New bandwidth selection criterion for Kernel PCA: approach to dimensionality reduction and classification problems.

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Diagnostics

Background:

• DNA microarrays are crucial for cancer diagnostics, treatment selection, and prognosis.
• High-dimensional data from microarrays presents challenges for prediction models.
• Dimensionality reduction techniques like feature selection and transformation are vital preprocessing steps.

Purpose of the Study:

• To develop a computationally efficient bandwidth selection criterion for Kernel PCA (KPCA).
• To propose a new prediction model integrating KPCA with Least Squares Support Vector Machine (LS-SVM).
• To evaluate the performance of the proposed model against established methods for cancer outcome prediction.

Main Methods:

• A novel data-driven bandwidth selection criterion for KPCA, linked to least squares cross-validation.
• Integration of the optimized KPCA with LS-SVM for a new prediction model.
• Comparative analysis using 9 case studies, evaluating Area Under the ROC Curve (AUC) and computational time.

Main Results:

• The proposed KPCA bandwidth selection method is computationally less expensive than existing approaches.
• The combined KPCA + LS-SVM model demonstrates competitive accuracy and reduced computational time.
• Performance comparison showed the proposed strategy's efficiency against methods like PCA, t-test, PAM, and Lasso.

Conclusions:

• Feature transformation and selection are effective for classification in medical diagnostics.
• The proposed strategy offers a more efficient approach to dimensionality reduction for cancer outcome prediction.
• The new method provides a practical advantage due to its reduced time complexity.