Figure 1.

Statistical framework to compare the three harmonization strategies and select the most age-related predictive features. Three harmonization strategies were applied to each of the two dataset (i.e., NC and ASD). For each resulting dataset, a nested feature selection was performed on the training set in each round of the k-fold validation. Subsequently, 100 stepwise regression models were trained by progressive increasing the training set size. A consensus ranking procedure was used to select the most stable features with the lowest mean absolute error. This procedure was repeated for each of the three regression algorithms—SVR, RF, and Lasso—resulting in nine combinations of “harmonization strategy–regression model”. The performance of the models were compared in order to select the most effective machine learning algorithm for age prediction. For the best regression model, the sets of features that resulted from different harmonization strategies are compared using a stability index to quantify the effects of the harmonization and support clinical evaluation.