Optimizing Feature Panels for Effective Model Training Toward Early Diagnosis of Alzheimer's Disease

In this work, we aim to design effective feature selection strategies to identify relevant features to expedite machine learning model training for diagnosing Alzheimer's disease. In the first method, we introduce ranks features based on their dependency on the diagnosis by employing metrics such as Mutual Information, Symmetric Uncertainty, and Cramer's V panels via iteratively selecting top features and gradually increasing the panel size. We then propose a second method that determines the relevant features by estimating the Euclidean distance between samples and class means for each feature, employing a threshold to filter out irrelevant features. Candidate panels identified using each method are extensively tested on two datasets. Even though panels formed using the first dataset fail to meet the minimum performance criteria of 75% sensitivity and specificity, those formed using the second set achieve a significantly high accuracy up to 99.53%, 100% sensitivity, and 95% specificity. The results demonstrate the viability of our two methods, potentially paving the way for low-cost, non-invasive early detection tools for Alzheimer's disease.