Artificial Intelligence Predictor for Alzheimer’s Disease Trained on Blood Transcriptome: The Role of Oxidative Stress

Abstract

Alzheimer’s disease (AD) is an incurable neurodegenerative disease diagnosed by clinicians through healthcare records and neuroimaging techniques. These methods lack sensitivity and specificity, so new antemortem non-invasive strategies to diagnose AD are needed. Herein, we designed a machine learning predictor based on transcriptomic data obtained from the blood of AD patients and individuals without dementia (non-AD) through an 8 × 60 K microarray. The dataset was used to train different models with different hyperparameters. The support vector machines method allowed us to reach a Receiver Operating Characteristic score of 93% and an accuracy of 89%. High score levels were also achieved by the neural network and logistic regression methods. Furthermore, the Gene Ontology enrichment analysis of the features selected to train the model along with the genes differentially expressed between the non-AD and AD transcriptomic profiles shows the “mitochondrial translation” biological process to be the most interesting. In addition, inspection of the KEGG pathways suggests that the accumulation of β-amyloid triggers electron transport chain impairment, enhancement of reactive oxygen species and endoplasmic reticulum stress. Taken together, all these elements suggest that the oxidative stress induced by β-amyloid is a key feature trained by the model for the prediction of AD with high accuracy.

Keywords

• Alzheimer’s disease
• data mining
• machine learning
• support vector machines
• neural network
• logistic regression