Knee osteoarthritis (KOA) is a high-impact degenerative disease whose diagnosis can be improved with artificial intelligence (AI). Existing models, however, often lack validation for the Brazilian population and efficient integration into clinical workflows. This study presents the development and validation of a multimodal model for KOA diagnosis using radiographic images and clinical epidemiological data from the ELSA-Brasil MSK study. The methodology employed a serial triage architecture: a clinical model (Logistic Regression optimized via Minimal Predictive Model Search [MPMS]) acts as a high-sensitivity filter, followed by a Deep Learning image model (DenseNet-161) for final confirmation. The image model achieved an Area Under the ROC Curve (AUC ROC) of 0.866, while the optimized clinical model achieved an AUC-ROC of 0.827. In a simulated triage scenario, this serial approach reduced the need for radiographic exams by 44.5% while maintaining a high overall sensitivity of 84.5% and achieving a final specificity of 99.7%. These results demonstrate that a multimodal AI approach can provide a robust, cost-effective, and validated tool for KOA screening in the Brazilian context.