This research is aimed at investigating if machine learning can be used to diagnose Alzheimer’s disease (AD) via fusion of extremely different types of data from the ADNI database: trajectories of blood biomarkers and PET brain scans.
The pipeline for blood biomarkers standardizes device family dependent measures, performs missing values imputation with KNN, and uses PCA to reduce the feature space before classification with an ensemble of classifiers (SVM, Naive Bayes, logistic regression, random forest, gradient boosting). Using this approach individually, we got an accuracy of 57%, constrained primarily by strong overlap between MCI and AD/CN classes. At the same time, PET scans (FDG, Amyloid, and Tau tracers) were registered, skull-stripped, smoothed and normalized and were classified individually with a VGG19 architecture trained from scratch with accuracy of 72.3%, again being constrained by difficulty of classification of MCI class.
Fusion of the odds provided a huge performance boost in comparison with individual approaches, resulting in 93.2% of overall accuracy, demonstrating that the information provided by imaging and blood-based approaches are indeed complementary.
Project done with Mikaela Buzdin for the Machine Learning module (CML3), under the supervision of Prof. Dr. Arzu Çöltekin and others.