Artificial intelligence (AI) technologies have been increasingly applied in healthcare. However, the "black-box" nature of AI models makes it difficult for medical experts to understand the logic behind model decisions. To accurately translate AI decision patterns into human-interpretable knowledge, many leading scientists have proposed that natural laws should fit in low‑dimensional manifold structures. However, no existing method manages to construct such a manifold that faithfully represent AI decisions and its underlying natural data laws.

In a study published in Nature Biomedical Engineering, Prof. CAI Yunpeng's team from the Shenzhen Institute of Advanced Technology (SIAT) of the Chinese Academy of Sciences developed a mathematical framework named Class-Association Manifold Learning (CAML), which efficiently maps the decision rules of black-box AI models into a low-dimensional manifold that friendly visualizes knowledge to human experts.

Using manifold learning and generative AI, researchers imposed a manifold decomposition that disentangles natural data into two complementary sub-manifolds: a low-dimensional class-association sub-manifold that encodes all features relevant to AI decisions and governs nearly all model behaviors, and a high-dimensional sub-manifold that encodes personalized features irrelevant to model decisions.

This method not only compresses AI decision rules into a low-dimensional manifold, but also enables targeted modification of data samples and synthesis of realistic counterfactual examples, allowing clinicians to intuitively interpret hidden patterns discovered by AI and translate them into actionable clinical knowledge, and gaining insights such as decision boundaries, subtype distributions, and lesion signatures.

Researchers validated CAML across diverse biomedical datasets including ophthalmic fundus images, Optical Coherence Tomography angiography, chest X-rays, brain tumor magnetic resonance imaging, electrocardiography and gene expression profiles.

The results showed that CAML compressed deep learning model decisions into an eight-dimensional space with only about 1%-3% accuracy loss, which is about 1/10 the loss and 1/3 to 1/10 the dimensionality compared to explainable methods. Local feature explanations significantly outperformed traditional methods. Extracted patterns aligned closely with established knowledge. Blind evaluations by experts consistently preferred CAML explanations by a large margin, supporting its applications for computer-aided diagnosis and medical knowledge discovery.

This work offers insights for improving the safety, compliance, and acceptance of AI-enabled medical devices, enhancing the quality of AI-assisted diagnosis, and accelerating AI-driven medical discovery.