Integrated meta-analysis and exploratory small-sample machine learning to evaluate curcumin against osteoporosis: a...

To systematically evaluate the anti-osteoporotic efficacy of curcumin in preclinical models and utilize an exploratory small-sample machine learning (ML) framework to examine potential contributors to heterogeneity and generate preliminary dose-duration hypotheses.

Following PRISMA guidelines, we searched eight databases for controlled in vivo animal studies in validated osteoporosis models that compared curcumin monotherapy with saline or vehicle controls and reported extractable bone-related outcomes, including bone mineral density (BMD) and trabecular microarchitecture. Methodological quality was assessed using the SYRCLE tool. In addition to random-effects meta-analysis, an exploratory ML framework incorporating SHapley Additive exPlanations (SHAP) and Gaussian Process Regression (GPR) was employed to examine associations between nine study-level features and effect estimates and to explore potential non-linear dose-response patterns after normalization to the Human Equivalent Dose (HED).

Twenty-three studies were included. Meta-analysis revealed that curcumin significantly elevated femoral BMD [standardized mean difference (SMD) = 2.73, P < 0.001], preserved trabecular microarchitecture, and enhanced biomechanical strength. Curcumin was also associated with changes in bone-remodeling and oxidative stress-related markers. Based on only 23 study-level observations, exploratory ML suggested that body weight had the largest model-dependent contribution (mean |SHAP| = 0.201). GPR modeling suggested a bell-shaped dose-response relationship, with the fitted response reaching a local maximum at an HED of approximately 32 mg/(kg·d) and an intervention duration of 8-12 weeks. These model-derived findings should not be interpreted as validated optimal dose or duration estimates.

Curcumin exerts potent, multi-target osteoprotective effects that are associated with improved bone remodeling and oxidative stress-related indices. The exploratory integration of ML with meta-analysis suggested that biological characteristics and dosage may contribute to variability in treatment effects. However, because the ML component was constrained by the limited number of study-level observations, these model-derived findings should be regarded as hypothesis-generating signals and should not be interpreted as independently validated predictors or clinical dosing recommendations.

https://www.crd.york.ac.uk/prospero/, identifier CRD420251269951.