Prefrontal Haemodynamics in Type 2 Diabetes with Cognitive Impairment Versus Healthy Controls – Modal Phenotyping During Cognitive Task

Background: Functional near-infrared spectroscopy (fNIRS) during cognitive tasks contains slow haemodynamic oscillations from neurovascular, superficial systemic and cardiorespiratory sources. We investigated whether output-only modal analysis can provide dynamic systems descriptors during a cognitive task in type 2 diabetes with cognitive impairment versus healthy controls. New method: Total haemoglobin (HbT) from our previously published Mini-Cog exercise-intervention dataset in older adults with type 2 diabetes and healthy controls was re-analysed using a harmonized modal phenotyping framework. Trigger-bounded INIT/LAST segments were processed by three estimators – multiverse Koopman dynamic mode decomposition (DMD), residual-validated Koopman DMD, and numerical algorithms for subspace state-space system identification/operational modal analysis (N4SID/OMA). Brain modes were classified using spatial, haemodynamic and short-separation evidence, cardiorespiratory modes were labelled by physiological bands and evaluated with internal automatic multiscale peak detection support. Results: The DMD-family estimators revealed a reproducible INIT to LAST increase in brain modal frequency and spatial structure across all groups. Multiverse DMD showed false discovery rate (FDR)-significant effects for all four primary brain metrics in all groups with Hedges’ (d_z=0.61-2.00), and residual-validated DMD reproduced the pattern. N4SID was conservative, yielding one sensitivity supported primary cell. Mixed models showed no FDR-significant univariate Group x Phase interaction. External fNIRS–ECG–respiration validation showed N4SID/OMA was most accurate for cardiac rate with mean absolute error 2.04–2.38 beats/min and (r=0.98) whereas respiration from prefrontal HbT was unreliable. Conclusions: Output only modal phenotyping provides a transferable, claim tiered approach for fNIRS dynamics. The data support a shared within-task brain-state transition, not a diabetes- or exercise-specific intervention effect.