<p>Personalized medicine advocates that physicians should provide medical treatments based on patients’ individual conditions. However, combining a patient’s medical history and current condition to develop a customized treatment plan and evaluate its possible outcomes is time-consuming. To address this issue, various outcome prediction methods and medical decision-making systems have emerged to assist physicians in the above process. Unfortunately, they cannot well support outcome-driven treatment optimization, where physicians can iteratively optimize treatment plans according to expected outcomes, aiming to provide the best possible treatment plan for patients. Therefore, this paper proposes an interactive visual analytics system, TOVis, to support outcome-driven treatment plan optimization using a neural network model and post-analysis technology. TOVis can guide physicians to optimize treatment plans in two manners: manual adjustment based on domain knowledge and experience, and automatic adjustment under the constraints of expected outcomes. We also design a provenance tree to organize the optimization history, which enables visual comparisons of different plans to make medical decisions. Finally, quantitative evaluation, case studies, and physician interviews on an acute kidney injury dataset demonstrate the usefulness and effectiveness of TOVis in treatment planning.</p> Graphical abstract <p></p>