Feature Down-Selection to Improve Supervised Classification by Machine Learning on Mass Spectrometry Imaging Data

Area of Science:

• * Computational Biology
• * Data Science
• * Analytical Chemistry

Background:

• * Mass spectrometry imaging (MSI) generates large datasets.
• * Machine learning (ML) is used to analyze MSI data.
• * Storing and handling large MSI datasets presents challenges.

Purpose of the Study:

• * Evaluate feature reduction strategies for large MSI datasets.
• * Minimize data storage while maintaining classification accuracy.
• * Guide researchers in effective data handling for MSI analysis.

Main Methods:

• * Tested two feature selection strategies on six MSI datasets.
• * Utilized XGBoost machine learning algorithm for classification.
• * Assessed feature selection based on average abundance and Student's t-test.

Main Results:

• * Feature selection by average abundance is effective for modest data reduction.
• * Feature selection by Student's t-test is suitable for aggressive data reduction.
• * Trends held regardless of training set size or cross-validation strategy.

Conclusions:

• * Provides insights into effective feature filtering for MSI data.
• * Helps determine when data reduction strategies are most beneficial.
• * Informs decisions on data reduction versus unrestricted data handling.