claims. They used Stan model which is similar to [20] in their experiments. Note that, they consider only provider level-fraud detection without considering clinical code based relations. Many of those methods use private datasets or different datasets with incompatible feature lists. Therefore, it is very difficult to directly compare these studies. In addition, HIPAA, GDPR and similar law enforce serious penalties for violations of the privacy and security of healthcare information, which make healthcare providers and insurance companies very reluctant to share rich datasets if not at all. For these reasons, we formulate the problem over a minimal, definitive claim data consisting of diagnosis and procedure codes. Under this setting we tackle the problem of flagging a procedure as legitimate or fraudulent using mixtures of clinical codes along with RNN and RPCA based encodings. Disadvantages Making false diagnoses to justify procedures that are not medically necessary. Fabricating claims for unperformed procedures. Performing medically unnecessary procedures to claiminsurance payments. Billing for each step of a procedure as if it is a separateprocedure, also called “unbundling”. Misrepresenting non-covered treatments as medicallynecessary to receive insurance payments, especially forcosmetic procedures.
We extend the MCC model using Long-Short Term Memory networks and Robust Principal Component Analysis. Our goal in extending MCC is to filter the significant concepts from claims and classify them as fraudulent or non-fraudulent. We extend MCC by using the concept weights of a claim as a sequence representation within a Long-Short Term Memory (LSTM) network. This network allows us to represent the claims as sequences of dependent concepts to be classified by the LSTM. Similarly, we apply Robust Principal Component Analysis (RPCA) to filter significant concept weights by decomposing claims into a low-rank and sparse vector representations. The low-rank matrix ideally captures the noise-free weights. Our unique contributions in this study can be summarized as follows. The system formulates the fraudulent claim detection problem over a minimal, definitive claim data consisting of procedure and diagnosis codes. The system introduces clinical concepts over procedure and diagnosis codes as a new representation learning approach. The system extends the mixtures of clinical concepts using LSTM and RPCA for classification. Advantages ? The proposed system uses Support Vector Machine (SVM) for classification with MCC. ? Multivariate Outlier Detection method is an effective method which is used to detect anomalous provider payments within Medicare claims data.