Alzheimer's disease (AD) is a progressive neurological disorder characterized by amyloid accumulation and behavioral changes across preclinical and clinical stages. Although AD is currently incurable, emerging machine-learning (ML) approaches are facilitating novel insights into disease etiology and treatment mechanisms. Our study utilized and built upon two open-source ML platforms, DeepLabCut (Mathis et al. 2018) and Variational Animal Motion Encoding (VAME) (Luxem et al. 2022), to track and analyze naturalistic behavior in App knock-in (KI) and 5xFAD transgenic mouse models of AD. New humanized App-KI models develop amyloid plaques and mimic preclinical disease mechanisms, but have an inconsistent behavioral phenotype when assessed by conventional methods. To bridge this technological gap, we use the unsupervised behavioral segmentation tool VAME to break down complex behavioral sequences into smaller postural units (motifs) and to assess the hierarchical structure (community dendrogram) of spontaneous behavior in the open field. Our analysis revealed significant behavioral alterations in motif usage and transitions in AppNL-G-F mice many months before the appearance of overt cognitive symptoms. To define spontaneous behavioral alterations in a clinical-stage model of AD and to determine whether these alterations are amenable to therapeutic intervention, we studied 5xFAD mice expressing wildtype or mutant fibrinogen (Fgg$\gamma$390–396A). 5xFAD-Fgg$\gamma$390–396A mice lack the inflammatory CD11b receptor binding site, thereby preventing CD11b receptor–mediated microglia interactions that are associated with cognitive decline in AD. We found that 5xFAD mice displayed severe spontaneous behavior abnormalities in motif usage and transitions, many of which were effectively rescued by blocking fibrinogen-microglia interactions in 5xFAD-Fgg$\gamma$390–396A mice. Notably, VAME outcomes (motif usage) outperformed conventional outcomes (distance, speed, time, zone) in a logistic regression classifier model, correctly labeling 100\% vs 74.5\% of subjects. Our study supports VAME’s utility as a sensitive, unbiased tool for discerning AD behavioral manifestations and treatment efficacy in preclinical and clinical mouse models.