Background: Early detection of mild cognitive impairment (MCI) is critical as global dementia burden grows. Extracellular vesicles (EVs) are promising biomarkers; these particles carry components indicative of their cell of origin and its physiological state, providing insight into brain cell interactions with amyloid, tau, and neurodegeneration (ATN) markers. Magnetic resonance imaging (MRI) can provide complementary anatomical brain measures for a more holistic understanding of patient conditions. This project explores the potential of multimodal models integrating blood EV and MRI markers to improve MCI detection.

Methods: 68 participants were selected from the Gait & Brain Cohort Study (mean age(SD)=70.66(7.187), M/F=27/41), diagnosed with normal cognition (n=35) or MCI (n=33). Blood collection, cognitive testing, and MRI were performed at time of recruitment. EVs in blood were co-labelled for neuron- or microglia-specific markers, amyloid beta (Aβ42), pTau181, and neurofilament light chain (NfL) and quantified using nanoscale flow cytometry. MRI scans were analyzed using a semi-automated pipeline to quantify structural features, white matter hyperintensities (WMH), and enlarged perivascular spaces. Logistic regression models were trained to query the performance of EV and MRI biomarkers to identify MCI either alone or in combination.

Results: Neuron- and microglia-specific EVs colabelled with pTau181 are significantly decreased and increased in MCI respectively, suggesting changes and cell-specific tau dynamics as MCI progresses. Correspondingly, pTau181+ neuronal and microglial EVs positively and negatively correlate with verbal working memory respectively as measured by Backwards Digit Span. White matter volume significantly decreased in MCI while WMH burden significantly increased. White matter volume is positively correlated with the visual search and processing speed as measured by the TMT-A. Multimodal logistic regression models trained using both EV and MRI metrics better distinguished identifying participants with MCI than models trained on either biomarker modality alone, with improved
ROC-AUC, corrected AIC, and Brier Score. 

Conclusion: Multimodal models integrating circulating blood EV and MRI markers identify MCI better than either modality alone, while also providing complementary information for clinical decisions. Future work with larger sample sizes and more diverse populations is needed to confirm findings.