This paper proposes a vertical federated learning framework incorporating differential privacy, which safeguarded confidential healthcare data through collaborative model training. The current literature employed analogous strategies, nevertheless, challenges such as considerable processing overhead, susceptibility to variable data distributions, and unstable convergence resulting from inflexible noise calibration were observed. A solution has been developed by combining adaptive noise calibration with repeated global aggregation and client-level denoising, thereby facilitating a balanced privacy-utility trade-off. The suggested methodology was executed utilizing a logistic regression model on heart disease datasets, yielding 97.36%, 87.60% and 95.75% final local accuracy of client-1, client-2 and client-3 respectively, with early stopping noted at round 5. The advantages of the proposed approach were evidenced by (1) the implementation of dynamically adjusted federated learning with differential privacy that reduces significant performance decline, (2) an iterative aggregation protocol for model accuracy convergence, (3) denoising at client-level to alleviate noise effects, (4) decreased communication overhead and computational complexity through effective parameter exchange, and (5) improved security against inference and reconstruction attacks. The proposed approach demonstrated efficacy in both performance and real-world applicability, thus providing an appealing option for the scalable and secure implementation of privacy-preserving vertical federated learning systems in healthcare as well as privacy-sensitive fields.