Abstract

Background: Alzheimer’s disease (AD) and dementia are growing global health challenges, yet public understanding of these conditions remains poorly characterized. Methods: This study analyzed 17,578 English-language tweets from Twitter/X collected throughout 2024 to characterize public discourse using dual sentiment analysis and topic modeling. Sentiment was assessed using two lexicon-based tools, VADER (Valence Aware Dictionary and sEntiment Reasoner) and TextBlob, with inter-method agreement evaluated via Cohen’s Kappa and confusion matrix analysis. Latent Dirichlet Allocation (LDA) with coherence-based model selection (c_v) identified latent discussion topics. Results: All reported sentiment proportions are tool-specific estimates rather than absolute characterisations of public discourse and should be interpreted in light of the fair inter-method agreement observed between VADER and TextBlob. VADER classified 43.3% of tweets as positive, 32.2% as negative, and 24.5% as neutral (mean compound score = 0.065, SD = 0.499; 95% CI: 0.058 to 0.073; note that the large SD relative to the mean reflects the broad, near-zero-centred distribution of compound scores). TextBlob produced a partially divergent distribution (45.9% positive, 36.2% neutral, 17.8% negative) with fair inter-method agreement (κ = 0.297), indicating that exact sentiment proportions depend meaningfully on the choice of classifier. Approximately 14% of tweets contained political content related to the U.S. presidential election cycle, which may represent an important source of noise in health discourse. LDA identified six optimal topics (c_v = 0.407): research studies, blood-based diagnostics, comorbidities and personal loss, research advocacy, clinical developments, and risk factors and treatment. Sentiment differed significantly across topics (Kruskal–Wallis H = 235.92, p < 0.001), with research and advocacy topics showing the most positive VADER-classified sentiment (50.0%) and comorbidity discussions showing the highest negative proportion (41.4%). Positive tweets received significantly higher engagement than negative tweets (Mann–Whitney p < 0.001), though follower count was the strongest predictor of engagement (β = 0.198). Conclusions: These findings, interpreted as tool-conditional estimates, have implications for health communication strategies, public education campaigns, stigma reduction, and the responsible design of social media-based health surveillance systems.

IPC Classification

Keywords