This paper addresses the multi-objective optimization problem in health science popularization short video recommendation by proposing a Multi-Stage Iterative Optimization Algorithm based on Disease Label Detection Mutation Strategy (MSIO-DLD). The algorithm introduces a multi-stage optimization framework that iteratively optimizes the diversity, professionalism, and accuracy of the recommendation list at each stage, aiming to achieve a comprehensive balance among these three core objectives. During the diversity optimization phase, MSIO-DLD incorporates a mutation strategy based on disease label detection, effectively mitigating the over-concentration of recommended content within a single category. Simultaneously, cosine similarity is employed to control the deviation between the mutated recommendation list and the original one. Experimental results demonstrate that MSIO-DLD outperforms seven representative baseline algorithms across multiple evaluation metrics, particularly excelling in professionalism and accuracy. Furthermore, MSIO-DLD exhibits superior performance in key indicators such as Hypervolume and Inverted Generational Distance, validating its effectiveness and robustness in addressing multi-objective optimization problems. Future research will focus on integrating domain knowledge from the medical field to further refine the recommendation process and explore more innovative strategies to enhance algorithm performance.