Bridging Neural and Hemodynamic Signals through EEG-bbNIRS Fusion for Early Alzheimer's Disease Classification

Early detection of Alzheimer's disease (AD) is crucial for therapeutic efficacy, yet noninvasive screening remains challenging. Electroencephalography (EEG) captures fast neural dynamics, whereas broadband nearinfrared spectroscopy (bbNIRS) provides hemodynamic and mitochondrial metabolic information, but each modality alone offers only a partial view of neurovascular-neurometabolic dysfunction. To address this, we propose a compact multimodal deep learning framework that explicitly models the coupling between electrophysiological activity and hemo-metabolic oscillations. The architecture comprises modality-specific encoders for EEG spectral features and bbNIRS-derived hemodynamic-metabolic signatures, whose embeddings are fused in a shared latent space regularized by a correlation-based consistency term. A focal loss is further employed to mitigate class imbalance and emphasize hard-to-classify samples in the smallsample regime. We evaluate the framework on resting-state EEG-bbNIRS recordings from healthy controls and AD patients under two protocols (combined eyes-open/eyesclosed vs. eyes-closed only). Empirically, the proposed fusion model consistently outperforms unimodal EEG or bbNIRS baselines and naïve feature concatenation, yielding improved and more stable classification performance across recording settings. These results demonstrate that jointly leveraging electrical and hemo-metabolic cues from EEG and bbNIRS can enhance noninvasive AD detection and suggest a promising direction for scalable, physiologygrounded early screening tools.