This study is a narrative review informed by bibliometric mapping [9,10]. Bibliometric analysis was used as a descriptive tool to visualize the international research landscape, and narrative synthesis was used to integrate the bibliometric findings with evidence from systematic and scoping reviews and with the researchers’ expertise in gerontological nursing. This narrative review was reported in accordance with the Scale for the Assessment of Narrative Review Articles (SANRA) [11].

Two databases were searched to capture complementary perspectives. For international literature, we searched PubMed from January 1, 2015 to December 30, 2025. Search terms combined dementia-related terms (“dementia”, “Alzheimer’s disease”, “cognitive impairment”) with AI-related terms (“artificial intelligence”, “machine learning”, “deep learning”, “chatbot”, “robot”) using Boolean operators, applied to title, abstract, and keyword fields. A total of 5,710 records were retrieved. For Korean literature, we searched the Research Information Sharing Service (RISS) using Korean equivalents of the same terms, yielding 265 peer-reviewed journal articles.

The same eligibility criteria were applied to both databases. Publications were included if they (1) focused on dementia, Alzheimer’s disease, or cognitive impairment; (2) involved AI-related methods or applications; and (3) were published between January 2015 and December 2025. Exclusion criteria were: (1) conference abstracts, editorials, letters to the editor, commentaries, and errata; (2) preprints and non–peer-reviewed reports; and (3) dissertations and theses. Cognitive impairment was included to capture research on early-stage and prodromal conditions, which represent a critical window for nurse-led preventive interventions.

Keyword co-occurrence analysis was conducted using VOSviewer version 1.6.20 [9]. Both Medical Subject Headings (MeSH) and keywords were extracted from the PubMed records. Synonymous and morphological variants (e.g., “Alzheimer’s disease” and “Alzheimer disease”) were merged using a thesaurus file. A minimum occurrence threshold of 15 was applied to retain keywords. Full counting was used as the counting method, and association strength normalization was applied to construct the keyword co-occurrence network [10]. The Leiden algorithm [12] was applied with a resolution parameter of 1.00. Each cluster was characterized by its representative keywords, mean publication year, and total link strength.

The 265 articles retrieved from RISS were reviewed by the authors and categorized according to study design, AI technology type, primary outcomes, and methodological limitations. Given the narrative approach and lack of access to individual study data, patterns are reported qualitatively rather than as formal frequency counts.

International publications on AI in dementia research increased modestly from 2015 to 2020, accelerated between 2021 and 2023, and expanded sharply after 2024 (Table 1, Figure 1). This recent increase may partly reflect growing interest in LLM-based chatbots, natural language processing tools, and generative AI–based clinical support systems.

Keyword co-occurrence analysis of the 15 keywords meeting the minimum threshold identified five thematic clusters (Figure 2).

Analysis of the 265 studies showed that qualitative studies and literature reviews predominated, while empirical studies testing intervention effectiveness were relatively scarce (Table 2). Technical proposal studies describing AI model architectures were also substantially represented. Intervention studies were concentrated on hardware-based conversational AI devices such as care robots and AI speakers (e.g., Aria, Clova), which were primarily used for emotional support, companionship, medication and schedule reminders, and cognitive stimulation. The most frequently measured outcomes were depression, loneliness, quality of life, and social isolation, reflecting a pronounced focus on psychosocial dimensions. Evaluation of cognitive function, ADL performance, and caregiver burden was comparatively limited. Methodological limitations were consistently observed, including the absence of control groups, non-randomized designs, and follow-up periods typically shorter than 3 months. These findings are consistent with the broader observation that AI research in older adult healthcare has generally emphasized performance, usability, and feasibility over clinical effectiveness [13].

Drawing on the keywords of Cluster 1 and integrating findings from recent systematic and scoping reviews on AI applications in nursing care [14,15], we identified four AI application domains most directly relevant to dementia nursing practice: monitoring technologies, digital biomarkers, AI chatbots for caregivers, and clinical decision support system (CDSS).

The first domain was monitoring technologies. AI-based monitoring is the most actively researched application domain in dementia care. Wearable sensors, ambient sensors, and camera-based systems have been used to monitor activity levels, sleep patterns, wandering behaviors, and ADL performance in both community-dwelling and institutionalized individuals with dementia [16]. AI algorithms can process these signals to detect real-time indicators of functional decline, safety risks, and BPSD exacerbation. However, most studies remain at the feasibility stage, and randomized controlled trials (RCTs) demonstrating effects on clinically meaningful outcomes such as fall rates, hospitalization, and caregiver burden are still lacking.

The second domain was digital biomarkers. Digital biomarkers derived from voice acoustics, linguistic features, gait kinematics, and keystroke dynamics have emerged as a promising area for dementia screening and longitudinal monitoring [17]. Unlike traditional blood- or neuroimaging-based biomarkers, digital biomarkers can often be collected passively through consumer-grade devices, enabling low-burden, continuous assessment in community settings. A recent scoping review mapped the landscape of AI-based digital biomarkers in Alzheimer’s disease and highlighted the need for clinical validation [18]. From a nursing perspective, the key value of digital biomarkers lies not in diagnosis per se but in the early detection of cognitive changes as a trigger for proactive nursing interventions.

The third domain was AI chatbots for caregivers. AI chatbots have evolved from rule-based systems focused on information retrieval and reminders to generative AI–enabled conversational agents capable of context-sensitive and personalized interaction. LLM-powered systems may support caregiver education, emotional support, and access to health and social resources [6]. However, dementia care chatbots remain in an early stage of development, with limited evidence-based content, concerns about response accuracy, and insufficient rigorous evaluation with end users [19]. Nurses are well positioned to advance this field by developing evidence-based response protocols, establishing criteria for professional escalation, and evaluating outcomes such as caregiver burden, caregiving self-efficacy, caregiving skills, and service use. Rigorous nurse-led evaluation is therefore a critical prerequisite for the safe and effective integration of AI chatbots into dementia care.

The fourth domain was CDSS. AI-based CDSS can support nursing assessment, care planning, and intervention timing. A recent scoping review identified diagnostic accuracy support, workflow optimization, and risk stratification as core contributions of AI-based CDSS to nursing decision making [20]. Reported applications in dementia care include antipsychotic prescription adjustment, risk stratification for BPSD-related behavioral events, and decision support for family caregivers regarding transitions to end-of-life care. Integration of CDSS into nursing workflows, however, requires careful attention to nurse trust in algorithmic recommendations, alert fatigue, and interoperability with electronic medical record systems; a systematic review identified alert inappropriateness and poor workflow fit as the most consistent barriers to effective CDSS use [21]. These considerations underscore the need for nurse-led evaluation of CDSS usability and clinical impact prior to large-scale implementation.

The contrast between international and Korean research patterns warrants attention. Internationally, research on multimodal AI systems integrating neuroimaging, voice, and behavioral data has expanded, whereas Korean research has tended to focus on hardware-based emotional support devices such as care robots and AI speakers. Rather than indicating a purely technological difference, this pattern may reflect contextual differences in dementia care priorities. In East Asian contexts, family-centered caregiving norms shaped by filial piety influence caregiving roles, help-seeking, and expectations for emotional and practical support [22]. In Korea, these cultural expectations coexist with a national dementia care infrastructure that includes Dementia Relief Centers and caregiver support services [23]. These contextual factors may help explain why domestic AI studies have emphasized socially interactive and supportive technologies.

The outcome profile of Korean AI-dementia research also warrants consideration. Affective and psychosocial outcomes—depression, loneliness, quality of life, and social isolation—were more frequently examined than cognitive, functional, or safety-related outcomes. Although these outcomes are important, this pattern may not fully capture the multidimensional effects of AI-based dementia interventions. Dementia intervention research has emphasized the need to assess outcomes across cognitive, functional, quality-of-life, behavioral, and caregiver-related domains [24]. Future studies should therefore combine psychological measures with standardized neuropsychological assessments, ADL/instrumental ADL (IADL) measures, and validated caregiver burden instruments.

As AI becomes more embedded in dementia care, several ethical considerations require deliberate attention. Individuals with moderate to severe dementia may have limited capacity to consent to continuous data collection through wearables or sensors; proxy consent procedures must therefore be clearly established and periodically reviewed in accordance with institutional review board requirements.

AI systems trained on underrepresented data may produce biased predictions for socioeconomically disadvantaged groups. In this context, digital ageism refers to the ways in which AI systems may encode, reproduce, or amplify age-related bias through underrepresentation of older adults in training data or through design decisions that overlook their needs [25]. Because dementia predominantly affects older adults, dementia-related AI systems may be particularly vulnerable to such biases. Algorithmic transparency, equity auditing, and the meaningful inclusion of older adults and their caregivers in AI development should therefore be embedded in the design and governance of AI tools for dementia care [25].

Based on the identified application domains, outcome gaps, and ethical considerations, we propose a four-component framework for nurse-led dementia AI research: Target, Outcome, Mechanism, and Integration. Target identifies whom the AI tool is designed to support, including persons living with dementia, family caregivers, nurses, or care teams. Outcome specifies what should be changed, such as BPSD frequency, ADL/IADL performance, falls, hospitalization, caregiver burden, caregiving self-efficacy, or quality of life. Mechanism clarifies how the AI tool is expected to produce change—for example through passive monitoring, risk detection, personalized feedback, caregiver education, or clinical decision support. Integration addresses how the tool is embedded in practice, including alert design, escalation pathways, staff training, interoperability, consent procedures, and equity monitoring. This framework can guide future studies beyond feasibility testing toward clinically meaningful, ethically governed, and practice-integrated implementation of AI tools in dementia care.

This review combines bibliometric mapping with narrative synthesis to link international and Korean AI-dementia research trends to gerontological nursing practice. However, because it was not designed as a systematic review, narrative interpretation may be subject to selection bias. The PubMed-based search may also have missed relevant studies indexed only in engineering or computer science databases. In addition, the four nursing-relevant application domains were identified through expert-informed interpretation rather than a pre-specified quantitative selection procedure. Thus, the findings should be interpreted as a trend-oriented overview rather than an exhaustive synthesis of the evidence.