AI-assisted modeling of attention quantifies engagement and predicts cognitive improvement in older adults

Abstract

Background: Cognitive training aims to prevent or slow cognitive decline in older adults, but outcomes vary widely. Engagement, describing how individuals allocate cognitive, affective, and physiological resources, is critical to training benefits, yet behavioral metrics lack real-time modeling of attention and do not reliably predict outcomes. We developed and validated a multimodal, AI-assisted biomarker that quantifies attentional states during computerized cognitive training and predicts cognitive improvements. Methods: We designed the Attentional Index using Digital measures (AID), leveraging a video-based facial expression encoder (pretrained on 38,935 videos), an ECG-based autonomic encoder (pretrained on 123,998 ECG samples), and a temporal fusion module. Using two processing speed/attention studies in older adults (> 65 years) with mild cognitive impairment, AID was trained and evaluated in BREATHE (n=50; ~300 hours from 368 sessions) and validated in FACE (n=20; ~150 hours from 219 sessions). Model training targeted session-level change in self-reported fatigue. Clinical validation tested relationships between AID scores and (1) behavioral attention, (2) cognitive outcomes, and (3) neural correlates. Findings: AID accurately detected engagement changes (BREATHE: accuracy 0.82, F1=0.81; FACE: accuracy 0.73, F1=0.74), outperforming unimodal models. AID scores were unrelated to session, task type, or demographics. In BREATHE, session-level AID scores significantly predicted executive function improvement (sessionxAID: Wald^2=7.85, p=0.005), whereas reaction time variability did not. Lower AID intercepts (B=-0.07{+/-}0.03, p=0.043) and steeper slopes (B=0.31{+/-}0.15, p=0.046) were associated with greater improvements. Post hoc analyses identified two engagement profiles linked to better attention: one characterized by low-RMSSD and focused periocular activation, and the other defined by coherent alignment between low-RMSSD and facial expression patterns. Interpretation: AID provides a reliable digital biomarker of effective engagement and predicts cognitive improvement beyond behavioral metrics. By capturing facial-autonomic dynamics of attention, AID offers a foundation for closed-loop cognitive intervention design. Funding NIH AG081723, NR015452, and AG084471; Stanford HAI seed funding.