Fig. 1

Prediction analysis with random forest method. A 10-tree random forest was used to predict the presence of OH; four different models were created using the following combinations of ABPM parameters: \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{ep}}^{\Delta 15/24\, {\mathrm{h}}}\) and \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{aw}}^{\Delta 15/24\, {\mathrm{h}}}\) (panel A), \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{ep}}^{\Delta 20/24\, {\mathrm{h}}}\) and \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{aw}}^{\Delta 20/24\, {\mathrm{h}}}\) (panel B), \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{ep}}^{\Delta 15/{\mathrm{DT}}}\) and \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{aw}}^{\Delta 15/{\mathrm{DT}}}\) (panel C), or \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{ep}}^{\Delta 20/{\mathrm{DT}}}\) and \({\mathrm{Hypo}}{\hbox{-}}{\mathrm{aw}}^{\Delta 20/{\mathrm{DT}}}\) (panel D), along with standard ABPM parameters. For each model, 2 × 2 tables were used to report accuracy, estimated classification by ABPM parameters, and clinical diagnosis as per in-office BP recording. The histogram represents the normalized predictive performance for each variable (blue: standard parameters; green: hypotensive parameters under investigation). Each classification model was validated through a random-labeling method (see methods section); after randomization of the outcome (presence or absence of OH), the accuracy was 64.6–67.3%; the difference between the accuracy on the randomly assigned outcome dataset and the accuracy on the real dataset (19.9–20.3%) confirmed the reliability of the classification analysis SBP: systolic blood pressure; DBP: diastolic blood pressure; MBP: mean blood pressure; w-BPV: weighted blood pressure variability; RD: reverse dipping pattern; PPH: postprandial hypotension; Hypo-ep: hypotensive episodes; Hypo-aw: awakening hypotension