Performance of a Partial Least Squares model to predict DaTscan SBR from the Midbrain surface deformation maps. As expected, the spatial loadings corresponding to the first PLS component in the predictive model follow a very similar pattern to the observed longitudinal changes in the Midbrain surface deformations. The optimal PLS model is composed of 42 components resulting from the removal of those components not actually contributing to the prediction of the DATscan density.
A Partial Least Squares Regression (PLSR) model was trained between the DaTscan SBR measurements from the striatum (response variables) and the local deformations from the surface midbrain (predictors).
The surface deformation weights corresponding to the first PLS component follow a spatial pattern very similar to the midbrain surface longitudinal changes and encode the relative contribution of each surface vertex to the first PLS score.
A leave-one-out (LOO) cross-validation analysis produced an optimal number of 42 PLS components with a Q2 value of 0.49. A model with more than 42 PLS components might be simply reflecting training data overfitting. The removal of orthogonal components increased the Q2 value to 0.52. This relatively large Q2 value provides an objective, out-of-sample measure of our PLS model's predictive power to generate generalizable patterns.