Fig. 2
From: Predicting self-perceived general health status using machine learning: an external exposome study
Top 30 variables from the variable importance rankings of RF models based on the 2012 and 2016 datasets. Legend: The left panel a shows the variable importance (VI) rankings of RF models performed on the entire 2012 dataset (n = 199,840, variables n = 81) and the right panel b the ranking based on the 2016 dataset (n = 244,557, variables n = 91). The VI is expressed as the difference between the original error of the RF model and the error after randomization of the exposures of interest. The VI procedure was performed in triplicate, and the average VI score is presented. Point shape represents the type of variable: personal characteristics, neighborhood characteristics and environmental characteristic. Variables listed with an * were only available within the 2016 dataset