Cardiovascular risk management is increasingly challenged by heterogeneous patient status and inconsistent recommendations from multiple clinical guidelines, which require structured, personalized, and safety-aware decision support. To address this issue, we propose a constraint-aware optimization framework that integrates patient state coding, multi-criteria candidate generation, learnable utility modeling, safety constraint filtering, and final decision optimization in a unified computational pipeline. The framework is evaluated on three cardiovascular datasets using 70/15/15 splits and repeated experiments. Experimental results show that the AUC of the proposed method is 0.88, the accuracy is 0.86, the F1-score is 0.83, the guideline consistency is 0.87, the personalization index is 0.54, and the safety violation rate is 0.06. It is better than the comparison methods in the prediction and decision support indicators. In conflict cases, the framework maintains an AUC of 0.86 and an accuracy of 0.84, and the decision stability at low risk, medium risk, high risk and very high risk levels is 0.88, 0.86, 0.83 and 0.81, respectively, which can output reliable and clinically consistent treatment recommendations.
