Created by Michael Riben
about 11 years ago
|
||
6 Main Steps to developing a predictivetools1. need to figure out timing and type of tool to develop and what role the predictive tool will play2. Diagnosis /treatment outcome determination3. Figure out what the predictors (clinical characteristics) that need to be identified and captured for the prediction tool4. Regression analysis/expert opinion --> quantification to quantify relationship between predictors and outcomes5. Test with data sets that contain the predictors and known outcomes6. Present the display to users in an applicable way Example prediction systems include DeDombal's for the diagnosis of acute abdominal painMost predictive systems are based on Bayes Theorem, Logistic regression, neural networks, and supervised pattern recognition
Development of a Prediction Rule by Regression Analysis TechniquesType of outcome determines type of regression analysis 1) For continuous outcomes, such as blood pressure, linear regression analysis is used2) For dichotomous outcomes, i.e. short term mortality, logistic regression 3) For dichotomous outcomes occurring over time long term mortality , failure-time analysis (eg. - Cox's proportional hazards regression analysisGeneral guidelines for evaluation of prediction rules:1) A clear definition of outcome and predictive variables2) Proper description of outcomes and patient population and study site to allow clinicians a comparison with their population3) A description of the mathematical technique that is used4) Availability of the accuracy or the misclassification rate of the rule 5) effects of use of the prediction tool4 Steps 1) Selection of variables, interaction terms, and classifications of inclusion in the regression model2) Estimation of regression coefficients3) Assessment of discriminative ability and goodness of fit4) Presentation table or graphical display of model results.
Variable Selection Method most frequently applied selecting a limited number of predictors --> stepwise selection --> method automatically selects variables on the basis of the amount of varianceStepwise selection may be applied in a forward a backward or combined backward-forward way--> The usual significance level for the selection of a variable in the model is 5%--> identical to significance level commonly used for the hypothesis testingWhen forward step wise is applied, variables with a p-value below 5% may be selected and the variable with the lowest P-value selected first. after one or more steps, none of the variables have P-values below 5% and the selection is completeAdvantages to stepwise selection: leads to limited number variables in the predictive model, widely available in most standard statistical computer packages. MAIN Disadvantage: Regression coefficients are overestimated = BIAS
New Page
Want to create your own Notes for free with GoConqr? Learn more.