k-Nearest neighbor models for microarray gene expression analysis and clinical outcome prediction

Area of Science:

• Genomic data analysis and bioinformatics
• Computational biology and machine learning in medicine

Background:

• Genomic data analysis is crucial for disease classification and outcome prediction.
• The k-nearest neighbor (KNN) modeling strategy shows performance variability in clinical applications, as observed in the MicroArray Quality Control Phase II (MAQC-II) project.

Purpose of the Study:

• To systematically evaluate factors influencing the performance of KNN models in clinical genomic data analysis.
• To identify and validate a robust KNN data analysis protocol for disease classification and outcome prediction.

Main Methods:

• Generated 463,320 KNN models by varying feature ranking, number of features, distance metric, neighbors, vote weighting, and decision threshold.
• Utilized clinical data from breast cancer, neuroblastoma, and multiple myeloma for model generation.
• Validated the optimized KNN protocol using an independent neuroblastoma patient dataset (n=478).

Main Results:

• Identified specific factors significantly contributing to performance variations in KNN models.
• Developed and validated a KNN data analysis protocol that improves model performance.
• Interpreted the biological and practical significance of the derived KNN models.

Conclusions:

• Systematic evaluation and protocol optimization are essential for reliable KNN modeling in clinical genomics.
• The validated KNN protocol offers a promising approach for disease classification and outcome prediction, comparable to existing clinical factors.