CNN-based health model using knowledge mining of influencing factors

In modern society, the number of chronic patients is increasing due to various causes, such as drinking, smoking, unhealthy lifestyles, and stress. Chronic diseases must be managed with constant care, but may get worse from various factors. With the development of information technology, healthcare technologies using health big data, machine learning, and reinforcement learning are attracting attention. Using these technologies, it is possible to predict potential diseases that may occur in the future by using data learning and clustering of similar data. To predict the potential for disease, we should research various models based on the convolutional neural network (CNN), which can identify knowledge objects from unstructured data such as medical data. However, the fully connected network structure of the CNN generally uses a large amount of memory. Another problem is that complexity increases with the number of layers. This causes the overfitting problem, which increases error. To solve this problem, this paper proposes a CNN-based health model using knowledge mining of influencing factors. The proposed method uses hidden layers of a double-layer structure within the CNN structure. The double-layer structure has the optimal conditions for classification, compared with a single layer that allows the AND/OR operations. First, the amount of data used is reduced by extracting influencing factors through multivariate analysis, and these influencing factors are used as input data. Significant influencing factors are extracted from the first hidden layer using the significance level. This improves accuracy, because it extracts data required for analysis. Common influencing factors appropriate for significance levels are extracted. Common influencing factors refer to correlated factors that can affect each other. In the second hidden layer, the correlations between influencing factors are discovered through a correlation coefficient, and they are classified into positive and negative factors. Furthermore, associated rules are discovered through knowledge mining from among the classified influencing factors. They are subdivided into influencing factors like obesity, high blood pressure, and diabetes through the rules of the discovered influencing factors. For performance evaluation, the root mean square error (RMSE) of the CNN model is evaluated according to the application of knowledge mining to the influencing factors. The evaluation of accuracy, computational load, complexity, and learning rate showed better results, compared with the existing method. Through the proposed health model, knowledge about the associations of various factors is derived. © 2019, Springer-Verlag London Ltd., part of Springer Nature.