Abstract: Current mainstream deep learning models, due to insufficient interpretability, struggle to meet the demands for decision transparency and controllability in risk-sensitive domains such as healthcare, justice, and finance. To address the limitation that existing supervised learning methods focus on classification prediction while neglecting the transformability of sample features, this research introduces Extenics theory and proposes a basic-element transformation algorithm based on decision trees. The algorithm first clarifies the essential differences between classification and transformation problems and establishes a formal mathematical model for the transformation problem. It then defines a transformation distance to quantify the difficulty of transforming between different feature values. On this basis, a general transformation mechanism is designed which can be embedded into arbitrary decision-tree models, enabling quantitative derivation of feature-level transformation paths. Experimental results demonstrate that the proposed algorithm can effectively mine samples in the positive extension domain and improve the overall compliance rate in personalized healthcare scenarios, verifying the effectiveness and generalizability of the method.