Research On Medical Data Mining Based On Topic Model

The explosive growth of medical data and the upsurge of data mining have aroused increasing concern about the implicit regularity in medical data. Using the data mining technology, people can extract valuable information from huge volume of medical data. This paper adopts the method of topic model to reveal the topic relationship between the patient symptoms and drugs and therefore discovers the possible syndromes and provides a reference for appropriate medicine treatment. With this method, hospitals will be provided with effective clinical paths and normative diagnosis and treatment process that may result in a scientific pattern of diagnosis and treatment.This paper firstly studies the research status of medial data mining and topic model, then analyzes medical data features, issues need to be paid attention to during medical data mining and medical data mining process, and further carries out medical data pre-processing for clinical data of a university infirmary such as attribute selection, noise processing, default value processing, etc. to acquire experimental data suitable for this research.Latent Dirichlet Allocation(LDA) topic model is applicable to text processing, which can perfectly mine the potential implications of words. Here, LDA topic model is applied in medial data mining to Put forward the relation model of “medical records- diseases-symptoms” and “medical records- diseases- drug”. Seen from the analysis of model perplexity and convergence rate, LDA topic model has a better effect in medical data mining.LDA topic model can carry out respective data mining for two categories of words in terms of “symptoms” and “drug” but not after mixing. Thus, this paper proposes a new topic model- Med-LDA topic model, and lists out Gibbs parameter deduction process. Med-LDA model, a Bayesian model with implicit parameters, has more powerful representation function than LDA model in medical data modeling. Finally experimental results show that, Med-LDA has a better effect in relationship mining of “medical records-diseases- symptoms and drug”.