Research And Application Of Neighborhood Information HMM On Predicting Individual Disease

Disease prevention is still important in contemporary. However, on the one hand, modern prediction method of TCM is still a manual prediction method based on individual experience, so the credibility and accuracy are very low; on the other hand, the problem of predicting individual disease featured with high complexity, few samples and multi-priori, so the traditional machine learning model cannot obtain a high accuracy as well. Therefore, this dissertation attempts to combine the prediction method of machine learning and the priori knowledge of TCM to design a model for predicting individual disease.The priori knowledge of TCM holds that the main reason for disease is that the internal state of an individual does not suited to the change of the external state. The external state can be reflected by local weather and the internal state can be reflected by individual meridian. In order to capture the change regularity of weather, this dissertation focuses on the study of sequence classification method based on Hidden Markov Model (HMM).There exists shortcomings on the speed and accuracy in the existing sequence classification methods based on HMM. Therefore, based on the conjecture of neighborhood’s similarity implies sequence’s similarity, this dissertation proposes a sequence classification methodology based on neighborhood information HMM. Firstly, the constrained HMM space defined by sample is transformed to the unconstrained HMM space. Next, the neighborhood information is extracted at the standard HMM, and then imported to the SVM. On the basis of the proposed methodology, two kinds of neighborhood information are proposed as well, which are derivative neighborhood information and frequency-domain neighborhood information. Specifically, the traditional methods to extract derivative neighborhood information often adopt approximation strategy due to the existence of mutual restraint parameters in the discrete HMM, which can only achieve poor accuracy and speed. To solve this problem, a derivation method based on the unconstrained HMM space is proposed. Experimental results show that compared with other existing sequence classification methods, the proposed methodology can indeed greatly improve the speed and accuracy. Meanwhile, the results validate the correctness of the original conjecture and the effectiveness of these two proposed neighborhood informations. In addition, it is easy to generalize and ensemble the proposed methodology.Takes the statistical analysis result of individual disease as guiding principle, and applies the proposed neighborhood information HMM as modeling tool, this dissertation designs a model for predicting individual disease which takes individual meridian vector and historical weather sequence as input, takes disease category number as output. Experimental results show that the designed model exhibits excellent predictive performance. Meanwhile, the results also validate the existence of the correlation between meridian, weather and disease.