<p>Machine Learning has achieved considerable success in clinical applications such as image-based diagnostics, predictive modeling for patient outcomes, and personalized treatment planning. However, the black-box nature of deep neural networks often results in poor interpretability and reliability of predictions. Traditional neural network architectures, focusing primarily on correlations, fall short in elucidating underlying causal medical mechanisms. Addressing this, causal discovery, aimed at elucidating the structure of causal graphical models from observational or experimental data, is gaining prominence in clinical fields demanding high reliability. Nevertheless, the complexity of search algorithms, the scarcity of real-world data, and the challenges in identifying unique results significantly hinder the reliability of these approaches. To overcome these challenges, we propose an iterative active structure learning approach to ensure reliable clinical causal analysis. Our method begins with the recovery of a causal structure, guided by a set of prior causal presence, followed by an iterative process of active refinement to enhance the output reliability. This involves using violations of known clinical mechanisms as structural constraints to guide successive rounds of learning, thereby correcting and refining the model iteratively. The process continues until there is a convergence between expertise and the data-derived solutions. Our experiments on real-world clinical data demonstrate that Our approach can improve the quality of causal findings and discover new causal associations beyond the basis of expert knowledge. Furthermore, our approach has yielded novel and significant insights from various datasets, which we explore in our discussion.</p>