Authors - Iroshan Aberathne, Don Kulasiri, Sandhya Samarasinghe Abstract - Manual techniques are not yet capable of tracking such tiny abnormalities in the early stages of AD due to the inherent complex nature of the structural changes in the brain. Therefore, we propose a novel model agnostic unsupervised feature selection algorithm called Prominent Feature Selection (PFS) to identify significant or vulnerable brain regions via volumetric measurements extracted from MRI images. Features selected by PFS are evaluated in two aspects. First, we compare the results with several other feature selection methods to assess the interpretability of the selected features against age and gender. The experimental results show that the PFS results better describe the structural changes in brain regions against age and gender. Secondly, binary class classification is performed to observe the discriminative power of the selected features against Cognitive Normal (CN) vs. Early Mild Cognitive Impairment (EMCI) subjects. Binary class classification performance measures demonstrate the significance of PFS. It has higher F1 score values despite a lower number of features compared to other feature selection methods. Since, PFS can identify and rank the most significant brain regions at early the stage of AD, diagnosis may become more specific. PFS can also be used as an effective feature selection technique not only to enhance the model performance but also to extend the model’s interpretability which would improve personalized evaluation in areas like eXplainable Artificial Intelligence (XAI).
