Learning Alzheimer’s disease signatures by bridging EEG with spiking neural networks and biophysical simulations

As the prevalence of Alzheimer’s disease (AD) increases, there is a growing need for interpretable biomarkers derived from non-invasive neural recordings. While deep learning approaches applied to electroencephalography (EEG) achieve strong classification performance, they often lack mechanistic interpretability and require large datasets. Spiking neural networks (SNNs) provide a biologically grounded alternative, yet their relationship to large-scale brain dynamics remains insufficiently understood, and their potential for low-power deployment on neuromorphic hardware remains largely unquantified in this domain. We propose a conceptual framework in which data-driven models and mechanistic simulations are linked through convergence on a common physiological axis, namely the excitation–inhibition (E/I) balance. Within this framework, EEG-derived functional connectivity, estimated using network-based statistics, is used as a connectivity prior to constrain large-scale spiking network simulations, enabling empirically informed circuit dynamics. In parallel, an SNN classifier trained on resting-state EEG achieves competitive performance (AUC = 0.804–0.844), with the aperiodic 1/f slope emerging as a discriminative marker corresponding to E/I imbalance, creating a further conceptual bridge. Deployed on SpiNNaker neuromorphic hardware, the latency-encoded SNN retains competitive classification performance (AUC = 0.793–0.839) while reducing estimated energy consumption per inference to approximately one third of a GPU-based baseline on SpiNNaker 2, demonstrating the practical viability of low-power neuromorphic deployment for this task. To mechanistically interpret the EEG findings, we simulate spiking networks with systematically varied inhibitory-to-excitatory ratios under both unconstrained and connectivity-constrained conditions. The resulting dynamics reproduce key spectral features of empirical EEG, including spectral slowing and alpha-band modulation, with functional connectivity priors improving alignment between simulated and observed spectral structure. Overall, by grounding both machine learning and biophysical simulations in a shared E/I balance framework, constraining circuit dynamics with empirical connectivity, and validating low-power deployment on neuromorphic hardware, this work provides a coherent, interpretable, and practically deployable account of EEG biomarkers in Alzheimer’s disease.