Deep Belief Network Model for Detection of an Outlier in Healthcare Data

Abstract

Sports wristbands provide a rich source of information for a thorough understanding of people's physical conditions in the light of the popularisation of intelligent wearable gadgets. Outlier detection is still important since there are unknown outliers in the multi-dimensional activity data it supplies. Traditional methods of density estimation are hindered by the "curse of dimensionality," resulting in poor detection results. A Gaussian mixture generative model (GMGM) health data detection method is employed to address this issue. To begin, the model trains the original data with a variational autoencoder (VAE) and recovers latent features by lowering the reconstruction error. The latent distribution and extracted attributes are then utilised to forecast the varied membership of the samples using a deep belief network (DBN). Then, to prevent the effects of model decoupling, the variational autoencoder, deep belief network, and Gaussian mixture model (GMM) are optimised together. The Gaussian mixture model predicts the sample density of each data set and considers samples with densities more than the threshold as anomalies during the training phase. On the ODDS standard dataset, the model's performance is tested. The results reveal that the AUC index of GMGM is enhanced by 5.5 percent points on average when compared to the deep autoencoder Gaussian mixture model (DAGMM). Finally, the method's usefulness is demonstrated by the experimental findings on real datasets.