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

Area of Science:

• Machine Learning and Artificial Intelligence
• Medical Imaging Analysis
• Computational Biology

Background:

• Multimodal studies often face challenges due to incomplete and inconsistent datasets.
• Existing decision tree methods struggle with data heterogeneity and missing features.

Purpose of the Study:

• To introduce scandent decision trees for handling incomplete multimodal data.
• To enhance decision forest performance using limited multimodal samples and large incomplete datasets.
• To develop tree-based feature transforms for classifying samples with partial feature sets.

Main Methods:

• Developed scandent decision trees that mimic data partitioning using feature subsets.
• Integrated scandent trees with tree-based feature transforms within a decision forest framework.
• Trained a model on multimodal MRI and PET data from the ADNI dataset.

Main Results:

• The proposed methodology significantly improved cognitive impairment staging compared to MRI-only models.
• Testing on cases with only MRI data demonstrated the effectiveness of the multimodal approach.
• Scandent trees and tree-based feature transforms outperformed other feature transform methods.

Conclusions:

• Scandent decision trees offer a robust solution for multimodal data analysis with missing features.
• The combined approach enables effective classification using limited multimodal data.
• This method shows promise for improving diagnostic accuracy in neurological disorders using medical imaging.