Facial phenotypes in Alzheimer's disease: from neurobiology to artificial intelligence

Facial analysis is increasingly being explored as a source of scalable behavioral signals relevant to Alzheimer's disease (AD) and AD-related cognitive impairment. In this narrative review, informed by a structured literature search, we summarize current evidence on the biological and behavioral basis of facial alterations in AD, with particular emphasis on affective expressivity, neuropsychiatric manifestations, and dynamic facial behavior. We also review representative artificial intelligence-based facial analysis methods, including commonly used datasets, feature representations, and modeling strategies, ranging from facial landmarks and texture descriptors to spatiotemporal video models, multimodal fusion, and language-enhanced frameworks. Current evidence remains limited by small and largely single-center cohorts, heterogeneity in acquisition settings and outcome definitions, inadequate control of confounding factors, limited external validation, poor calibration reporting, and persistent concerns regarding interpretability and clinical specificity. Within the evolving biomarker-based diagnostic framework of AD, facial analysis is better viewed as a candidate, non-specific, and context-dependent tool for auxiliary risk stratification, triage support, and longitudinal monitoring rather than as stand-alone diagnostic tests. Future progress will depend on standardized data acquisition, integration with clinical and biomarker data, improved explainability, and prospective real-world validation.