Figure 5
Performance of different classifier types (linear support vector machine, SVM, random forests, logistic regression, linear discriminant analysis, LDA) in assigning sex in the validation data set. The classifiers were trained in a 100-fold nested cross validation setting with subsets randomly drawn from the 80% training data set separated at the beginning of the analyses. Training was performed with all variables (d = 11 pain-threshold related variables) as “full” feature set, d = 4 variables that had resulted from the feature selection steps shown in Fig. 4 as “reduced” feature set (darker blue columns Fig. 4A), or d = 1 variable that had resulted from further narrowing the feature set to the “sparse” feature set (darkest blue column Fig. 4A). In addition, the sex assignment task was repeated using permuted data for algorithm training or using the unselected features, i.e., the pain-threshold information that was not found to be informative for sex inference (light blue columns in Fig. 4A). The boxes show the 25th, 50th and 75th percentiles balanced accuracy (BA) and roc-auc for the classification performance in the 20% validation data separated before feature selection and classifier training from the whole data set and not used for feature selection and algorithm training. Whiskers span the 95% confidence interval from the 2.5th to the 97.5th percentiles. The vertical dashed red lines mark the 50% guessing level that must not be touched by the confidence interval of the classification performance measures if the algorithm can be considered successfully trained. The vertical dotted blue lines mark the respective best median classification performance observed across all data set variants and algorithms. The figure has been created using the software package R (version 4.2.2 for Linux; https://CRAN.R-project.org/6) and the R libraries "ggplot2" (https://cran.r-project.org/package=ggplot27) and "ComplexHeatmap" (https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.html58.
