Polysomnography is the gold standard for sleep staging, but its multisensor setup limits routine real-world monitoring. We evaluated functional near-infrared spectroscopy (fNIRS) for differentiating nonREM sleep stages using frequency-domain biomarkers, with EEG-based manual scoring as the reference framework. Eighteen healthy adults underwent simultaneous recording with a 16-channel EEG and a compact prefrontal four-channel fNIRS montage. After signal-quality screening, sixteen participants were retained, and 10 with complete N1–N3 coverage were included in repeated-measures inference. Oxyhemoglobin ([Formula: see text]HbO) and deoxyhemoglobin ([Formula: see text]HbR) signals were preprocessed with motion-artifact handling and PCA-based aggregation, then analyzed in 30-s epochs aligned to standard sleep scoring. Frequency-domain features were extracted from band-limited relative power spectra spanning slow (very low to low), respiration-related, cardiac-related, and cardiac-harmonic ranges (up to 4.9 Hz). Statistical analysis used repeated-measures ANOVA with sphericity correction and Benjamini–Hochberg FDR control. [Formula: see text]HbO showed the most reproducible stage differentiation, primarily in the cardiac and harmonic ranges, where peak-related features remained robust after correction. In contrast, [Formula: see text]HbR exhibited more selective effects that were most evident in respiration-related oscillations. Notably, N3 displayed a distinct spectral organization characterized by slow-band modulation and enriched local subpeak structure around cardiac-related activity, albeit with inter-individual heterogeneity in the slow range. Together, these results indicate that nonREM sleep depth is reflected in frequency-domain fNIRS organization across a broad physiological bandwidth: [Formula: see text]HbO shows robust modulation of cardiac and harmonic structure, while [Formula: see text]HbR contributes selectively through respiration-related oscillations. N3 further emerges as a distinct spectral regime combining slow-band modulation with enriched local subpeak structure.
Background/Objectives: This study introduces a reproducible and interpretable graph-based framework for resting-state functional near-infrared spectroscopy (fNIRS) that enables network-level biomarker discovery for Alzheimer’s disease (AD). Although resting-state fNIRS is well suited for task-free assessment, most existing approaches rely on static channel-wise features or conventional functional connectivity, limiting insight into coordinated network dynamics and reproducibility. Methods: Resting-state prefrontal fNIRS signals were represented as subject-level graphs in which edges captured coordinated fluctuations of nonlinear signal complexity across channels, computed using sliding-window analysis. Graph neural networks (GNNs) were employed as analytical tools to identify disease-stage-related network patterns. Interpretability was assessed using edge-level importance measures, and reproducibility was evaluated through fold-wise stability analysis and consensus network construction. Results: The proposed complexity–fluctuation-based graph representation consistently outperformed conventional amplitude-based functional connectivity. Statistically supported prefrontal network biomarkers distinguishing mild cognitive impairment (MCI) from healthy aging were identified, with statistically significant group differences (p = 0.001). In contrast, network patterns associated with Alzheimer’s disease were more heterogeneous and less consistently expressed. Consensus analysis revealed a subset of prefrontal connections repeatedly selected across cross-validation folds, and attention-based network patterns showed strong spatial correspondence with statistically derived biomarkers. Conclusions: This study establishes a reproducible and interpretable framework for resting-state fNIRS analysis that emphasizes coordinated complexity dynamics rather than classification accuracy. The results indicate that network-level alterations are most consistently expressed at the MCI stage, highlighting its role as a critical transitional state and supporting the potential of the proposed approach for longitudinal monitoring and clinically applicable fNIRS-based assessment of neurodegenerative disease.