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

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

LSTM mentioned above relying on memory cells to learn long-range dependence information. As we know, each human physiological parameter is not independent, they are interrelated, and this relationship is difficult to be found by simple coding or logistic regression algorithm. Therefore, we need a deep learning model that can learn the relationship better of far apart data to complete the task of processing text-based medical data. Simply, the LSTM neural network takes the original text-based medical data as input, and then use many special neurons to extract the joint features automatically from the original data, and finally use the classification function to classify the samples automatically to achieve the purpose of automatic diagnosis of diseases. This architecture makes it possible to process medical text data which have complex internal relationships, deep learning technology has been widely used in various fields. 15, 16 The deep learning technology-based text data classification method replaces the mathematical distance-based traditional clustering method, which greatly improves the performance of the model in processing text data. The key elements of traditional automatic diagnosis method using medical text data are: (1) patient description pathological characteristics, (2) researchers extract features manually based on patient descriptions or patient's EHR, (3) the extracted features are encoded according to the requirements, (4) classification algorithm is used to classify the coded physiological features. 17 The traditional automatic diagnosis method needs to extract features manually, and the quality of extracted feature vectors is greatly affected by the researcher's clinical experience and professional level, so it has uncertainty. At the same time, the traditional method will lose some original information artificially in the process of feature extraction, which may lose some joint features of physiological parameters, so the traditional method also has a degree of one-sidedness. Although it is not only the pathological features described by patients, EHR are also widely used in the research of automatic diagnosis. But there is no objective and unified standard to evaluate the quality of EHR. This is one of the important factors that limit the performance of automatic diagnosis algorithms using EHR. 18, 20 Deep learning algorithm has the ability of feature extraction automatically, so it overcomes the one-sidedness caused by manual feature extraction, saves labor resources and improves the efficiency and accuracy of automatic diagnosis. [20] [21] [22] Another challenge in applying deep learning algorithms to auxiliary diagnosis is the interpretability. Up to now, the deep learning model is still a black-box model, which cannot explain exactly which kind of physiological parameters plays a vital role in the process of data processing. The development of disease diagnosis technology depends not only on the improvement of diagnostic accuracy but also on the discovery of more effective diagnostic markers and the relationship between different physiological parameters (diseases). It is difficult to meet the requirements above by the auxiliary diagnosis system which only gives the diagnosis results. As we know, the human brain tends to have an attention focus when processing things, and it is able to find out important features purposefully according to the environment, this mechanism is called the attention mechanism. The co