Discriminant analysis and feature selection in mass spectrometry imaging using constrained repeated random sampling -...

Area of Science:

• Biomedical Imaging
• Analytical Chemistry
• Computational Biology

Background:

• Mass Spectrometry Imaging (MSI) is increasingly used for biomarker identification in clinical settings.
• Analyzing hyperspectral MSI data presents challenges due to complex spectral and spatial variables.
• Classification model validation is crucial for statistically significant biomarker discovery.

Purpose of the Study:

• To introduce and evaluate Constrained Repeated Random Subsampling Cross-Validation (CORRS-CV) for MSI data.
• To compare CORRS-CV with traditional k-fold CV for validating classification models on multi-image MSI datasets.
• To enhance the robustness of biomarker candidate selection using CORRS-CV combined with rank products.

Main Methods:

• Application of CORRS-CV for internal validation of classification models on multi-image MSI data.
• Utilizing biological replicates to define data splits in CORRS-CV.
• Combining CORRS-CV with rank products for feature selection.
• Consideration of factors like image size, CORRS-CV parameters, spatial pixel similarity, and computation time.

Main Results:

• CORRS-CV provides more accurate model performance estimates than k-fold CV, particularly with scarce biological replicates.
• CORRS-CV is advantageous when independent test sets are unavailable, preventing data wastage.
• The combination of CORRS-CV and rank products improves the robustness of discriminant feature selection for biomarker candidates.
• This approach addresses variabilities inherent in multi-image MSI analysis, especially from clinical human tissues.

Conclusions:

• CORRS-CV is a valuable tool for robust classification model validation in Mass Spectrometry Imaging.
• It offers superior performance estimation compared to k-fold CV in scenarios with limited biological replicates.
• The integration of CORRS-CV with rank products strengthens the reliability of identifying potential biomarkers from complex MSI data.