Distributional Transformation Improves Decoding Accuracy When Predicting Chronological Age From Structural MRI

When predicting a certain subject-level variable (e.g., age in years) from measured biological data (e.g., structural MRI scans), the decoding algorithm does not always preserve the distribution of the variable to predict. In such a situation, distributional transformation (DT), i.e., mapping the predicted values to the variable's distribution in the training data, might improve decoding accuracy. Here, we tested the potential of DT within the 2019 Predictive Analytics Competition (PAC) which aimed at predicting chronological age of adult human subjects from structural MRI data. In a low-dimensional setting, i.e., with less features than observations, we applied multiple linear regression, support vector regression and deep neural networks for out-of-sample prediction of subject age. We found that (i) when the number of features is low, no method outperforms linear regression; and (ii) except when using deep regression, distributional transformation increases decoding performance, reducing the mean absolute error (MAE) by about half a year. We conclude that DT can be advantageous when predicting variables that are non-controlled, but have an underlying distribution in healthy or diseased populations.