Revealing the Latent Structure of Health Literacy in Online Peer-to-Peer Communication with Uncertainty-aware Bayesian Deep Learning.

Health literacy is a multidimensional construct essential for health decision-making, yet its computational assessment from naturalistic online discourse remains limited by categorical classifications that fail to capture its latent, continuous, and context-sensitive nature. To address this gap, we introduce an uncertainty-aware Bayesian deep learning framework that probabilistically infers latent health literacy from social media text while systematically quantifying predictive uncertainty. Using a large corpus of English-language health forum posts (N = 342 k), we operationalized five theoretical dimensions of health literacy-Functional, Communicative, Critical, Digital, and Expressed-through validated NLP features. A Bayesian Variational Autoencoder with Monte Carlo Dropout models health literacy as a continuous latent variable and provides epistemic uncertainty estimates. The framework recovers a robust three-factor latent struc- ture: Core Integrated Proficiency (merging Critical, Communicative, and Expressed dimensions), Digital Proficiency (exhibiting an inverse association with Functional Literacy), and Applied Functional Literacy. The model achieves strong reconstruction performance (MSE = 0.109) with uncertainty estimates reliably correlated to prediction error (r = 0.438). From the continuous latent representations, we derive three distinct user profiles-Balanced, Specialized, and Transitional-revealing heterogeneous patterns of health literacy expression and adaptive communication behavior. This work advances computational health literacy assessment by providing a probabilistic, uncertainty-aware framework that moves beyond static categorization, with direct implications for personalized public health communication and hybrid human-AI assessment systems.