Alzheimer’s disease and related dementias (ADRD) represent a growing public health challenge, and informal caregivers regularly report high levels of stress. Prior research has typically treated caregiver burden, unmet needs, and system failures as separate domains. The authors sought to determine whether and how these domains co-occur within individual caregivers using large-scale, real-world online forum data. By mapping co-occurrence patterns, the study aims to reveal interconnected challenge structures that could inform how support services are designed.
The analysis used 7,198 posts from three ALZConnected caregiver forums (general forum, spouse/partner forum, and adult child caregiver forum). Posts were publicly available and de-identified. The study applied a large-language-model (LLM) classification framework, specifically Claude 3.5 Sonnet, to assign category labels at the post level. Stanford University Institutional Review Board approval was obtained (IRB Protocol #83818), and informed consent was waived due to the use of public de-identified data under minimal-risk classification.
Each post was coded for the presence of three domains: burden, unmet needs, and system failures. The authors used predefined category sets for each domain: nine categories for burden, twelve categories for unmet needs, and ten categories for system failures. The source reports these domain and category counts but does not list the full item-level category definitions in the abstract; those details are likely in the full text or supplementary materials.
The primary analyses compared post-level expression rates for each domain across caregiver role (primary versus secondary caregiver) and relationship to the care recipient (spouse/partner versus adult child). In addition to simple frequency comparisons, the authors built post-level co-occurrence networks to quantify how categories cluster within and across the three domains. Network density metrics and counts of significant cross-domain connections were used to summarize connectivity patterns.
Co-occurrence networks were constructed at the post level to capture how specific categories within one domain co-occur with categories in the same or different domains. The networks were assessed for within-domain clustering (reported network densities) and for cross-domain connections. The authors identified statistically significant cross-domain links and reported the strongest pairwise co-occurrences observed across the dataset.
Across the 7,198 posts, burden was present in 89.0% and unmet needs in 93.3% of posts; system failures appeared in 34.8% of posts. Comparing caregiver roles, primary caregivers expressed burden more often than secondary caregivers (91.6% vs. 84.7%). Conversely, secondary caregivers reported higher rates of unmet needs (94.6% vs. 92.5%) and higher reporting of system failures (37.2% vs. 33.4%).
When stratified by relationship to the care recipient, adult child caregivers reported higher rates than spousal caregivers across all three domains (burden, unmet needs, and system failures). Network analyses showed dense within-domain clustering, with reported densities in the range of 0.61–0.65. There were 84 statistically significant cross-domain connections overall. The strongest cross-domain links were specific and interpretable: behavioral or safety-related burden most strongly co-occurred with safety-management needs (co-occurring in 21.7% of posts), and emotional burden co-occurred strongly with emotional-support needs (20.9% of posts).
The authors interpret these findings to mean that caregiver challenges do not occur in isolation but form an interconnected ecosystem of issues that vary by caregiver role and relationship. Dense within-domain clustering suggests coherent subdomains of problems, while multiple cross-domain links indicate that a single expressed problem (for example, behavioral burden) is likely to be accompanied by related unmet needs or experiences of system failures. From this perspective, interventions and support services designed for caregivers may be more effective if they account for these co-occurring patterns rather than targeting single domains in isolation.
The Institutional Review Board of Stanford University approved the research (IRB Protocol #83818) and classified it as minimal risk; informed consent and data use authorization were waived because the analysis used publicly available, de-identified forum posts. The underlying forum posts are publicly available on ALZConnected. The processed dataset used for the analysis is not publicly available; the authors state that researchers interested in the data may contact the corresponding author. The work declares no competing interests and reports funding from the Scan Foundation.