Abstract The bimodal separation of unstructured clinical narratives and organized laboratory data frequently impedes Cardiovascular Disease (CVD) risk classification. To close this gap, we provide a Decoupled Dual-Expert Late Fusion architecture that uses Clinical BERT for semantic signal extraction and LightGBM for physiological pattern recognition. Prior to merging output through a weighted fusion layer, we maintained modality-specific feature hierarchies by training independent unimodal experts. The framework greatly outperformed the tabular-only baseline (AUROC: 0.988) with an almost flawless AUROC of 1.00 and an F1-score of 0.940 after analyzing 10,004 clinical records. The application of Disagreement Analysis, which found a 5.90% conflict rate where narrative "latent" symptoms conflicted with stable physiological markers, is a crucial contribution. Additionally, customized risk profiles were created to change AI from a "black box" to a therapeutic safety net that can be understood. These findings demonstrate that a decoupled multimodal strategy provides a more reliable, comprehensible pathway for early cardiovascular intervention. Keywords: Multimodal Machine Learning, Late Fusion, Clinical BERT, LightGBM