A multi-method AI framework integrating explainable machine learning, Bayesian multilevel modeling, and latent profile analysis for childhood stunting surveillance in Indonesia

Applying AI-driven analytical methods to real-world hierarchical health data presents methodological challenges—including latent constructs, nested structures, and non-linear relationships—that no single algorithm can fully resolve. This study develops and evaluates a multi-method AI triangulation framework integrating causal discovery, explainable machine learning with SHAP, Bayesian Multilevel Structural Equation Modeling (SEM), and Latent Profile Analysis (LPA), applied to district-level stunting prevalence and its determinants across 514 districts nested within 38 provinces in Indonesia. Among the key methodological findings, we document the conditions under which causal discovery algorithms exhibit limited sensitivity in real-world settings: the Bayesian Network Structure Learning and the Peter-Clark (PC) algorithm recovered only 33.3% (2 out of 6) of theoretically specified edges, attributable to weak effect sizes, measurement error attenuation, and level-of-analysis mismatches—providing preliminary considerations for responsible AI deployment. The Bayesian Multilevel SEM achieved full convergence (R-hat = 1.000), identifying provincial poverty as the dominant stunting determinant (standardized γ = 0.630; unstandardized γ = 0.356) and explaining 42.3% of between-province variance. The Random Forest–SHAP pipeline explained 38.95% of district-level variance and detected a non-linear protective effect of population density (percentage increase in mean squared error = 29.825%) undetectable by linear methods, demonstrating diagnostic value of cross-method divergence in AI ensembles. Latent Profile Analysis identified three structurally distinct district typologies; Profile 2 (N = 102 districts) exhibited the highest stunting burden with co-occurring deficits across demographic and economic dimensions, while Profile 3 (N = 92 districts) showed favorable outcomes with strong service access, offering exploratory typologies to support discussions on tailored AI-assisted policy targeting. This framework advances responsible integration of explainable AI methods, structural equation modeling, and probabilistic clustering for hierarchical public health surveillance.