Open Collections

Scandent tree : a decision forest based classification method for multimodal incomplete datasets

University of British Columbia

Incomplete and inconsistent datasets often pose difficulties in multimodal studies. A common scenario in such studies is where many of the samples are non-randomly missing a large portion of the most discriminative features. We introduce the novel concept of scandent decision trees to tackle this issue in the context of a decision forest classifier. 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 use the forest resulting from ensembling these trees as a classification model. We test the proposed method on a real world example of the target scenario, a prostate cancer dataset with MRI and gene expression modalities. The dataset is imbalanced with many MRI only samples and few with MRI and gene expression. Using scandent trees, we train a classifier that benefits from the large number of MRI samples at training time, and of the presence of MRI and gene expression features at the time of testing. The results show that the diagnostic value of the proposed model in terms of detecting prostate cancer is improved compared to traditional methods of imputation and missing data removal. The second major contribution of this work is the concept of tree-based feature maps in the decision forest paradigm. The tree-based feature maps 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. This has important clinical implications: one can benefit from an advanced modality to train a classifier, but use it in a practical situation when less expensive modalities are available. We use the proposed methodology to build a model trained on MRI and PET images of the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, and then test it on cases with only MRI data. We show that our method is significantly more effective in staging of cognitive impairments compared to a model trained and tested on MRI only, or one that uses other kinds of feature transform applied to the MRI data.

Text

2016-04-26

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/57848

Biomedical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/