| Aged health care population has made remarkable contributions to the Chinese Communist Party and national reconstruction. Prolonging the life span and improving the quality of the aged health care population and providing a systematic, solid and active health care model for them is responsibility of the health maintenance organizations using advanced medical technology, information technology and management strategy in modern society.With the rapid development of social economy and increasing improvement in the people’s life quality, ischemic cardiovascular disease (ICVD) has become the first killer. At present,the diagnosis and treatment of ICVD most depend on clinical experience, including too much subjective judgment and lack of precise forecasting for the incidence rate of disease and necessary quantization . In order to make up for the deficiency of human decision-making, kinds of prediction models appear Recently, most models are used to predict the risk rate of the middle aged people usually about 10 years without the background of cardiovascular diseases. However, the expected models of the aged without disease background are limited, especially for the ischemic cerebrovascular disease.The aged health care population holds high expectations to the medical health care. However, due to their old age, various elementary illness factors and longtime exposure to the dangerous factors, every factor is in complicated relationship with each other. So the establishment of the illness-risk-calculating model for the aged should take the integrity, complexity, dynamic characteristics of the illness and the cooperative relations among these factors into account. The man-made nerve approach is characterized with its comprehensive, systematic, cooperative, self-taught, self-organized properties and the extreme capacity to the mistakes. Moreover, it has made extraordinary achievement in such fields as recognizing dangerous factors, processing information and aiding decision duing to its peculiar structures and information-processing methods. It is one of the main methods to process the complicated non-linear problems. Using the classification and calculation functions of the research of the man-made nerve net, the research develops the illness-risk-calculating model for the aged health care population, thus to solve the key problems of the prior-warning of ischemic cerebrocardiac disease for the aged health care population By doing so, we hold the expectation that we will give a reasonable safeguard scheme for the aged health care population and the corresponding aged health management.Objective and significance:1. To study the risk factors and descriptive characteristics of ICVD for the aged health care population .2. Using Back Propagation(BP) artificial neural network to establish the prediction models of ICVD for the aged health care population.3. To establish Cox’s proportional hazard (PH) regression model of ICVD for the aged health care population .4. To establish the prediction models of ICVD for high-ranking military based on BP artificial neural network. comparing the predictive value with the actual situation, testing model’s discrimination ability, prediction accuracy, the predictive power of group level.The early prediction model of ICVD for the aged are rare at home and abroad, the aim of this study was to found the prediction models based on BP artificial neural network with these risk factors as independent variable, utilizing and sorting the characteristic medical record about cadre health care population in our country, construction a scientific and utility information databases, focused on the study of the early warning model in the common serious disease with health care population, and found the model which is better to healthy accord with the characteristic of this population, exploration and development of a series of supplementary procedures standardization of treatment, comprehensive prevention and control measures and scientific management methods, which can provide theoretical basis and operational work guide for clinical evidence-based care and standard medical care, meanwhile, to develop the macro-strategy for health care information technology support and protection for decision-making bodies.Methods:Baseline population is all the total aged health care population that was born before 01.01.1938 (>65 years) that was recorded in the database of the hospital.. All the data were from physical examination data at 05, 2003, hospitalization data of past years, questionnaire and telephone return visit .The deadline of follow-up is 10.2009, and the follow-up period is 6years and 4 months. Data collection through combination of electronic medical record and traditional paper medical record, objective index measurement and questionnaire, field survey and telephone return visit, with emphasis on quality control of physical examination and questionnaire data.According to the latest development of researches in home and overseas, risk factors involved in the model of this study: Baseline age,body mass index (BMI),systolic blood pressure (SBP),serum total cholesterol (TC),serum triglyceride levels (TG),serum high density lipoprotein levels (HDL-C),serum creatinine levels (Scr), plasma apolipoproteinα1(ApoAⅠ), diabetic mellitus, smoking.In order to reduce bias, female and male patients underwent ICVD at the baseline time were removed at the establishment of the model. Randomly selected baseline population with a proportion of 4:1, generates module group and verification group. Using BP neural network model and COX PH regression model respectively to fit the optimal model. To verify distinguished ability with area under ROC curve; to verify the predictive veracity through comparing the predictive incidence rate and actual incidence rate of every deciles group by Hosmer-Lemeshow test. To verify the predictive veracity of the prediction model at population level through comparing predictive 6 years incidence rate of ICVD with actual 6 years accumulative incidence rate of ICVD and calculating error rate.Epidata 3.1 was used to input the data, Stata 9/SE was used to clean data,analyze data, draw chart and establish COX PH regression model, Matlab 7.0 was used to establish BP neural network model.Result:1. The baseline population of this study is 2271 cases of old male, and The accumulated personnel year is 12852.8 in the observation. Of that, 523 cases were hospitalized for ICVD (23.03%), 1499 hospitalized for other reasons (66.01%) and 249 without hospitalization experiences (10.96%). 81 died for ICVD (3.57%), and 370 died for other reasons (16.29%). The disease incidence of personnel year of ICVD was 41.63/1000 personnel years and the cumulative life-incidence was 23.56%. The cumulative death rate was 3.57% and death rate of personnel year was 6.30/1000 personnel years. According to the system invaded, the highest rate was diseases of digestive system followed by ICVD, ischemic cerebrovascular diseases and circulatory diseases (ICVD is excluded) one by one. Of all the cases, 34.91% was the incidence of cardiovascular diseases. Of the cases with cardiovascular diseases, 69.86% was ICVD with a relatively long observation. Cerebral infarction incidence was first, followed by myocardial infarction closely. However, the main death reason was disease of respiratory system and disease of digestive system with the shortest observation. In the whole, the reasons and its proportion of death cases were separately cerebral infarction (35 cases, 7.76%), and myocardial infarction (32 cases 7.10%), upper respiratory infection (17 cases 3.77%), and chronic bronchitis (17 cases 3.77%).2. By life table method, we conclude the incidence of ICVD rises year by year, with the survival rate decreasing meanwhile. The invalid rate reaches the top in the 5th year and 6th year with high incidence of ICVD.3. By mono-factor analysis of risk factors and ICVD, we conclude variables which are positively correlated with ICVD are SBP, BMI, TG, TC, ApoAⅠ, diabetes, smoking, and variables which are negatively correlated with ICVD are HDL-C、Scr.4. Construction of BP neural network model Training data was at random divided into training data (1400 persons) and efficacy data (417 persons). By Matlab7.0, we realized random separation, variable normalization, network initialization, network training, network simulation, output function design in training data. Neuronic count of input layer was 10 as the same as input variable. The count of hidden layer was 1. We tried to construct predictive model with hidden units from 5 to 15, taking ROC area under curve as the index which was used to screen the optimal model. After training repeatedly, we conclude that when the hidden unit is 8, network speed is fastest, with the smallest vibration amplitude and the largest index of ROC area under curve. Therefore, the network can reach the expected training error very fast. However, after that, the network decreases gradually, indicating the decrease of net discriminating power. When the hidden unit is 12, the network begins to increase slowly again. The difference of ROC area under curve between training data and efficacy will reach the smallest index when the hidden node is 8. Therefore, when the hidden count is 8, we can obtain the optimal model of BP neural network model 1 output layer and 3 output variables was designed: the number 1 meant that disease occurred within 6 years, and the number 2 meant no incidence within 6 years; from baseline to eclipse period of ICVD; cumulative life-incidence of ICVD within 6 years. This network can obtain good simulation effect. When the training reached 7401 steps, descending gradient became 0, which could meet the requirement. The training error was 0.0923879. When hidden unit was 8, the effect of the network was best, with the index of R reaching 0.914.5. Construction of COX PH regression models. The factor of age was divided into advanced age group (≥75) and old age group (<75). Except the factor of age, Proportional Hazards(PH)analysis was done for natural logarithm of other continuous variables that did not obey to Gaussian distribution. Independent variables in every layer must obey to presumption. As a result, by the introduction of all the risk factors according to age intervals as layer standard, COX PH regression models was constructed finally. The result of old age group revealed the risk factors of statistical significance were age, ln (SBP)、ln (Scr)、ln (Blood-fasting sugar) and the conservative factor was ln (HDL-C); the result of advanced age group revealed the risk factors of statistical significance were ln (BMI)、ln (SBP)、ln (TC)、ln (Scr)、ln (Blood-fasting sugar) and the conservative factor was ln (HDL-C). According to the results of mono-factor analysis and multiple factors analysis, we rejected the independent variables without statistical significance and then constructed COX PH regression models. Finally, the risk factors induced into the COX PH regression model in the old age group were ln (BMI)、ln (SBP)、ln (TC)、ln (HDL-C)、ln (Scr)、ln (Blood-fasting sugar); the risk factors in the advanced age group were ln (BMI)、ln (SBP)、ln (TC)、ln (HDL-C)、ln (Scr)、ln (Blood-fasting sugar). Baseline risk possibility for 6 years was 0.254 in the old age group and was 0.328 in the advanced age group.6. Verification and comparison of the model. Baseline data of verification group was put into the optimal BP neural network model and COX PH regression model respectively to anticipate the disease incidence of ICVD for 6 years. According to the contrast of the anticipated index and actual index, ROC curve was constructed. ROC area under curve (AUC) of BP neural network model was 0.892 (0.870~0.914) (95% CI), and ROC AUC of COX PH regression model was 0.723 (0.687~0.759) (95% CI). Thus, the discrimination ability of BP neural network model is considered to better than that of COX PH regression model. Individual cumulative incidence rate of ICVD anticipated was sequenced from the lowest point to the top and then was divided into groups in decile. Mean value of anticipated incidence rate in every group was compared with actual cumulative incidence rate. Hosmer-Lemshow test was used respectively in BP neural network model (χ2=0.82,P =0.896) and COX PH regression model (χ2=1.43,P =0.786). Except the anticipated incidence rate of 10th group, other groups had slightly lower anticipated incidence rate than the actual index. However, generally speaking, both models were relatively accurate. The actual cumulative incidence rate was 26.43%. In contrast, the anticipated mean incidence rate of COX PH regression model was 25.84% with an error rate of﹣2.23%, and that of BP neural network model was 26.42% with a tiny error rate of﹣0.04%. Therefore, the anticipation ability of BP neural network model at population level is better than that of COX PH regression model.Conclusions:1. ICVD become the most important chronic diseases in the aged, and in the aged health care population the incidence of ICVD is higher, shorter latency, longer duration and heavier burden. Therefore, prevention of ICVD is one of the most important works of the aged health care population.2. SBP, fasting blood glucose, TC are major risk factors of ICVD events, HDL-C is the most important protective factor, therefore, the necessary measures to prevent future ICVD are control of blood pressure and blood sugar, total cholesterol levels, increase high density cholesterol level .3. This study applyed BP neural network to forecast the 6-year incidence of ICVD in the aged health care population. Compared with Cox’s PH regression model, artificial neural network has the advantages of highlights the holistic, systematic, nonlinear, parallel, self-learning, self-organization and strong fault tolerance, and the discriminate ability, prediction accuracy, the predictive power of groups to be superior to the level of Cox’s proportional hazard regression model, but also less demanding on the raw data, and the process is simple, so it have better ability to spread and apply. |
PDF Full Text Request