| The significance of medical data mining is to establish a model based on the historical patient data that can predict the occurrence and development of some diseases and give reference to doctor on diagnosis. When warning signs occurred in the patient, the hospital can take appropriate preventive measures timely to prevent the exacerbation or to save the patient’s life.The model is built in this article based on real historical patient cases of a hospital in ShanDong province to predict whether patients have Deep Venous Thrombosis (DVT). It can help the hospital to implement real-time monitoring of patient condition and calculate the probability of occurrence of DVT. The hospital can use the predicted probability to narrow the range of high-risk population and take specific treatment for them.There are two aspects of meaning of this empirical analysis. First, DVT is very dangerous, and even threaten lives. The research indicates that the occurrence of DVT can be controlled effectively through prevention. Second, this work can provide evidence and experience for evaluation models of other disease. It is an exploratory work and represents a number of disease early warning models.The data included the information of the patient’s symptoms and whether or not the occurrence of DVT. And it contains eight months hospitalized pa-tients with a total of 69779 records including 108 DVT. The 65 characteristic field of the patient’s medical record was extracted as the predictor variable to predict the risk of DVT.The study belongs to the two classification data modeling problem. There are many modeling tools. Some can offer the classification results only. Some can give a scoring mechanism that provide more information and practical val-ue. The primary models used in this article are logistic regression model, BP neural network model, and support vector machine model. They can give the corresponding prediction scoring mechanism. The process of building logistic model used stepwise regression and combined with the medical conclusion to select the variables. Besides, taking into account the interaction between vari-ables may have an impact on the results, association rules algorithm is used to select and construct interaction effects as candidate prediction variables.There are two methods of selecting modeling samples. One is using the first four months data (large sample) as training samples. The other is us-ing the first four months data which removed a part of the negative patient randomly as training samples. In this way, the training samples have a high incidence. Test samples are last four months cases. Mostly, the model with large sample can predict better slightly than small samples.Viewing from the forecasting ability for test samples, these three mod-els have no big difference, and mostly, logistic model has a slight advantage. Besides, the principle of logistic model is easily to understand. And because the forecasting result is the incidence probability, its practical meaning is clear and it has better interpretation.Data construction is the key to optimize the model. Moreover, with the improvement of data, we can consider the combination of multiple models to improve the prediction accuracy. |
