Analysis And Forecast Of Medical Data Via Nonlinear Dynamics

There are many organs of human being belong to nonlinear systems,for example,the brain.Chaos theory is a key theory in nonlinear kinetics.This thesis focus on the forecast of medical data with chaotic characteristic(the chaotic time series).Chaotic time series with the characteristics of irregular,unstable,long-term unpredictable but short-term predictable are widely found in business,agriculture,meteorology,biological sciences,ecology,and many other fields.The discovery of useful knowledge from chaotic time series has become one of the hotspots in data mining.In this work,we used electroencephalogram as the sample to forecast the seizure.The main contents of this thesis are as follows:First,the traditional model has some common shortcomings,one of them is that it could not work well when deal with cross-patient's prediction.Based on traditional method,our first model(Model 1)improves this situation by introducing transfer learning.Meanwhile,we used RQA and SMO algorithms in relevant steps for better result.Finally,the AUC of Model 1 is about 0.799,much higher than the AUC of the traditional model(0.648).Second,the other defect of traditional method to forecast electroencephalogram is that it could not extract high-level characteristic.Our second model(Model 2)addresses this issue by introducing convolutional neural network structure and using special preprocessing operations.Finally,the AUC value of Model 2 is about 0.810,the performance is far better than the traditional model.Third,considering the timing properties of time series,we combined convolutional neural network and bi-directional recurrent neural network to propose a new model,BRCNN,which shows a remarkable forecast effect.Its AUC value can reach about 0.904,the performance is obviously superior to the first two models.Finally,we forecast the medical data(electrocardiogram)via BRCNN to prove the universality of this new model on the forecast of chaotic time series,results demonstrated that this model is also suitable for electrocardiogram prediction.This work is significative,and it also provides a great convenience for other researchs on related field.