Dimensionality reduction is a crucial process in medical data mining, especially for predictive analytics involving high-dimensional clinical and biomarker datasets used in lung cancer diagnosis. These datasets often suffer from redundancy, noise, and the “curse of dimensionality,” making effective feature extraction necessary for accurate prediction and computational efficiency. This study proposes a non-linear dimensionality reduction approach using Uniform Manifold Approximation and Projection (UMAP) to extract the most relevant features while preserving the intrinsic structure of lung cancer data. The method constructs a weighted k-nearest neighbor graph and optimizes low-dimensional representations through cross-entropy minimization, capturing both local and global relationships more effectively than traditional linear techniques such as Principal Component Analysis (PCA). Experimental evaluation demonstrates that UMAP improves classification accuracy, model interpretability, and execution time for lung cancer diagnosis and subtype identification. By retaining essential clinical and biomarker patterns, UMAP strengthens early detection models and enhances clinical decision-support systems. The results highlight UMAP’s broader applicability to other complex biomedical datasets, with future work focusing on integration with deep learning frameworks for precision medicine.
