A Feature Space Exploration for Improving Prediction of Demographics from Multimodal Neuroimaging Data

Abstract

The gender and age classification based on multimodal brain imaging with LEMON and HCP datasets are being explored in this thesis. The main feature engineering pipelines include feature connectivity and network graph parameter analysis in feature selection. Feature connectivity provided the main contribution to be a dominant feature which resulted in achieving the majority of accuracy ratio. More to add, the network parameters feature had added an enhancement to the accuracy in a novel way, they have been tuned in a unique way by reflecting the original meaning of graph parameters on brain concept (considering the brain regions as network nodes). While LEMON dataset presented difficulties due to small sample size and gender class imbalance, balanced data augmentation combined with PCA for dimensionality reduction and the integration of network parameters improved classification accuracy. Gender classification on the LEMON dataset has been boosted to 78.5% accuracy via SVM by using MSDL atlas and balanced augmentation. PCA also served to enhance accuracy to 84%, and 82% for SVM and FFNN on LEMON dataset as well. Concerning age classification, the accuracy was high at 92% by using SVM and MSDL atlas with the consideration of merging the network graph metrics with connectivity features. The HCP dataset outperforms with balanced classes and a large sample size by about 96% accuracy in gender classification using connectivity features and network parameters. Underlining the importance of dataset size, feature extraction, and balancing in improving predicted performance. The thesis emphasizes the effect of merging the feature connectivity with network graph parameters achieving excellent classification results. Future works may investigate possibilities of multi-modal integration, advanced methods of augmentation, and more sophisticated machine learning models for generalization and clinical application.