Predicting Alzheimer’s Disease Using Deep Neuro-Functional Networks with Resting-State fMRI

• Published in: Electronics (Basel) 2023-02, Vol.12 (4), p.1031
• Main Authors: Sethuraman, Sambath Kumar, Malaiyappan, Nandhini, Ramalingam, Rajakumar, Basheer, Shakila, Rashid, Mamoon, Ahmad, Nazir
• Format: Article
• Language: English
• Subjects: Accuracy; Advertising executives; Age; Algorithms; Alzheimer's disease; Automation; Brain diseases; Brain research; Classification; Cognitive ability; Computer-aided medical diagnosis; Customization; Deep learning; Diagnosis; Frequencies; Machine learning; Magnetic resonance imaging; Medical research; Medicine, Experimental; Neural networks; Parameter modification; Time series
• ISSN: 2079-9292; 2079-9292
• DOI: 10.3390/electronics12041031

Resting-state functional connectivity has been widely used for the past few years to forecast Alzheimer’s disease (AD). However, the conventional correlation calculation does not consider different frequency band features that may hold the brain atrophies’ original functional connectivity relationships. Previous works focuses on low-order neurodynamics and precisely manipulates the mono-band frequency span of resting-state functional magnetic imaging (rs-fMRI). They specifically use the mono-band frequency span of rs-fMRI, leaving out the high-order neurodynamics. By creating a high-order neuro-dynamic functional network employing several levels of rs-fMRI time-series data, such as slow4, slow5, and full-band ranges of (0.027 to 0.08 Hz), (0.01 to 0.027 Hz), and (0.01 to 0.08 Hz), we suggest an automated AD diagnosis system to address these challenges. It combines multiple customized deep learning models to provide unbiased evaluation, and a tenfold cross-validation is observed We have determined that to differentiate AD disorders from NC, the entire band ranges and slow4 and slow5, referred to as higher and lower frequency band approaches, are applied. The first method uses the SVM and KNN to deal with AD diseases. The second method uses the customized Alexnet and Inception blocks with rs-fMRI datasets from the ADNI organizations. We also tested the other machine learning and deep learning approaches by modifying various parameters and attained good accuracy levels. Our proposed model achieves good performance using three bands without any external feature selection. The results show that our system performance of accuracy (96.61%)/AUC (0.9663) is achieved in differentiating the AD subjects from normal controls. Furthermore, the good accuracies in classifying multiple stages of AD show the potentiality of our method for the clinical value of AD prediction.