This trained classifier is intended to predict the biological sex on the basis of an MRI image. It has a measured balanced accuracy of 94.01% on a validation set of N=351.
Individualized prediction of the biological sex was assessed by the use of a support vector classifier as implemented in the Scikit-learn toolbox. CAT12 whole-brain gray matter images were used as a classifier input.
To reduce dimensionality by preserving maximal localized morphometric differences, gray matter images were resliced to a voxel size of 3 mm × 3 mm × 3 mm. To strictly separate the training process from the evaluation a random validation set of 20 % (N=351, female=219, male=132) was selected, which was not used during classifier training and testing. The remaining data set of N=1402 subjects was balanced with a random undersampling procedure (N=1218, female=609, male=609) and used in a 10-fold split procedure resulting in balanced training sets of 1096 subjects in each stepfold.
Then we conducted a principal component analysis to further reduce the dimensionality of the data. The maximum number of principal components is limited to 1096, the number of subjects resulting from the 10-fold split. We carried out a Bayes statistic based hyperparameter optimization for the SVC (Scikit-Optimize) nested in the 10-fold cross-validation.
The parameter search included choice of the kernel (radial basis function (rbf) or linear), the C parameter (10^(-2) to 10^2, non-discrete log-scale), which influences penalties for misclassification, and the gamma parameter (10^(-6) to 10, non-discrete log-sale), influencing the curvature of the decision boundary.
In this iterative Bayes approach a total of 100 parameter combinations were evaluated. Quality and classifier performance are reported by area under the ROC curve (AUC).
The training of the classifier lead to two results. The first result was the estimation of a hyperparameter set, determined with the Bayes optimization method. The hyperparameter optimization estimated a rbf kernel, C=27.3 and gamma=2.4 × 10-05 for the SVM as optimal approximation for the present problem.
The second result was the classification outcome of the held out 20% validation set, which gave a performance indication for the trained classifier. The balanced accuracy for the validation set classification was 94.01%.
See file use_pretrained_classifier.py code file contained in archive