CNN-based health model for regular health factors analysis in internet-of-medical things environment

Abstract

Remote health monitoring applications with the advent of Internet of Things (IoT) technologies have changed traditional healthcare services. Additionally, in terms of personalized healthcare and disease prevention services, these depend primarily on the strategy used to derive knowledge from the analysis of lifestyle factors and activities. Through the use of intelligent data retrieval and classification models, it is possible to study disease, or even predict any abnormal health conditions. To predict such abnormality, the Convolutional neural network (CNN) model is used, which can detect the knowledge related to disease prediction accurately from unstructured medical health records. However, CNN uses a large amount of memory if it uses a fully connected network structure. Moreover, the increase in the number of layers can lead to an increase in the complexity analysis of the model. Therefore, to overcome these limitations of the CNN-model, we propose a CNN-regular target detection and recognition model based on the Pearson Correlation Coefficient and regular pattern behavior, where the term 'regular' denotes objects that generally appear in similar contexts and have structures with low variability. In this framework, we develop a CNN-regular pattern discovery model for data classification. First, the most important health-related factors are selected in the first hidden layer, then in the second layer, a correlation coefficient analysis is conducted to classify the positively and negatively correlated health factors. Moreover, regular patterns' behaviors are discovered through mining the regular pattern occurrence among the classified health factors. The output of the model is subdivided into regular-correlated parameters related to obesity, high blood pressure, and diabetes. Two distinct datasets are adopted to mitigate the effects of the CNN-regular knowledge discovery model. The experimental results show that the proposed model has better accuracy, and low computational load, compared with three different machine learning techniques methods.