Alzheimer's disease prediction using deep learning and XAI based interpretable feature selection from blood gene expression data.

Alzheimer's disease (AD), a type of neurodegenerative disorder, has seen an increase in cases over the past decade, necessitating the construction of a comprehensive early detection method. Existing methods are typically invasive and costly, so our research concentrates on blood gene expression as a possible biomarker for early diagnosis of AD. Many research directions using machine learning and deep learning techniques exist in the literature for AD diagnosis. However, most of them use MRI scans as the primary data, and very few studies have been carried out on the use of blood gene biomarkers. The analysis of blood gene expression data is complicated by its high dimensionality and limited sample size. In this paper, we attempt to address these issues by applying multiple feature selection methods to identify the critical genes that act as biomarkers for AD diagnosis. To select the genes linked to AD and identify AD patients, we employ four feature selection approaches, including Chi-square, ANOVA, Recursive Feature Elimination (RFE), and ElasticNet, and build two deep learning models for AD classification. The selected genes are assessed with nested five-fold cross-validation to avoid overfitting. Further, we employ SHapley Additive exPlanations (SHAP), an Explainable AI (XAI) model, for ranking the selected genes and explaining why the feature selection algorithms predict a subset of genes as probable biomarkers. Generative Adversarial Network (GAN)-based data augmentation is used to address the issue of small sample size and improve model generalization. We demonstrate the results of feature selection and AD classification on three blood gene expression datasets, namely GSE63060, GSE63061, and ADNI, and their integrated version. Experimental results indicate that the deep neural network classifier achieved an accuracy of 91% and a precision of 95% in identifying AD samples. Feature selection along with data augmentation has significantly enhanced the precision and interpretability of early detection of AD using blood gene expression.