Characterizing Alzheimer’s disease with reservoir computing

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by impairments in memory, cognition, and behavior. Resting-state functional magnetic resonance imaging (rs-fMRI) is one of the valuable techniques for studying brain function in AD, providing insights into neural mechanisms and aiding early diagnosis. Traditional rs-fMRI analyses, such as dynamic functional connectivity and time series analysis, aim to explore spatiotemporal relationships between brain regions. However, the short duration of real-world rs-fMRI data often limits their ability to capture nonlinear features that are critical for early-stage AD detection. To overcome these obstacles, this study develops a surrogate model based on the techniques of reservoir computing and compressed sensing (CS-RC), which expands data through time series predictions and extracts nonlinear dynamical characteristics of AD. Two indicators are further proposed based on the developed model, the maximum Lyapunov exponent and the phase locking values, indicating a reduced dynamical complexity in AD and the key AD-affected brain regions in the frontal and parietal lobes, respectively. The classification accuracy based on these two indicators can reach as high as 87% across different datasets, validating the CS-RC analysis approach. Our framework improves the use of rs-fMRI data and may support further theoretical investigations in neurodegenerative disease research.