Table 1.

Summary of related works on fraud detection machine learning (ML) and deep learning (DL) algorithms in health insurance claims.

Algorithm under study	Number of fraud scenarios detected	Considering privacy and security	Considering bias issue	ML or DL	Throughput	Latency	CPUa use	Memory use	Data set	Metrics
MHAMFDb [16]	NRc	Xd	X	DL	X	X	X	X	Medical-1: balanced data set with a ratio of positive to negative samples of 1:2 Medical-2: unbalanced data set with a ratio of positive to negative samples of approximately 1:70	Medical-1: Accuracy: 0.8961 F1-score: 0.8694 Medical-2: F1-score: 0.8361 Recall: 0.8764 Precision: 0.9194
GSVMse [17]	NR	X	X	ML	X	X	X	X	100-claim data set 300-claim data set 500-claim data set 750-claim data set 1000-claim data set	100-claim data set accuracy: 71.43% 300-claim data set accuracy: 95.45% 500-claim data set accuracy: 99.18% 750-claim data set accuracy: 82.56% 1000-claim data set accuracy: 90.91%
WMTDBCf [18]	NR	X	X	ML	X	X	X	X	The data set used in the study was the claims data submitted by health care providers under the US Medicare CMSg Part B health care program	Overall accuracy ranged from 0.857 to 0.946.

^aCPU: Central Processing Unit.

^bMHAMFD: Multilevel Hierarchical Attention Mechanism for Fraud Detection.

^cNR: not reported.

^dX: not considered.

^eGSVM: Genetic Support Vector Machine.

^fWMTDBC: Weighted MultiTree Density-Based Clustering.

^gCMS: Centers for Medicare & Medicaid Services.