Feature Selection improves classification of Alzheimer's Disease using multimodal MRI data: Knowing more by forgetting more

Introduction Machine-learning (ML) models trained on MRI data provide promising results for Alzheimer’s Disease (AD) diagnosis. However, extracting features from an MRI scan can take up to 25 hours, due to the multitude of feature types, and different software programs used for calculating all these types. To make ML-based AD classification more accessible to doctors and researchers, this study aims to use Feature Reduction (FR) techniques to reduce the number of features needed to accurately classify AD. Therefore, the thesis question is: 'Can feature selection techniques reduce the amount of data needed to accurately classify Alzheimer’s disease from multimodal MRI scans?'. Materials and Methods For this study, 76 clinically diagnosed AD patients and 173 healthy elderly control subjects were scanned. The structural MRI scans were processed using three different programs. Cortical Thickness, Area, and Curvature data were all obtained using the Freesurfer software. Grey Matter...Show moreIntroduction Machine-learning (ML) models trained on MRI data provide promising results for Alzheimer’s Disease (AD) diagnosis. However, extracting features from an MRI scan can take up to 25 hours, due to the multitude of feature types, and different software programs used for calculating all these types. To make ML-based AD classification more accessible to doctors and researchers, this study aims to use Feature Reduction (FR) techniques to reduce the number of features needed to accurately classify AD. Therefore, the thesis question is: 'Can feature selection techniques reduce the amount of data needed to accurately classify Alzheimer’s disease from multimodal MRI scans?'. Materials and Methods For this study, 76 clinically diagnosed AD patients and 173 healthy elderly control subjects were scanned. The structural MRI scans were processed using three different programs. Cortical Thickness, Area, and Curvature data were all obtained using the Freesurfer software. Grey Matter Density was obtained using FSL VBM, and subcortical volumes were obtained using FSL FIRST. The functional MRI scans were used to calculate connectivity between resting state networks using a pipeline of different software programs. The data resulting from the four different software programs / pipeline were analysed separately. Three different classification algorithms were used: Elastic Net Regression, Gradient Boost Machines, and Random Forests. The three different algorithms individually went through Recursive Feature Reduction. Here, the models are trained and tested on the data with 5-fold cross validation and get an AUC score, after which the least important feature is removed. This is repeated until there is there is only one feature left. Results The ENR model paired with the data from Freesurfer gave the overall best scores. With 100% of data, the AUC score was only 0.899. With just 5% of data, 10 features, the AUC score improved to 0.952. The maximal AUC score was 0.981, with 18% of data. Thus, it was found that feature reduction not only preserved AUC scores, but improved them significantly. This was true in all cases, with the improvement ranging from 2%-28% when trained on only 5%-30% of features. Furthermore, when training and testing on just 5% of features, the AUC scores had a 6.6% decrease to a 20% increase. Discussion This study has found using Recursive Feature Elimination (RFE) not only preserves AUC scores, but often increases them significantly, ranging from 2% to 28%. This result exceeded the expectations of the hypothesis. Additionally, it was found that training on only 5% of features still preserved and sometimes even increased the height of AUC scores. This means that the goal of the study was exceeded, as feature reduction not only preserves, but improves accuracy scores when classifying AD. Furthermore, it was found that using an Elastic Net Regressor on the 10 outlined features of the Freesurfer Program seems most effective when classifying AD. These findings have external validity, because the areas associated with these features are also commonly linked to atrophy in AD.Show less

Leiden University Student Repository

Documents

In Collections

Feature Selection improves classification of Alzheimer's Disease using multimodal MRI data: Knowing more by forgetting more