PROJECT SUMMARY/ABSTRACT Thrombectomy has significantly improved stroke outcomes. Nearly 80% of our clinic population now present with small strokes and low NIH Stroke Scale scores. However, greater than half endorse significant problems with attention, executive function, and processing speed. For many, significant recovery is seen by 6 months, but up to one third experience persistent vascular cognitive impairment. A biomarker to robustly predict who will exhibit long-standing deficits would enable us to initiate early interventions to slow or even prevent decline. Our work with MEG suggests global disruption of cognitive networks irrespective of stroke size or location; however, the compensatory mechanisms that allow some to recover but fail in others are poorly understood. There is a critical need for a noninvasive, inexpensive screening tool that can be widely implemented. The scientific premise of this proposal is two-fold: (i) using MEG and EEG we can determine functional network characteristics affecting both those with transient post-stroke cognitive impairment (psMCI) and persistent vascular cognitive impairment (VCI) as well as the compensatory mechanisms responsible for recovery, and (ii) a novel deep learning model that performs multimodal (MEG and EEG) learning to find shared signatures of VCI, but ultimately yields a model that needs affordable EEG-only data, will yield a powerful biomarker that can predict conversion of psMCI to VCI early after stroke. This proposal will pursue three specific aims. 1) Identify neurophysiologic similarities between transient psMCI and persistent VCI; 2) Identify specific features of functional connectivity that prognosticate conversion to VCI; 3) Design a digital biomarker that predicts conversion using functional brain networks that can be extended from MEG to EEG. To achieve these aims, we will collect both MEG and EEG data from 200 patients with minor stroke, evaluate their signals with expert neurophysiologists, and monitor the patient’s yearly conversion rate to VCI. We will then design and validate a deep learning model called Siamese Multiple Graph to Gauss (SMG2G), which performs multimodal learning on MEG and EEG network (graph) data but ultimately yields a model that needs EEG-only data to make predictions of conversion to VCI. The final product will be an EEG digital biomarker that can be readily measured and widely employed across the country. The research proposed in this application is innovative because it is the first to use functional network signals to design a biomarker for VCI that is inexpensive and widespread, yet robust, and achieves this by cutting edge machine learning. It is also significant because it will advance the field vertically both scientifically and clinically by enabling large-scale, early detection of VCI. Our team is well-prepared to undertake this project, with clinical and engineering expertise, strong collaborations, preliminary data supporting the aims, an...