| Ankle Brachial Index (ABI)?0.9 is considered the diagnostic criteria for peripheral arterial disease(PAD). Because PAD is related to many kinds of other diseases, ABI is often used as a predictor for morbidity or mortality. But, it is still controversial for the predictive value of ABI. In this study, the predictive value of ABI for all-cause and coronary heart disease death was evaluated with Cox Proportional-Hazards Model and neural network, and the issues related to statistical analysis were explored on the modeling process. Related cohort studies were collected. Pooled hazard ration (HR) as the effect indicator and 95% confidence interval (CI) were used to estimate the association of low ankle brachial index and all-cause death risk.The research detail was as follows:1. Based on cohort study in pupulation with 2 or more risk factors of cardiovascular diseases, high level of PAD prevalence (25.8%) was observed. Flexible Cox regression model and artificial networks were used to establish predictive models for population and all-cause and coronary heart disease death were used as events of health outcomes. Relative risk and absolute risk were estimated. The results showed that ABI was an important predictor of death for this population, and Nomogram to predict objects' 1 year and 3 years absolute survival probability worthed application.2. Meta-analysis resulted in pooled HR about 1.7, when the HR of this cohort study was added, the pooled HR was about 1.6.Wiht sensitivity analysis, the results was very steady.3. The PH assumption of the traditional Cox PH model was checked with schoenfeld residual plots and hypothesis test. The linear assumption was checked with restricted cubic spline. Both nonlinear and time-dependent effects were estimated in flexible Cox model. The method was proved flexible and effective.However, when there were many variables, a large amount of work was needed.4. Least Absolute Shrinkage and Selection Operator (LASSO) was applied to establish an accurate and simplified forecasting model. The results showed that the estimated results from LASSO can be easily interpreted. LASSO should be recommended to apply.5. Multilayer perception neural network was explored to analyze survival data. Neural network had not any limitation, and can be applied flexible, and achieved easily. For large samples, the prediction performance of neural network was good. But, the results from it were difficultly interpreted. Moreover, for small samples, it was cautious to use neural network model.6. ROC curve was applied to evaluate the model performance. For smaller proportion of positive outcome events or small samples, other index such as sensitivity and specificity should be considered.To integrate the information showed above, there was high level of PAD prevalence in pupulation with 2 or more risk factors of cardiovascular diseases.Results from cohort study and Meta analysis both showed that, ABI was an important indicator for all-cause and CHD death, lower ABI can increase the risk of death. It was proposed that ABI should be routine physical examination for population who had risk factors of PAD and other cardiovascular diseases, which made early detection, intervention and treatment of PAD possible. The two assumptions of traditional Cox PH model should not be ignored. If violated, restricted cubic spline was recommended to apply to estimate the nonlinear and time-dependent effect. LASSO selected variables and estimated simultaneously, and the model was simple and its performance was good, it should be recommended. Neural network model was flexible and can be applied conveniently, it was recommended to use in large samples, but results of Neural Network was not easily to interprete. |
PDF Full Text Request