Dementia diagnosis remains a critical challenge in clinical practice, particularly in distinguishing Alzheimer’s disease (AD) from Dementia with Lewy Bodies (DLB). The purpose of this study is to develop a lightweight, privacy-preserving framework for quantifying nonverbal medical findings in dementia screening. The key idea is to integrate browser-based facial recognition with a dynamics analyzer that extracts temporal metrics such as expressiveness, reaction speed, and asymmetry. Our proposed method leverages pre-trained models within a client-side architecture, avoiding the transfer of sensitive video data to external servers. Controlled simulations were conducted in which a medical professional enacted AD-like and DLB-like behaviors under standardized tasks, including a written instruction to close the eyes and a pareidolia test with an embedded emotional stimulus. Experimental results demonstrate clear differences between the two simulated conditions. The DLB simulation exhibited higher variability ( \(E=0.087\) vs. 0.042), faster reaction dynamics ( \(R=0.036\) vs. 0.018), and greater asymmetry ( \(A=0.029\) vs. 0.011) compared to the AD simulation. These findings suggest that nonverbal cues, when systematically quantified, provide meaningful distinctions between dementia subtypes that extend beyond verbal and cognitive markers.