Learning in data-limited multimodal scenarios: Scandent decision forests and tree-based features

Abstract

Incomplete and inconsistent datasets often pose difficulties in multimodal studies. We introduce the concept of scandent decision trees to tackle these difficulties. Scandent trees are decision trees that optimally mimic the partitioning of the data determined by another decision tree, and crucially, use only a subset of the feature set. We show how scandent trees can be used to enhance the performance of decision forests trained on a small number of multimodal samples when we have access to larger datasets with vastly incomplete feature sets. Additionally, we introduce the concept of tree-based feature transforms in the decision forest paradigm. When combined with scandent trees, the tree-based feature transforms enable us to train a classifier on a rich multimodal dataset, and use it to classify samples with only a subset of features of the training data. Using this methodology, we build a model trained on MRI and PET images of the ADNI dataset, and then test it on cases with only MRI data. We show that this is significantly more effective in staging of cognitive impairments compared to a similar decision forest model trained and tested on MRI only, or one that uses other kinds of feature transform applied to the MRI data.