Neural features underlying high-accuracy classification of amnestic mild cognitive impairment using multi-sensory-evoked potentials.

Early detection of amnestic mild cognitive impairment (aMCI) is crucial for timely interventions. This study combines scalp recordings of lateralized auditory, visual, and somatosensory stimuli with a flexible and interpretable support vector machine learning pipeline to differentiate individuals diagnosed with aMCI from healthy controls. Event-related potentials (ERPs) and functional connectivity (FC) matrices from each modality successfully predicted aMCI. Reduced ERP amplitude in aMCI contributed to classification. The analysis of FC using phase-locking value revealed higher FC in aMCI than controls in frontal regions, which predicted worse cognitive performance, and lower FC in posterior regions from delta to alpha frequency. We observe optimal classification accuracy (96.1%), sensitivity (97.7%), and specificity (94.3%) when combining information from all sensory conditions rather than when using information from a single modality. The results highlight the clinical potential of sensory-evoked potentials in detecting aMCI, with optimal classification using both amplitude and oscillatory-based FC measures from multiple modalities.