Journal of Biometrics & Biostatistics

Developing Prediction Models from Results of Regression Analysis: Woodpecker^™ Technique

Abstract

Alexander Goldfarb Rumyantzev, Ning Dong, Sergei Krikov, Olga Efimova, Lev Barenbaum and Shiva Gautam

Background: Developing medical prediction models remains time and labor consuming. We propose the approach where information collected from published epidemiological outcome studies is quickly converted into prediction models. Methods: We used general expressions for regression models to derive prediction formulae defining the probability of the outcome and relative risk indicator. Risk indicator (R) is calculated as a linear combination of predictors multiplied by regression coefficients and then is placed on the scale of 0 to 10 for interpretability. Prediction expression for the probability (P) of the outcome is derived from general expression for logistic regression and proportional hazard models. The intercept is calculated based upon the outcome rate in the population and the risk indicator assigned to a subject representing mean characteristics of the population (Ã?Ë?Ã?â??). We also consider linear expression where probability of outcome is the product of risk indicator and the ratio of observed outcome rate and Ã?Ë?Ã?â??. Results: These models were explored and compared in a numeric simulation exercise and also using real data obtained from NHANES dataset. All three expressions generate very similar predictions in the lower categories of risk indicator. In the groups with the higher value of risk indicator linear expression tends to predict lower probability than exponential expressions and also lower than observed. Conclusions: We demonstrated simple technique (named Woodpecker™) that might allow deriving functional prediction model and risk stratification tool from the report of clinical outcome study using multivariate regression model.