Study On The Predictive Model Of Malignant Tumor Progression Based On The Individual Symptom Characteristic Sequence

Statistics from World Health Organization (WHO) show that lung cancer is amalignant tumor with the highest incidence and mortality rates worldwide. Over onethird of deaths caused by lung cancer around the world is from China. Nowadays,more and more terminal cancer patients are facing the overtreatment problem.Moreover, the reality of blind doctor seeking not only increases the difficulty fordoctors to diagnose patient’s conditionscompressively, but also delays the timelytreatment recommendations for disease progression. Recently, with the continuousdevelopment of electronic medical records, the study on patients’ tumor progressionmodeling has become a new research topic for the majority of health researchers.From the perspective of system engineering, interventions will first impact theeasily observed macro performance indicators, which will reflect the progress oftumor growth. Based on experiencesfrom clinical doctors, the characteristic pattern ofindividual patient’s symptom changes can reflect the disease progression to someextent. Therefore, this paper focus on the modeling and analysis of advancednon-small lung cancer patients. And the clinical testing indicators are compared as thestate observer vectors in the control system. Making the analysis of cancerprogression as a start point, this paper models the control effects of interventions totumor’s progression as a dynamic relation model between all previous clinical testingseries to the tumor progression. This paper also proposes the tumor progressionprediction model, which is based on the characteristic sequence ofindividualsymptoms and is independent of interventions.Therefore, main research contents of this paper can be divided into four parts.First, the incompleteness problem of clinical data is inevitable in practicalclinical data collection.The reasonable and effective fill-up of incompleteness clinicaldata is the basis for problem modeling.Based on the analysis of the characteristics ofclinical data and input variables, this paper proposes different fill-up methods forincompleteness clinical data preprocessing, such as mean fill-up method, single valuefill-up method and Last Observation Carried Forward method, by considering both theadvantages and limitations of existing fill-up methods.This paper utilizes the proposedexpectation maximization algorithm based multiple values incompleteness clinicaldata fill-up method to fill up clinical data with different incompleteness ratios. Resultsprove that the proposed multiple values fill-up method is the most efficient methodwhen the incompleteness ratio of clinical data is from30%to50%. By using theproposed clinical data fill-up methods, the complete clinical data set is achievedwhich is the basis for researches below in this paper. Second, in order to unveil the mapping relation between symptomssequencepattern of the individual patient and the disease progression, this paper proposes amethod of combining both Traditional Chinese Medicine (TCM) and westernmedicine treatments. The proposed method utilizes theautocorrelation featureofcross-sectional and time series combination which is hidden in the longitudinal data,and establishesa prediction model using regression analysis method. Moreover, thispaper proposes a new coding method for patients’ symptoms changing mode by usingthe classification soft indicator. The logistic regression prediction model is alsoestablished by using the symptoms variation values of the individual patient as inputand the disease progression as output. Results show that the prediction accuracy ofdisease progression for terminal lung cancer patients achieves90.7%. And botheffectiveness and accuracy of the proposed model are verified by using ROC curves inthis paper.Third, in order to explore the symptom variation features of individualpatient’sclinical data, this paper combines the discrete clinical data sequence into acountable data model and explores the mapping relation between the countablemodeland disease progression or survival status of patients. In this paper, thehierarchical clustering and classification clustering methods are applied to cluster andanalyze the time series clinical data of TCM soft indicator. Furthermore, theadvantages and limitations of both methods are compared from theoretical andpractical perspectives. Finally, the countable data model is achieved to presentsymptom variation features.Fourth, to build the prediction model of malignant tumors based on thesymptomscharacteristic ofthe individual patient without interventions, this paper proposes anovel feedforward artificial neural network (ANN) based prediction model.Optimization approaches are designed in the network topology and training sequencestrategy to overcome the limitations of existing ANN. Moreover, this paper proposesthe correlation coefficient method for variable selection to improve the learningefficiency of ANN model and introduces the momentum factor to improve theconvergence rate. Based on10-fold cross-validation method, results show that theaccuracy of disease progression prediction of terminal lung cancer patients achieves88.39%, verifying the effectiveness and applicability of proposed ANN basedprediction model.