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

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

Further, the above studies on the auxiliary diagnostic system show good performance in the test set, but do not provide the basis for model classification data. The development of medical diagnosis depends not only on the improvement of diagnosis accuracy but also on the researching diagnostic markers, diagnostic basis or the influence of different physiological parameters on diseases. Compared with previous work, this paper proposed a deep learning model that integrated the attention mechanism. By using the attention mechanism, we could observe which physiological parameters are more important for disease diagnosis. Model can automatically provide disease diagnostic markers while diagnosing diseases. When using only LSTM, we found that the model reached 92% ACC in the same test dataset. We suspect that this is because the use of the attentional mechanism can help the model process more efficient information purposefully, thus alleviating the problem of over-fitting. In addition, the use of attention mechanism is more convenient for the study of diagnostic markers, which can effectively reflect the importance of different physiological parameters for disease diagnosis. The performance is very close. We speculate that this is because the data is not very complex. In the future, we will study and use more types of physiological parameters to identify more complex diseases. As shown in Figure 10 , in the process of diagnosing hyperlipidemia, the model mainly judged hyperlipidemia according to cholesterol and triglyceride. This phenomenon coincides with the diagnostic criteria of hyperlipidemia. At the same time, HDL was also found to be associated with hyperlipidemia. We speculate that this phenomenon is due to the fact that HDL functions as a carrier of cholesterol in the surrounding tissues, so it has a close relationship with hyperlipidemia. 40, 41 The model mentioned above not only shows a high correlation between hyperlipidemia and direct markers, but also provide indirect markers. This phenomenon not only shows that the model proposed in this paper can learn the relationship between different physiological parameters but also shows that the model has great potential to discover new diagnostic markers. Although the model does not give new diagnostic markers using limited data, the prediction results of the model are in line with the gold standard, which proves the reliability of the model, and the model has the potential to reasonably analyze more evidence for disease diagnosis. As shown in Figure 10 , although the model pays little attention to the remaining items, in fact, the attention is not zero. We speculate that this is due to there is a correlation between human different physiological parameters. The model shows a strong concern for the physiological parameters directly related to disease, but does not give high attention to the physiological parameters not related to disease, such as red blood cells, which further proves the reliability of the model. By using the visualization method, the diagnostic basis of the auxiliary diagnostic model can be clearly presented, which improve a certain degree of transparency to the black-box model. The AI diagnosis system proposed in this paper not only provided accurate and robust diagnosis results but also provided the diagnostic basis of diseases (94% ACC, 97.48% AUC, 96% sensitivity and 92% specificity with test dataset). It not only increases the intelligence of the model but also bro