Deep Learning-Based Method of Diagnosing Hyperlipidemia and Providing Diagnostic Markers Automatically

Liu, Yuliang; Zhang, Quan; Zhao, Geng; Liu, Guohua; Liu, Zhiang

We also compared the performance of SVM and fully connected neural network with our model. The type of SVM is C-SVC, and the kernel function is RBF. It achieved 63% ACC with the same testing dataset. The SVM which used sigmoid kernel function polynomial kernel achieved 50% ACC and 81% ACC in the same testing data, respectively. Fully connected neural network achieved 89% ACC with test data mentioned above. We speculated that it is due to the fact that the traditional classification method is difficult to learn the relationship between different physiological parameters and cannot learn the importance of different physiological parameters for disease. Each physiological parameter is not independent. Actually, the parameters interact with each other. One physiological parameter is the same, while the other physiological parameters are different, which may reflect different health conditions. Different physiological parameters are interrelated in physiological mechanism. Like the semantic environment, the same words have different interpretations in different semantic environments. For example, when both lipids and HDL are high, patients may experience a temporary increase in blood lipids due to diet rather than hyperlipidemia. Moreover, HDL reflects the synthesis of lipid metabolism, and it is not the higher the better. We also compared the performance of the simple recurrent neural network (RNN) with the model proposed by this paper. This RNN model also used the Adam algorithm to update global parameters. It achieved 93% ACC in the test dataset mentioned above. The performance of these two models is very close. However, LSTM can better synthesize the relationship between different physiological parameters to give a judgment, and the simple RNN model only considers the state at the nearest moment. The more complex the data processed, the more obvious the difference in performance between the two models. We also found similar work in the limited range. Manjeevan Seeraa 37 and his colleges classify transcranial Doppler signals using individual and ensemble RNN, it archives 85.52% AUC. These works have also achieved good results in the test set. However, we speculate that human physiological features are not independent, it is not sufficient to consider only one parameter which is the reason for the better performance of LSTM that can analyze joint characteristics. 38, 40 Therefore, LSTM is a better choice for dealing with physiological parameter sequences with complex intrinsic relationships, similar to the recognition of semantic environments or voice signals.